Industrial robot fault prediction method and system based on industrial internet of things big data

Abstract

The application provides an industrial robot fault prediction method and system based on industrial internet of things big data, and relates to the field of electric digital processing, comprising: collecting historical physical examination data of single-axis sequential operation of an industrial robot during production idle time through industrial internet of things, extracting multi-dimensional features and associated features of the historical physical examination data by using a time series anomaly detection trigger, inputting a LightGBM model for iterative training, including histogram binning discretization, split point selection based on gradient information, decision tree residual fitting and feature weight adaptive optimization, and obtaining an industrial robot fault prediction model; inputting shaft physical examination data of a to-be-tested industrial robot during production idle time into the trained industrial robot fault prediction model to obtain a fault prediction result of the to-be-tested industrial robot. The application can reduce data collection quantity, improve fault prediction accuracy, share fault prediction models among similar industrial robots, and reduce model maintenance cost.

Description

Technical Field

[0001] This invention relates to the field of electronic digital processing technology, and in particular to a method and system for predicting industrial robot faults based on big data from the Industrial Internet of Things. Background Technology

[0002] Currently, industrial robots are increasingly widely used in manufacturing, and their operational stability directly affects production efficiency and maintenance costs. However, traditional industrial robot fault prediction methods mainly rely on vibration sensor data acquisition in a laboratory environment, constructing predictive models by monitoring parameters such as motor current, torque, and temperature.

[0003] While these methods perform well under laboratory conditions, they face numerous challenges in real-world industrial scenarios. First, large-scale deployment of vibration sensors leads to soaring costs, and issues such as installation location, power supply, and signal cabling are difficult to resolve. Second, industrial robots typically operate in high-concurrency production networks, and high-frequency data acquisition places enormous pressure on storage and computing. For example, in a car manufacturing plant, if 2000 robots collect data at a 2-second interval, the daily data volume would reach 586TB, making long-term storage and analysis highly impractical. Furthermore, the operating trajectories of industrial robots vary significantly depending on load, task, and other working conditions; modeling each device individually would require substantial computing resources and manpower.

[0004] Therefore, how to reduce the complexity of data acquisition and model training while ensuring prediction accuracy has become a key problem that urgently needs to be solved in the field of industrial robot fault prediction. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a method and system for predicting industrial robot faults based on big data from the Industrial Internet of Things, in order to eliminate or improve one or more defects existing in the prior art and solve the problems of high-concurrency data acquisition, storage and network load in the prior art.

[0006] On the one hand, the present invention provides a method for predicting industrial robot faults based on big data from the Industrial Internet of Things, the method comprising the following steps:

[0007] Acquire the axis inspection data of the industrial robot under test during idle production. The axis inspection data is the operating parameters of each axis collected based on the Industrial Internet of Things when the robot under test is controlled to run one of the multiple motion axes in sequence according to the preset inspection process.

[0008] The split-axis physical examination data is input into a pre-trained industrial robot fault prediction model to obtain the fault prediction result of the industrial robot under test.

[0009] The industrial robot fault prediction model is trained based on a robot fault prediction training set. The training data in the robot fault prediction training set includes: historical health check data collected by the Industrial Internet of Things, which shows that each industrial robot runs sequentially along a single axis during idle production according to the preset health check process, and labels corresponding to the historical health check data. The labels are the running states corresponding to the historical health check data, and the running states include normal running state and fault state.

[0010] The industrial robot fault prediction model includes a time-series anomaly detection trigger and a LightGBM model. The time-series anomaly detection trigger can extract features from the historical physical examination data or the axial physical examination data input to the industrial robot fault prediction model to obtain multi-dimensional features, and generate associated features based on the multi-dimensional features. In each iteration of the training of the industrial robot fault prediction model, the parameters of the time-series anomaly detection trigger are fixed. The LightGBM model is used to: discretize the features in the multi-dimensional features and the associated features using histogram binning technology, select the optimal split point for splitting based on the gradient information of the prediction residual, and construct a decision tree; fit the prediction residual using each decision tree, and calculate the weight of each feature in the multi-dimensional features and the associated features through the split gain value to adaptively adjust the priority of each feature when splitting the decision tree; integrate the fault prediction results output by the decision tree in the current round, and update the parameters of the LightGBM model with the goal of minimizing the loss value between the fault prediction result and the label; wherein, the multi-dimensional features include statistical features, model fitting features, and frequency domain transformation features.

[0011] In some embodiments of the present invention, the historical medical examination data includes at least current parameters, torque parameters, and temperature parameters, and the historical medical examination data is obtained based on the following methods:

[0012] Taking PLC-controlled automation islands as units, the central control system issues health check instructions to the industrial robots within the automation islands;

[0013] The operating parameters of each industrial robot are collected as it runs sequentially along a single axis according to the preset health check process. These operating parameters are then used as historical health check data and transmitted to a server for storage via the industrial Internet of Things.

[0014] In some embodiments of the present invention, if the industrial robot or the automation island fails to transmit data and the automatic marking check is not completed, the check will be performed first in the next idle cycle.

[0015] If the industrial robot is in a state awaiting maintenance, a personnel occupancy marker is set, and the preset inspection process in the corresponding automation island is suspended to allow maintenance personnel to work first.

[0016] In some embodiments of the present invention, the multidimensional features are obtained based on the following steps:

[0017] The statistical characteristics are obtained by statistically analyzing the historical physical examination data in the robot fault prediction training set within the preset sliding window.

[0018] The historical physical examination data are fitted using a preset time series model, and the coefficients of the preset time series model are extracted as the model fitting features.

[0019] The time-domain signal of the historical physical examination data is converted into a frequency-domain representation to obtain the frequency-domain transformation feature;

[0020] The statistical features, the model fitting features, and the frequency domain transformation features are output in the form of binary states or confidence scores to obtain the multidimensional features.

[0021] In some embodiments of the present invention, when the time series anomaly detection trigger extracts the multidimensional features, and the extracted features of the same dimension contain multiple parameters, the time series anomaly detection trigger calculates the 2-norm of the multiple parameters, and compresses the multiple parameters into a single value based on the 2-norm.

[0022] In some embodiments of the present invention, the correlation features include hysteresis triggering features, cross-sensor linkage features, and time-series statistical enhancement features, which are obtained based on the following steps:

[0023] Based on the multidimensional features, a threshold is set for each dimension of the feature. When a feature exceeds the corresponding threshold, the time series anomaly detection trigger is marked as activated.

[0024] Based on the multidimensional features, the activation frequency of the time series anomaly detection trigger within a preset time period is statistically analyzed to obtain the hysteresis triggering feature;

[0025] When the time series anomaly detection trigger is marked as active, the data change results of other related sensors within the same time window are statistically analyzed to obtain the cross-sensor linkage feature;

[0026] Within a preset time window after the time series anomaly detection trigger is marked as activated, the historical physical examination data corresponding to the multidimensional features are segmented and statistically analyzed to obtain the time series statistical enhancement features.

[0027] In some embodiments of the present invention, the method further includes training the industrial robot fault prediction model, selecting the optimal split point for splitting based on the gradient information of the prediction residual, and constructing a decision tree, including:

[0028] Calculate the residual between the fault prediction result and the label to obtain the first-order gradient;

[0029] Calculate the loss function between the fault prediction result and the label, and calculate the second derivative of the loss function with respect to the fault prediction result to obtain the second gradient;

[0030] The multidimensional features and the associated features are used as candidate features. The gain of each candidate feature and its split point after splitting is calculated based on the first-order gradient and the second-order gradient. The candidate feature with the largest gain and the split point are selected for node splitting to construct a decision tree. The cumulative split gain of each candidate feature is recorded.

[0031] In some embodiments of the present invention, training the industrial robot fault prediction model further includes:

[0032] The LightGBM model employs a feature randomization strategy, where each decision tree samples from the multidimensional features, the associated features, and the original signal features contained in the historical physical examination data according to a preset ratio.

[0033] The LightGBM model calculates the weights based on the cumulative split gain. The higher the weight, the higher the probability that the feature corresponding to the weight will be retained during the decision tree sampling process.

[0034] In some embodiments of the present invention, obtaining the fault prediction results of the industrial robot under test includes:

[0035] The LightGBM model performs a weighted summation of the outputs of all decision trees to obtain the predicted value, wherein the weight of each decision tree is controlled by a preset learning rate.

[0036] The predicted value is converted into the fault prediction probability of the industrial robot under test by the LightGBM model through the Sigmoid function;

[0037] If the fault prediction probability exceeds a preset threshold, the industrial robot under test is determined to have a fault risk; otherwise, the industrial robot under test is determined to be operating normally.

[0038] On the other hand, the present invention also provides an industrial robot fault prediction system based on industrial Internet of Things big data, including a processor, a memory, and a computer program / instructions stored in the memory. The processor is used to execute the computer program / instructions. When the computer program / instructions are executed, the system implements the steps of any of the methods mentioned above.

[0039] This invention provides a method and system for industrial robot fault prediction based on big data from the Industrial Internet of Things (IIoT). The method includes: collecting historical health check data of an industrial robot during idle periods by sequentially operating each axis via the IIoT; extracting multi-dimensional and correlation features from the historical health check data using a time-series anomaly detection trigger; inputting this data into a LightGBM model for iterative training, including histogram binning discretization, gradient-based split point selection, decision tree residual fitting, and adaptive optimization of feature weights, to obtain an industrial robot fault prediction model; and inputting the axis-by-axis health check data of the industrial robot under test during idle periods into the trained industrial robot fault prediction model to obtain the fault prediction result for the industrial robot under test. This invention significantly reduces the amount of data collected by establishing a health check process for industrial robots during idle periods and adopting a sequential collection of operating parameters for each axis instead of concurrent collection across all axes. Simultaneously, it avoids network congestion caused by high-frequency data collection during production periods due to low network utilization during idle periods, thus optimizing network load. The LightGBM model integrates multiple features, improving the accuracy of fault prediction. Industrial robots of the same type can share predictive models, eliminating the need for separate modeling for each robot and greatly reducing model maintenance costs. At the same time, the use of histogram binning technology compresses the model size and supports incremental updates, saving storage costs and achieving a significant reduction in operation and maintenance costs.

[0040] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0041] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0042] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings:

[0043] Figure 1 This is a schematic diagram illustrating the steps of an industrial robot fault prediction method based on industrial IoT big data in one embodiment of the present invention.

[0044] Figure 2 This is a flowchart illustrating an industrial robot fault prediction method based on big data from the Industrial Internet of Things, according to one embodiment of the present invention.

[0045] Figure 3 This is a flowchart illustrating a preset physical examination process in one embodiment of the present invention.

[0046] Figure 4 This is a schematic diagram of the process of an industrial robot performing a physical examination according to a preset procedure during idle time in one embodiment of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0048] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0049] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0050] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0051] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0052] It should be emphasized here that the step markers mentioned below are not a limitation on the order of the steps, but should be understood as meaning that the steps can be executed in the order mentioned in the embodiments, or in a different order than in the embodiments, or several steps can be executed simultaneously.

[0053] To address the challenges of high-concurrency data acquisition, storage, and network load in existing technologies, this invention provides a method for predicting industrial robot faults based on big data from the Industrial Internet of Things (IIoT). Figure 1 As shown, the method includes the following steps S101~S102:

[0054] Step S101: Obtain the axis inspection data of the industrial robot under test during idle production. The axis inspection data is the operating parameters of each axis collected based on the Industrial Internet of Things when the robot under test is controlled to run sequentially according to the preset inspection process.

[0055] Step S102: Input the split-axis physical examination data into the pre-trained industrial robot fault prediction model to obtain the prediction results of the industrial robot under test.

[0056] Among them, the industrial robot fault prediction model is trained based on the robot fault prediction training set. The training data in the robot fault prediction training set includes: historical health check data of each industrial robot collected through the Industrial Internet of Things, which runs sequentially on a single axis according to a preset health check process during idle production, and the labels corresponding to the historical health check data. The labels are the running status corresponding to the historical health check data, including normal operation status and fault status.

[0057] The industrial robot fault prediction model comprises a time-series anomaly detection trigger and a LightGBM model. The time-series anomaly detection trigger extracts features from historical or axis-specific physical examination data input to the model, yielding multi-dimensional features, and generates associated features based on these features. In each iteration of the industrial robot fault prediction model's training, the parameters of the time-series anomaly detection trigger are fixed. The LightGBM model is used to: discretize the features in the multi-dimensional and associated features using histogram binning technology; select the optimal split point based on the gradient information of the prediction residuals to construct a decision tree; fit the prediction residuals using each decision tree and calculate the weights of each feature in the multi-dimensional and associated features using the split gain value to adaptively adjust the priority of each feature during decision tree splitting; and integrate the fault prediction results output from the current round of decision trees to update the parameters of the LightGBM model, aiming to minimize the loss between the fault prediction result and the label. The multi-dimensional features include statistical features, model fitting features, and frequency domain transformation features.

[0058] like Figure 2 The diagram shows a flowchart of an industrial robot fault prediction method based on big data from the Industrial Internet of Things.

[0059] Industrial robots typically consist of 6 or 7 axes, each controlled by a motor. The standard operating range of each axis is generally consistent. Testing has confirmed that when similar industrial robots perform the same movement on a single axis, slight differences in data may occur due to varying loads, but the overall trend remains consistent. Therefore, a standardized robot health check procedure can be established, requiring all robots to operate according to prescribed actions and collecting data during this process as the basis for the health check results. Furthermore, considering that robot failures are usually single-axis faults, such as motor, motor-driven gearbox, or transmission linkage failures, the health check process can minimize data interference from simultaneous multi-axis testing by employing a sequential, single-axis testing approach.

[0060] In some embodiments, such as Figure 3The diagram illustrates the process of a single industrial robot undergoing a pre-set health check. Its core principle is to collect key parameters of each axis step-by-step by activating them sequentially, avoiding data interference caused by simultaneous operation of multiple axes. Specifically, the industrial robot is moved to a designated position to ensure it does not collide with its own mechanical structure, interfere with external equipment / environment, or trigger software limit alarms during the check. The control system activates individual joint axes sequentially in a preset order (e.g., axis 1 → axis 2 → ... → X-axis), while the remaining axes remain stationary. Key parameters for each axis are collected, such as motor parameters and gearbox / transmission component parameters. Motor parameters include current, torque, temperature, and speed, while gearbox / transmission component parameters include vibration spectrum, noise signal, and hysteresis error. During the check, each axis performs a uniform preset action (e.g., uniform rotation, sinusoidal trajectory) to ensure data comparability.

[0061] In some embodiments, if data acquisition of a certain axis of an industrial robot fails (e.g., signal loss), the system automatically marks it and prioritizes retesting in the next cycle to avoid omissions.

[0062] Industrial robots need to respond to production commands in real time during normal production, resulting in high network load and data collection being easily interfered with. Therefore, this invention proposes an off-peak health check mechanism, which utilizes production breaks, such as lunch breaks, nighttime shutdowns, and weekend maintenance periods, to collect data, thereby avoiding production network congestion, ensuring data integrity, and reducing the impact on production.

[0063] In some embodiments, such as Figure 4 The diagram illustrates the process of an industrial robot performing a pre-set health check during idle production. Specifically, it first determines whether the operating conditions are met. These conditions include: whether it is during a production break or maintenance period, whether the automated island is ready to perform the health check, whether it is not occupied by maintenance personnel, and whether the automated island is in automatic operation mode. If these conditions are not met, the unmet automated island is recorded, a condition-unmet prompt is triggered, and adjustments are made to achieve the desired state. If the conditions are met, the health check instructions are issued to each industrial robot within each PLC-controlled automated island (typically containing 20-30 industrial robots) through the central control system, executing the check as described above. Figure 3 The preset physical examination process shown obtains physical examination data, which is then transmitted to a server for storage via the Industrial Internet of Things (IIoT). An automation island is a typical production unit division method in the field of industrial automation. It refers to an independent work area governed by a central controller (usually a PLC), containing a group of collaboratively working devices (such as industrial robots, conveyor belts, sensors, etc.) capable of completing specific production processes (such as welding, assembly, and painting).

[0064] In some embodiments, the number of automated islands for parallel physical examinations can be dynamically adjusted according to the real-time network load. For example, when the network is idle, multiple automated islands can perform physical examinations simultaneously.

[0065] In some embodiments, if data transmission fails for an industrial robot or an automation island, the system automatically marks the health check as incomplete and prioritizes a re-check in the next idle cycle.

[0066] In some embodiments, for industrial robots that have not completed the physical examination, the physical examination task can be triggered again by setting a time interval, or it can be carried out in the next round of testing. If the physical examination is not carried out multiple times, the system will issue an alarm prompt.

[0067] In some embodiments, the PLC is pre-set with a personnel occupancy flag. By setting the personnel occupancy flag, the corresponding inspection process in the automation island of the PLC can be paused, allowing maintenance personnel to work with priority.

[0068] In some embodiments, the idle-time health check mechanism provided by the present invention supports breakpoint resumption. If the health check is interrupted due to production tasks, it can be resumed from the breakpoint during the next idle time.

[0069] Based on the pre-defined health check process and idle-time health check mechanism described above, historical health check data (such as motor current, torque, temperature, and other single-axis operating parameters) of each industrial robot are collected via the Industrial Internet of Things (IIoT) during idle production hours, following the pre-defined health check process and sequentially running each axis. This data, along with corresponding labels, forms a training set used to train the industrial robot fault prediction model. The labels represent the operating status corresponding to the historical health check data, including normal operating status and fault status.

[0070] A fault prediction model for industrial robots is then constructed. This model includes a time-series anomaly detection trigger and a LightGBM model. The time-series anomaly detection trigger is used to extract features from historical inspection data, and the LightGBM model is used to generate a prediction result on whether the industrial robot is faulty based on the extracted features.

[0071] Specifically, the time series anomaly detection trigger is a multi-method parallel feature extraction and anomaly discrimination mechanism. Its core objective is to transform the collected time series operation data of industrial robots (such as current, vibration signals, etc.) into standardized features for subsequent LightGBM model training and prediction.

[0072] Historical health check data of each axis motor of the industrial robot in the training set are organized into time-series signals and input into a time-series anomaly detection trigger. Multidimensional features are extracted and output as binary signals 0 / 1 or confidence scores. For example, time-series anomaly detection triggers include threshold triggering, mutation detection, and trend drift detection.

[0073] In some embodiments, multidimensional features include statistical features, model fitting features, and frequency domain transformation features.

[0074] In some embodiments, the time-series anomaly detection trigger processes historical physical examination data in parallel using three methods: statistical methods, model fitting, and frequency domain transformation, to extract multidimensional features. Specifically, statistical analysis is performed on the time-series signals of historical physical examination data within a preset sliding time window to obtain statistical features, including mean, variance, range, band energy, and power spectral density. A preset time-series model, such as an ARMA model or an EMG model, is used to fit the time-series signals of the historical physical examination data, and the parameters of the preset time-series model are extracted as model fitting features, such as the coefficients of the autoregressive moving average model and the parameters of the exponentially modified Gaussian distribution. The time-series signals of the historical physical examination data are converted from time-domain signals to frequency-domain representation, including principal component analysis to extract the weights of the first K principal components, wavelet transform, and high-frequency / low-frequency capability coefficients after multi-scale decomposition, to obtain frequency-domain transformation features.

[0075] In some embodiments, when the time series anomaly detection trigger extracts multidimensional features, and the extracted features of the same dimension contain multiple parameters, the time series anomaly detection trigger calculates the 2-norm of the multiple parameters, and compresses the multiple parameters into a single value based on the 2-norm.

[0076] In some embodiments, the time series anomaly detection trigger performs secondary statistics based on multidimensional features and other sensor data to generate associated features, including hysteresis trigger features, cross-sensor linkage features, and time series statistical enhancement features.

[0077] Specifically, based on the extracted multidimensional features, a threshold is pre-set for each dimension of the feature. When a feature exceeds the corresponding threshold, the time series anomaly detection trigger is marked as active. For example, when the current value of 3.2A is greater than the threshold of 3A, the time series anomaly detection trigger is marked as active.

[0078] Multidimensional features are processed to statistically analyze the activation frequency of time-series anomaly detection triggers within a preset time period, yielding hysteresis trigger features, such as the number of temperature exceedance triggers in the past hour. When a time-series anomaly detection trigger is activated, it statistically analyzes the data changes of other associated sensors within the same time window, obtaining cross-sensor linkage features. Within a preset time window after the time-series anomaly detection trigger is activated, the historical physical examination data time-series signals corresponding to the multidimensional features are segmented and statistically analyzed to obtain time-series statistical enhancement features, such as the mean slope and peak fluctuation within 5 minutes after triggering.

[0079] The training set constructed above is used to iteratively train the industrial robot fault prediction model to be trained. It should be noted that during training, only the LightGBM model is trained, and the parameters of the time series anomaly detection trigger are fixed.

[0080] Specifically, in one iteration, the time-series anomaly detection trigger extracts multi-dimensional features from historical physical examination data in the training set based on the method described above. Correlation features are then derived from these multi-dimensional features, and both multi-dimensional and correlation features are input into the LightGBM model. The LightGBM model employs histogram binning technology, discretizing the continuous feature values ​​contained in the input multi-dimensional and correlation features into a fixed number of histogram bins to reduce computational complexity. The binning process for different features can be performed in parallel. Based on the gradient information of the preceding model obtained from the previous iteration, the optimal split point is selected for splitting, constructing a decision tree. Each decision tree fits the prediction residual of the preceding model, gradually approximating the optimal solution of the loss function through gradient descent. The weights of each feature in the multi-dimensional and correlation features are automatically adjusted by calculating the split gain, assigning higher weights to high-frequency effective features. The fault prediction results output by the current round of decision trees are integrated, with the goal of minimizing the loss between the fault prediction results and the labels, to update the parameters of the LightGBM model. When the preset number of iterations is reached or the loss value converges to a preset threshold, the iteration stops, and all current decision trees are integrated into a decision tree ensemble model, which serves as the trained LightGBM model. The trained LightGBM model (decision tree ensemble model) can generate fault prediction results for industrial robots based on multi-dimensional features and correlation features.

[0081] It should be noted that in the first round of training, since no decision tree has been generated yet, i.e. there is no preceding model, the LightGBM model will pre-set an initial prediction value as a baseline.

[0082] In some embodiments, in each iteration of LightGBM, the algorithm constructs a new tree using the Gradient Boosting Decision Tree (GBDT) framework. Its core is to dynamically select the optimal splitting features and splitting points using the first-order gradient (residual) and second-order gradient (curvature), while automatically adjusting feature weights using splitting gain. The entire process includes:

[0083] The first gradient is obtained by calculating the residual between the prediction result of the preceding model and the true label. The second gradient is obtained by calculating the loss function between the prediction result of the preceding model and the true label, and calculating the second derivative of the loss function with respect to the failure prediction result of the preceding model. The first gradient guides the direction of residual fitting (e.g., the larger the residual, the more attention the current decision tree should pay to that sample), and the second gradient is used for the selection of weighted split points (splitting regions with greater curvature yields higher benefits).

[0084] The input multidimensional features and associated features are used as candidate features. Based on the first-order gradient and the second-order gradient, the gain of each candidate feature and its split point is calculated, as shown in formula (1):

[0085] ; (1)

[0086] in, Indicates gain; Indicates the first-order gradient; Indicates the second-order gradient; This represents the regularization coefficient.

[0087] Select the feature with the largest gain and the split point to split the node and construct a decision tree.

[0088] The cumulative split gain of each feature across all decision trees is recorded, and the weight of each feature is adjusted based on the cumulative split gain to reflect the importance of the corresponding feature. For example, high-frequency effective features, such as vibrational burst triggers, are preferentially selected due to their high split gain. In subsequent iterations, the split points of features with high split gain are explored more frequently.

[0089] The goal of each new decision tree is to fit the negative gradient of the preceding model, i.e., the residual.

[0090] In some embodiments, a feature randomization strategy is employed, where each decision tree samples from multidimensional features, associated features, and original signal features contained in historical physical examination data according to a preset ratio. Preferably, each decision tree randomly selects 70% of the multidimensional features and associated features, and 30% of the original signal features from the historical physical examination data to prevent overfitting. During feature sampling, features with high splitting gain, i.e., high weights, are more likely to be retained.

[0091] In some embodiments, the outputs of all decision trees are integrated to output the final fault prediction result, specifically including:

[0092] The weights of each decision tree are controlled by a preset learning rate (e.g., 0.1). For example, the predicted value output by each decision tree is multiplied by the learning rate (e.g., 0.1) before being accumulated into the ensemble model, preventing a single tree from excessively influencing the overall prediction. It should be noted that the output of each decision tree for the sample is an unscaled predicted value, which may exceed [0,1]. The outputs of all decision trees are weighted and accumulated to obtain the predicted value. Then, the predicted value is converted into an industrial robot fault prediction probability of [0,1] using the Sigmoid function. A preset threshold, such as 0.5, is set. If the fault prediction probability is greater than or equal to 0.5, the industrial robot is considered faulty; otherwise, it is considered normal.

[0093] Based on the industrial robot prediction model obtained from the above training, only the axis inspection data of the industrial robot under test during its idle production period needs to be input to output the fault prediction results of the industrial robot under test.

[0094] That is, in step S101, the axis-by-axis physical examination data of the industrial robot under test is obtained. The axis-by-axis physical examination data is the running parameters of each axis collected by the Industrial Internet of Things when the robot under test is controlled to run sequentially according to the preset physical examination process. This data is used for subsequent prediction.

[0095] In step S102, the split-axis physical examination data is input into the industrial robot fault prediction model trained above to obtain the fault prediction result of the industrial robot under test, and to determine whether the industrial robot under test is in normal operation or fault state.

[0096] Corresponding to the above method, the present invention also provides an industrial robot fault prediction system based on industrial Internet of Things big data, including a processor, a memory and a computer program / instructions stored in the memory. The processor is used to execute the corresponding computer program / instructions. When the computer program / instructions are executed, the system implements the steps of the aforementioned method.

[0097] Corresponding to the above method, the present invention also provides an industrial robot fault prediction system based on industrial Internet of Things big data, the system comprising:

[0098] The data processing module is used to acquire the axis inspection data of the industrial robot under test during its idle production period.

[0099] The fault prediction module contains a pre-trained industrial robot fault prediction model, which is used to input the axis-by-axis physical examination data of the industrial robot under test into the industrial robot fault prediction model to obtain the fault prediction results of the industrial robot.

[0100] Corresponding to the above method, the present invention also provides an electronic device including a computer device, the computer device including a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the electronic device performs the steps of the method as described above.

[0101] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned edge computing server deployment method. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0102] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0103] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0104] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0105] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting industrial robot faults based on big data from the Industrial Internet of Things, characterized in that, The method includes the following steps: Acquire the axis inspection data of the industrial robot under test during idle production. The axis inspection data is the operating parameters of each axis collected based on the Industrial Internet of Things when the robot under test is controlled to run one of the multiple motion axes in sequence according to the preset inspection process. The split-axis physical examination data is input into a pre-trained industrial robot fault prediction model to obtain the fault prediction result of the industrial robot under test. The industrial robot fault prediction model is trained based on a robot fault prediction training set. The training data in the robot fault prediction training set includes: historical health check data collected by the Industrial Internet of Things, which shows that each industrial robot runs sequentially along a single axis during idle production according to the preset health check process, and labels corresponding to the historical health check data. The labels are the running states corresponding to the historical health check data, and the running states include normal running state and fault state. The industrial robot fault prediction model includes a time-series anomaly detection trigger and a LightGBM model. The time-series anomaly detection trigger can extract features from the historical physical examination data or the axial physical examination data input to the industrial robot fault prediction model to obtain multi-dimensional features, and generate associated features based on the multi-dimensional features. In each iteration of the training of the industrial robot fault prediction model, the parameters of the time-series anomaly detection trigger are fixed. The LightGBM model is used to: discretize the features in the multi-dimensional features and the associated features using histogram binning technology, select the optimal split point for splitting based on the gradient information of the prediction residual, and construct a decision tree; fit the prediction residual using each decision tree, and calculate the weight of each feature in the multi-dimensional features and the associated features through the split gain value to adaptively adjust the priority of each feature when splitting the decision tree; integrate the fault prediction results output by the decision tree in the current round, and update the parameters of the LightGBM model with the goal of minimizing the loss value between the fault prediction result and the label; wherein, the multi-dimensional features include statistical features, model fitting features, and frequency domain transformation features.

2. The industrial robot fault prediction method based on industrial IoT big data according to claim 1, characterized in that, The historical medical examination data includes at least current parameters, torque parameters, and temperature parameters, and the historical medical examination data is obtained based on the following methods: Taking PLC-controlled automation islands as units, the central control system issues health check instructions to the industrial robots within the automation islands; The operating parameters of each industrial robot are collected as it runs sequentially along a single axis according to the preset health check process. These operating parameters are then used as historical health check data and transmitted to a server for storage via the industrial Internet of Things.

3. The industrial robot fault prediction method based on industrial IoT big data according to claim 2, characterized in that, If the industrial robot or the automation island fails to transmit data, the system will automatically mark the physical examination as incomplete and prioritize the physical examination in the next idle cycle. If the industrial robot is in a state awaiting maintenance, a personnel occupancy marker is set, and the preset inspection process in the corresponding automation island is suspended to allow maintenance personnel to work first.

4. The industrial robot fault prediction method based on industrial IoT big data according to claim 1, characterized in that, The multidimensional features are obtained by the time series anomaly detection trigger based on the following steps: The statistical characteristics are obtained by statistically analyzing the historical physical examination data in the robot fault prediction training set within the preset sliding window. The historical physical examination data are fitted using a preset time series model, and the coefficients of the preset time series model are extracted as the model fitting features. The time-domain signal of the historical physical examination data is converted into a frequency-domain representation to obtain the frequency-domain transformation feature; The statistical features, the model fitting features, and the frequency domain transformation features are output in the form of binary states or confidence scores to obtain the multidimensional features.

5. The industrial robot fault prediction method based on industrial IoT big data according to claim 1, characterized in that, When the time series anomaly detection trigger extracts the multidimensional features, and the extracted features of the same dimension contain multiple parameters, the time series anomaly detection trigger calculates the 2-norm of the multiple parameters, and compresses the multiple parameters into a single value based on the 2-norm.

6. The industrial robot fault prediction method based on industrial IoT big data according to claim 1, characterized in that, The correlation features include hysteresis triggering features, cross-sensor linkage features, and time-series statistical enhancement features, which are obtained based on the following steps: Based on the multidimensional features, a threshold is set for each dimension of the feature. When a feature exceeds the corresponding threshold, the time series anomaly detection trigger is marked as activated. Based on the multidimensional features, the activation frequency of the time series anomaly detection trigger within a preset time period is statistically analyzed to obtain the hysteresis triggering feature; When the time series anomaly detection trigger is marked as active, the data change results of other related sensors within the same time window are statistically analyzed to obtain the cross-sensor linkage feature; Within a preset time window after the time series anomaly detection trigger is marked as activated, the historical physical examination data corresponding to the multidimensional features are segmented and statistically analyzed to obtain the time series statistical enhancement features.

7. The industrial robot fault prediction method based on industrial IoT big data according to claim 1, characterized in that, The process involves outputting the gradient information of the predicted residuals, selecting the optimal split point for splitting, and constructing a decision tree, including: Calculate the residual between the fault prediction result and the label to obtain the first-order gradient; Calculate the loss function between the fault prediction result and the label, and calculate the second derivative of the loss function with respect to the fault prediction result to obtain the second gradient; The multidimensional features and the associated features are used as candidate features. The gain of each candidate feature and its split point after splitting is calculated based on the first-order gradient and the second-order gradient. The candidate feature with the largest gain and the split point are selected for node splitting to construct a decision tree. The cumulative split gain of each candidate feature is recorded.

8. The industrial robot fault prediction method based on industrial IoT big data according to claim 7, characterized in that, Training the industrial robot fault prediction model also includes: The LightGBM model employs a feature randomization strategy, where each decision tree samples from the multidimensional features, the associated features, and the original signal features contained in the historical physical examination data according to a preset ratio. The LightGBM model calculates the weights based on the cumulative split gain. The higher the weight, the higher the probability that the feature corresponding to the weight will be retained during the decision tree sampling process.

9. The industrial robot fault prediction method based on industrial IoT big data according to claim 1, characterized in that, The fault prediction results obtained for the industrial robot under test include: The LightGBM model performs a weighted summation of the outputs of all decision trees to obtain the predicted value, wherein the weight of each decision tree is controlled by a preset learning rate. The predicted value is converted into the fault prediction probability of the industrial robot under test by the LightGBM model through the Sigmoid function; If the fault prediction probability exceeds a preset threshold, the industrial robot under test is determined to have a fault risk; otherwise, the industrial robot under test is determined to be operating normally.

10. An industrial robot fault prediction system based on industrial Internet of Things big data, comprising a processor, a memory, and a computer program / instructions stored in the memory, characterized in that, The processor is configured to execute the computer program / instructions, and when the computer program / instructions are executed, the system implements the steps of the method as described in any one of claims 1 to 9.

Images