Intelligent monitoring and predictive maintenance system for industrial robot cluster operation status

Abstract

The present invention discloses an intelligent monitoring and predictive maintenance system for the operating status of industrial robot clusters. The system comprises a cloud-based main system and multiple subsystems, each of which is configured to correspond to a regional robot cluster. The cloud-based main system includes a core platform layer, a communication layer, a logic layer, and a storage layer. The subsystems include a regional communication layer, a regional logic layer, and a regional storage layer. The regional communication layer is used to transmit data within the subsystem, between subsystems, and between the subsystem and the cloud-based main system. The regional logic layer is equipped with a data preprocessing module, a multi-source data fusion and alignment module, a regional anomaly monitoring module, a robot regional collaboration and optimization module, and a predictive maintenance module. The system aims to improve cluster monitoring and maintenance efficiency, reduce unplanned downtime, and ensure stable and efficient operation of the production line.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial robots, and specifically relates to an intelligent monitoring and predictive maintenance system for the operating status of an industrial robot cluster. Background Art

[0002] In modern industrial production, particularly in the automotive industry, industrial robot swarms have become critical equipment. Their efficient collaborative operation plays a crucial role in improving production line efficiency and reducing unplanned downtime. Robots such as industrial six-axis manipulators undertake core tasks such as welding, assembly, and handling, ensuring the stability and efficiency of the production process. However, traditional monitoring and maintenance methods often rely on fixed maintenance plans and periodic inspections, lacking real-time response capabilities and targeted anomaly detection methods, making it difficult to effectively address the dynamic operating conditions of robot swarms. In complex and volatile industrial environments, traditional monitoring systems face numerous challenges, such as inconsistencies in multi-source data, noise interference, and inadequate data integration technologies. These limitations limit fault prediction and response speed, increasing the risk of production downtime and maintenance costs. In particular, for anomaly monitoring in swarm robots, there is currently no comprehensive technology that enables collaborative learning and real-time anomaly sharing among robots in different areas. Existing systems are typically limited to fault detection for individual robots or local areas, making it difficult to rapidly identify and disseminate anomaly information across the entire swarm. The lack of effective collaborative monitoring strategies leaves swarm robot anomaly recognition and response in swarms unreliable, further limiting the reliability and efficiency of robot swarms in industrial production. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide an intelligent monitoring and predictive maintenance system for the operating status of an industrial robot cluster, aiming to improve the efficiency of cluster monitoring and maintenance, reduce unplanned downtime, and ensure the stable and efficient operation of the production line.

[0004] In order to achieve the above object, the present invention provides the following technical solutions:

[0005] An intelligent monitoring and predictive maintenance system for the operating status of an industrial robot cluster, comprising a cloud-based main system and multiple subsystems, wherein the subsystems are arranged corresponding to regional robot clusters;

[0006] The cloud-based main system includes a core platform layer, a communication layer, a logic layer, and a storage layer; the core platform layer is used to process computing tasks, coordinate data analysis and collaborative learning of each subsystem, and support efficient computing and task scheduling of the system; the communication layer is used to realize data transmission between the cloud-based main system and the subsystems, ensuring real-time synchronization and reliable transmission of data; the logic layer is used to receive abnormal data, similarity thresholds, and similar area information from the subsystems, confirm the abnormal data, and further confirm similar areas based on the similarity thresholds. It is also responsible for coordinating the initiation of cross-regional collaborative learning and promoting collaborative optimization between similar areas; the storage layer is used to store abnormal data, historical monitoring data, similarity thresholds, similar area information, and confirmed abnormal events from each subsystem for subsequent query, analysis, and model update.

[0007] The subsystem includes a regional communication layer, a regional logic layer and a regional storage layer. The regional communication layer is used to realize data transmission within the subsystem, between subsystems, and between the subsystem and the cloud host system;

[0008] The regional logic layer is equipped with a data preprocessing module, a multi-source data fusion and alignment module, a regional anomaly monitoring module, a robot regional collaboration and optimization module, and a predictive maintenance module; the data processing module is used to fill missing values ​​to ensure data integrity, and perform data cleaning, outlier processing, and data standardization on the data after missing values ​​are filled; the multi-source data fusion and alignment module is used to align and normalize data from different sources; the regional anomaly monitoring module is used to capture abnormal behavior; the robot regional collaboration and optimization module is used to identify the most similar area after detecting an anomaly, share abnormal data, and realize cross-regional collaborative learning and parameter optimization; the predictive maintenance module is used to predict the future equipment status to issue early warnings;

[0009] The regional storage layer is used to store robot sensor data.

[0010] Furthermore, the method for filling missing values ​​in the data preprocessing module is:

[0011]

[0012] Where: X filled is the missing value to be filled; X LagI is the filling value obtained by Lagrange interpolation method; X nearest is the effective data point filling value obtained by the adjacent data filling method; X Grap hSAGE is the effective data point filling value obtained by the GraphSAGE filling method; ω1 and ω2 are weights, and: ω1+ω2=1;

[0013] The Lagrange interpolation method is:

[0014]

[0015] Wherein: is the supplementary value adopting the Lagrange filling strategy at time point t md ; and are respectively the data values corresponding to the first three time points t md −3, t md −2 and t md −1 for selecting the missing area;

[0016] The adjacent data filling method is as follows:

[0017]

[0018] Wherein: is the supplementary value obtained by the adjacent data filling strategy at time point t mk ; t c is the central position time point in the data segment of the missing area; s represents the length of the extracted data segment;

[0019] The GraphSAGE filling method is as follows:

[0020]

[0021] Wherein: is the supplementary value predicted by Pred at time point t mk ; is the feature representation of time point t mk at the p-th layer;[[ID=​​​​​​​​​​​​

[0027] Where: Z is the data obtained by data standardization; X is the data after missing data filling; μ and σ are the mean and standard deviation of the data respectively.

[0028] Furthermore, the multi-source data fusion alignment module uses a dynamic time warping method to align the time series of multi-source data to eliminate the dimension, frequency and format differences of different robot data in the region. The method steps are as follows:

[0029] S11: Construct time series data: Organize the preprocessed data into a time series format. Each time series contains a series of data points, and each data point corresponds to a timestamp.

[0030] S12: Define reference template and test template: select one time series as the reference template and other time series as the test templates, and align them with the reference template;

[0031] S13: Calculate the distance matrix: Construct an n*m matrix grid, where n and m are the lengths of the reference template and the test template respectively; each element (i, j) of the matrix represents the distance d between the i-th point in the reference template and the j-th point in the test template i,j :

[0032]

[0033] Where: x i is the horizontal coordinate of the i-th point in the reference template; y j is the ordinate of the jth point in the test template;

[0034] S14: DTM method: Given two time series P and Q of length LP and LQ respectively:

[0035] P=(P1,…,P i ,…,P LP )

[0036] Q=(q1,…,q i ,…,q LQ )

[0037] Where: P i represents the value of the i-th data point in the time series P; q i represents the value of the i-th data point in the time series q;

[0038] The alignment between P and Q is represented by the matching path W, which is expressed as:

[0039] W=(w1,…,w l ,…,w k );wl =(α l ,β l ),k≥max(LP,LQ),1≤α l ≤LP,1≤β l ≤LQ

[0040] Where: (α l ,β l ) represents the αth l The value matches the βth value in Q l value; w l Indicates the path matched in step l;

[0041] The goal of the DTW method is to find the optimal matching path W so that the cumulative distance Minimum is achieved under the following three constraints:

[0042] (1) Boundary constraint: This constraint restricts the path to start at (1,1) and end at (LP,LQ) when aligning sequences, that is:

[0043] (α1,β1)=(1,1)

[0044] (α l ,β l )=(LP,LQ)

[0045] (2) Continuity constraint: This constraint restricts the temporally changing curved paths when aligning sequences. Under this condition, the path must move forward one step at a time, i.e.: l ―α l+1 ≤1 and β l ―β l+1 ≤1;

[0046] (3) Monotonicity constraint: This condition preserves the temporal order of the points and restricts the path to move forward without decreasing, i.e.: l ≥α l+1 And β l ≥β l+1 ;

[0047] Assume D∈R LP×LQ Represents the cumulative distance, then the global minimum cumulative distance can be obtained by the following dynamic programming:

[0048] D i,j =C i,j +min(D i―1,j ,D i―1,j―1 ,D i,j―1 )

[0049] Where: Initial value D 1,1 =C 1,1; The DTW distance of two sequences P and Q is expressed as DTW(P,Q), which is equivalent to D LP,LQ ,The optimal alignment path can be obtained by backtracking in the dynamic ,programming process.

[0050] Furthermore, the regional anomaly monitoring module combines the ETS model and the GAT-GRU model to comprehensively analyze the operating status of the industrial robot cluster, accurately detect and predict abnormal behaviors, and use an optimized fusion loss function to efficiently identify and warn of potential faults, providing reliable support for maintenance decisions and the stable operation of the production line.

[0051] The GAT-GRU model includes a graph attention network (GAT) and a gated recurrent unit (GRU); the graph attention network (GAT) dynamically calculates the importance weights between robot features and time through an attention mechanism, and combines the dual-channel design of feature GAT and time GAT to capture the complex spatial and temporal dependencies of industrial robot cluster data; the gated recurrent unit (GRU) takes the graph embedding vector generated by the graph attention network (GAT) as input, and the gated recurrent unit (GRU) network dynamically adjusts the hidden state by updating the gate and resetting the gate based on the time series modeling process.

[0052] Furthermore, the ETS model predicts future data points using exponential smoothing. The principle is:

[0053] Smoothing equation:

[0054]

[0055] in: is the predicted value for the next time step; y t is the current observation value; is the current predicted value; T t is the trend term; α is the smoothing coefficient, and 0<α<1;

[0056] Trend Adjustment:

[0057]

[0058] Where: T t+1 is the trend component of the next time step; β is the trend smoothing parameter;

[0059] Seasonal Adjustment:

[0060]

[0061] Where: S t+m is the seasonal component after m steps; γ is the seasonal smoothing parameter; and m is the seasonal period.

[0062] Furthermore, the graph attention network (GAT) captures the complex spatial and temporal dependencies of industrial robot cluster data by:

[0063] Based on feature GAT: the number of key parameter features n inside the robot is used as a node, the length of the time series intercepted by the sliding window function is used as the number of features in each node, and the attention weight is calculated:

[0064]

[0065] in, is the attention coefficient between feature GAT node i and adjacent node j; LeakyReLU is a nonlinear activation function; W Fe is the weight matrix based on the feature GAT; α is the learned weight vector; || represents the connection operation; h i and h j They represent the feature vectors of node i and node j in the feature GAT, representing all key parameter features of node i and node j;

[0066] Use the softmax function to normalize the attention weights of the features within the group:

[0067]

[0068] in: is the normalized attention weight between node i and its neighboring node j; N(i) is the set of neighboring nodes of node i;

[0069] Use attention weights to perform weighted averaging on features and update feature representation:

[0070]

[0071] in: It is the overall influence of other nodes on node i based on the feature GAT;

[0072] Based on time GAT: the length l intercepted by each sliding window function is used as a node, the key parameter feature number n is used as the feature in each node, and the learning parameter is adjusted to γ ​​and the weight matrix is ​​adjusted to W Ti ; Calculate weight normalization through the softmax function and use attention weights to update node representations:

[0073]

[0074] in: It is the overall influence of other nodes on node i in the temporal GAT; is the normalized attention weight between node i and its neighboring node j in the time-based GAT; is the feature vector of node j in the time-based GAT, representing all key parameter features of node j;

[0075] Combining the outputs of parallel feature-based GAT and the output of the time-based GAT Fusion is the output of the graph attention network GAT

[0076] Furthermore, the principle of the gated recurrent unit GRU is:

[0077] Update Gate:

[0078] z t =σ(W z ·[h t―1 ,x t ]+b z )

[0079] Where: z t is the update gate at time step t, controlling the previous hidden state; σ is the sigmoid function; W z is the weight matrix of the update gate; h t―1 is the hidden state of the previous time step; x t is the input of the current time step; b z is the bias term;

[0080] Reset the gate:

[0081] r t =σ(W r ·[h t―1 ,x t ]+b r )

[0082] Where: r t It is the reset gate that determines how much information of the previous state is forgotten;

[0083] Candidate hidden states:

[0084]

[0085] in: is a candidate hidden state; W h is the weight matrix of the candidate hidden state; b h is the bias term; * represents the element-wise product, meaning that the output of the reset gate affects which parts of the hidden state should be retained;

[0086] Final hidden state:

[0087]

[0088] Where: ht is the hidden state at the current time step t, which is used for prediction or further processing;

[0089] Prediction output: GRU’s final hidden state h t It will be converted into the final output prediction through a fully connected layer

[0090]

[0091] Where: W o and b o are the weights and biases of the output layer.

[0092] Furthermore, the prediction output of the ETS model is set to The predicted output of the GAT-GRU model is The true observation value is y t , then the fusion loss function is defined as:

[0093] L=ε·L ETS +(1―ε)·L GRU

[0094] Where: ε is a weight parameter, and 0<ε<1, which is used to control the contribution ratio of the ETS model and the GAT-GRU model to the total loss; I ETS and L GRU are the prediction errors of the ETS model and the GAT-GRU model, respectively, expressed as:

[0095]

[0096] Where: y t is the actual value; is the predicted value calculated using the statistical model ETS; N is the total number of samples;

[0097]

[0098] in: is the predicted value obtained using the GAT-GRU model.

[0099] Furthermore, the robot regional collaboration and optimization module identifies the most similar region after detecting an anomaly, shares anomaly data, and implements cross-region collaborative learning and parameter optimization in the following steps:

[0100] S21: Similarity calculation: Assume that area A contains n robots, denoted as A1, A2, ...A n , area B contains m robots, denoted as B1, B2, ... B m ; then robot A i and Robot Bj The cosine similarity is:

[0101]

[0102] Where: cos(θ ij ) is the i-th robot A i With the jth robot B i Similarity; A i,k and B j,k Robot A i and B i The value of the kth feature; d is the dimension of the feature vector;

[0103] Averaging the similarities of all robot pairs gives the overall similarity between region A and region B:

[0104]

[0105] Where sim(A,B) is the average similarity between region A and region B; n and m represent the number of robots in region A and region B, respectively;

[0106] S22: Collaborative learning mechanism: Determine whether the average similarity sim(A,B) between region A and region B is higher than the set similarity threshold: If so, there is similarity between region A and region B, and the confirmed abnormal data is broadcast from region A to region B to trigger collaborative learning in region B; if not, there is no similarity between region A and region B;

[0107] The dynamic update method of the similarity threshold is:

[0108]

[0109] Where: t is the similarity threshold at time step t; e t is the error of the current time step, that is, the deviation between the current false alarm rate or missed alarm rate and the set value; K p is the proportional coefficient, which is used to adjust the immediate impact of the error; K i is the integral coefficient, which is used to adjust the cumulative effect of the error; K d is the differential coefficient, which is used to adjust the rate of change of the error; is the cumulative error from the initial time step to the current time step;

[0110] S23: Multi-objective optimization: Apply the NSGA-II multi-objective optimization algorithm to dynamically optimize the model parameters of each region. The objective function of the multi-objective optimization is:

[0111] min f(θ)=(f1(θ),f2(θ))

[0112] Where θ represents the set of model parameters to be optimized; f1(θ) is the accuracy target of the anomaly monitoring model, and the optimization goal is to maximize the accuracy; f2(θ) is the response time target, and the optimization goal is to minimize the response time;

[0113] Construct a Pareto optimal solution set to achieve a balance between multiple objectives; select the optimal parameter combination θ in the Pareto optimal solution set * Used to update the model;

[0114] S24: Dynamic parameter update: The loss function L(θ) is used to measure the prediction error of the model and is combined with the regularization term to prevent the model from overfitting:

[0115]

[0116] Where: L(θ) is the loss function used to evaluate the prediction accuracy of the model; N is the number of samples, that is, the number of training data points used; y i is the actual observation value of the i-th sample; is the predicted value of the model for the i-th sample under the parameter θ; λ is the regularization coefficient, which is used to control the impact of the regularization term on the loss and avoid overfitting of the model; Ω(θ) is the regularization term, which is used to penalize excessively large parameter values; in the process of minimizing the loss function L(θ), the regional anomaly monitoring module in each area will update the parameters to enhance the monitoring accuracy while maintaining the stability of the model.

[0117] Furthermore, the method for predicting the future operating state of the industrial robot cluster in the predictive maintenance module is:

[0118] Construct a TCN model and use it to predict the future operating status of an industrial robot cluster, obtaining a sequence of key parameters that can characterize the robot status over a period of time in the future.

[0119] Based on historical data and expert experience, the warning threshold range for each key parameter is set, and the warning threshold is divided into two levels: the first warning threshold and the second warning threshold;

[0120] If the predicted value of the key parameter is greater than the set first-level warning threshold, a first-level warning is triggered; if the predicted value of the key parameter is between the first-level warning threshold and the second-level warning threshold, a second-level warning is triggered; or,

[0121] Using the comprehensive scoring method, the warning signals of each parameter are integrated into a global warning signal, and the global warning index is defined as:

[0122]

[0123] Where: i is the importance weight of each parameter; is the predicted value of the i-th parameter; It is the set second-level warning threshold; when the global warning index CWI exceeds the set global warning threshold, a global warning signal is generated.

[0124] The beneficial effects of the present invention are:

[0125] The intelligent monitoring and predictive maintenance system for the operational status of industrial robot clusters, presented in this paper, aims to improve the operational reliability and maintenance efficiency of robot clusters, reduce the risk of unplanned downtime, and optimize production line efficiency. The system comprises multiple subsystems and a cloud-based master system, and its overall design incorporates a cluster mutual learning strategy. Each subsystem is responsible for collecting, monitoring, and performing preliminary analysis of data within a region, while also enabling cross-regional collaborative management through the cloud. The subsystems comprise five modules: a data preprocessing module, a multi-source data fusion and alignment module, a regional anomaly monitoring module, a robot regional collaboration and optimization module, and a predictive maintenance module, forming a comprehensive regional monitoring and maintenance system. The cloud-based master system is responsible for global data management, anomaly identification, and cross-regional collaborative learning. It further identifies and identifies shared anomaly data in similar regions based on similarity thresholds, optimizes models, and enhances overall detection and response capabilities. Within the subsystems, the data preprocessing module is responsible for cleaning, gap filling, and stabilization to ensure the accuracy and consistency of data input. The multi-source data fusion and alignment module integrates data from different sources, resolving dimensionality and acquisition frequency inconsistencies and providing high-quality data for subsequent monitoring. The regional anomaly monitoring module performs detailed monitoring of multidimensional data and identifies anomaly signatures. When an anomaly is detected, the robot regional collaboration and optimization module identifies similar regions based on cosine similarity. It then uploads the anomaly data, similarity threshold, and similar region information to the cloud-based main system and broadcasts it to these regions to trigger collaborative learning, thereby optimizing model parameters. The predictive maintenance module implements future state predictions and fault warnings, supporting optimized maintenance plans and reducing the risk of sudden failures. This system enables comprehensive monitoring and collaborative optimization of robot clusters, providing intelligent maintenance support for enterprises and significantly improving the operational stability of robot clusters and production line efficiency.

[0126] The intelligent monitoring and predictive maintenance system for the operating status of industrial robot clusters of the present invention has the following advantages:

[0127] 1. Advantages of system architecture design:

[0128] (1) Improvement of system data processing quality: The system of the present invention includes a cloud-based main system and multiple subsystems. By configuring a data preprocessing module in the regional logic layer of each subsystem and filling in missing values ​​in the data, the integrity and consistency of the data are ensured; by performing data cleaning, outlier processing and data standardization on the data, the data quality can be effectively improved, providing a reliable data basis for subsequent anomaly monitoring and fault prediction; (2) Accurate anomaly monitoring and fault prediction: The regional anomaly monitoring module in the subsystem architecture adopts the ETS model and the GAT-GRU model, combining the stability of the statistical model with the adaptability of the deep learning model, and can realize refined monitoring under multi-dimensional data, effectively capture abnormal patterns in complex scenarios, improve the accuracy and timeliness of fault detection, and improve the reliability of monitoring results.

[0129] 2. Advantages of Subsystem Modular Design

[0130] (1) Efficient collaborative learning and dynamic optimization: With the support of the robot regional collaboration and optimization module, abnormal data between regions is shared. Similar regions are identified through the cosine similarity analysis of the subsystem, and abnormal data is propagated to regions with similar characteristics, thereby achieving cross-regional collaborative learning. The NSGA-II multi-objective optimization algorithm within the system further balances monitoring accuracy and response time, dynamically adjusts the model parameters within the region, and ensures efficient operation of each region in the cluster; (2) Optimize maintenance plans and resource allocation: Through the predictive maintenance module in the subsystem, the long-term operating characteristics of the robot are captured, and the equipment status can be predicted in advance, providing data support for maintenance decisions. This method makes maintenance plans more scientific and resource allocation more reasonable, significantly reducing unnecessary maintenance costs.

[0131] 3. Enhancement of overall system performance

[0132] (1) Reduce unplanned downtime: Based on the system's accurate fault prediction and real-time warning functions, the failure risk of robot clusters can be significantly reduced, reducing unplanned downtime, thereby improving the stability and efficiency of the production line, reducing production losses, and improving overall production efficiency. (2) Enhance the scalability and adaptability of the system: The system's modular design is flexible and can be easily integrated into industrial environments of different types and sizes to meet the needs of various application scenarios. The system structure is convenient for future technology upgrades and expansion adjustments, and has high long-term application value and good scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0133] In order to make the purpose, technical solutions and beneficial effects of the present invention more clear, the present invention provides the following drawings for illustration:

[0134] Figure 1 This is an architectural diagram of the industrial robot cluster intelligent monitoring and predictive maintenance system of the present invention;

[0135] Figure 2 This is the software architecture diagram of the cloud main system; Figure 3 It is the main interface for monitoring the robot cluster of the cloud main system;

[0136] Figure 4 Provide detailed analysis of regional robots for the cloud main system; Figure 5 This is the subsystem software architecture diagram;

[0137] Figure 6 This is a flow chart of the overall cluster mutual learning strategy of the system;

[0138] Figure 7 An example diagram of filling in missing data for the subsystem data preprocessing module;

[0139] Figure 8 This is a DTM flow diagram of the subsystem's multi-source data fusion and alignment module;

[0140] Figure 9 This is the ETS model workflow diagram for the subsystem regional anomaly monitoring module;

[0141] Figure 10 This is a schematic diagram of the GAT model structure of the subsystem regional anomaly monitoring module;

[0142] Figure 11 This is a schematic diagram of the GRU model structure of the subsystem regional anomaly monitoring module;

[0143] Figure 12 Provide a flow chart for the subsystem area collaborative optimization module;

[0144] Figure 13 This is a schematic diagram of the overall structure of the TCN model of the subsystem predictive maintenance module;

[0145] Figure 14 This is a flow chart of the five modules included in the subsystem. DETAILED DESCRIPTION

[0146] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0147] like Figure 1As shown, the intelligent monitoring and predictive maintenance system for the operating status of industrial robot clusters in this embodiment includes a cloud-based main system and multiple subsystems, and the subsystems are set corresponding to the regional robot clusters. Each subsystem is regarded as an independent operating unit, which contains five major modules: data preprocessing module, multi-source data fusion and alignment module, regional anomaly monitoring module, robot regional collaboration and optimization module, and predictive maintenance module. Each subsystem is responsible for the collection, processing, monitoring and maintenance of robot data in its area, and realizes the integration of module functions and the independence of monitoring within the area through a three-layer architecture (regional communication layer, regional logic layer, and regional storage layer). The cloud-based main system undertakes the tasks of global management and data coordination to ensure the coordinated operation and optimization of each subsystem.

[0148] 1. System architecture design

[0149] In the system design of this embodiment, each robot zone is considered an independent subsystem. The subsystem's main components include: a data preprocessing module, a multi-source data fusion and alignment module, a regional anomaly monitoring module, a robot zone collaboration and optimization module, and a predictive maintenance module. Each subsystem is responsible for data collection, monitoring, and maintenance of the robots in its zone, and global collaboration and intelligent management are achieved through an integrated cloud-based master system. Building on this architecture, the system also integrates a cluster mutual learning strategy, enabling efficient sharing of anomaly knowledge among subsystems, improving the overall anomaly detection efficiency of the cluster.

[0150] 1.1. Cloud Main System Architecture Design

[0151] The cloud main system is responsible for the overall management and coordination of the cluster level, realizing collaborative learning and unified control among regions. The cloud main system software architecture is as follows: Figure 2 In this embodiment, the cloud main system includes a core platform layer, a communication layer, a logic layer, a storage layer, and a front-end display layer.

[0152] The core platform layer handles computing tasks. It features high-performance computing servers (such as GPU servers) and a technology stack supporting both the front-end and back-end platforms, including Vue.js, JavaScript, and HTML (front-end presentation frameworks), Django (back-end framework), and a MySQL database. The core platform layer is responsible for handling complex computing tasks, data management and processing, model optimization, and scheduling within the cloud-based main system, ensuring system efficiency and scalability.

[0153] The communication layer is used to realize data transmission between the cloud main system and subsystems, ensure the real-time transmission of abnormal data, monitoring results and optimization instructions, and maintain the synchronization and integrity of the system.

[0154] The logic layer is the core functional module of the cloud-based main system, receiving anomaly data, similarity thresholds, and similar area information from each subsystem. The logic layer first verifies the validity of the uploaded anomaly data with equipment engineers and further identifies similar areas based on the set similarity thresholds. Once confirmed, the anomaly data is transmitted to these similar areas. The logic layer is responsible for coordinating the initiation of cross-region collaborative learning, promoting collaborative optimization and model updates among similar areas.

[0155] The storage layer is used to store abnormal data, historical monitoring data and confirmed abnormal events from each subsystem, supporting model optimization, historical tracing and further analysis, and providing a reliable data foundation for long-term optimization and abnormality analysis of the system.

[0156] Front-end display layer: provides a real-time interactive interface. The main page displays the distribution of all robots in each area, as well as overall operation information (such as the number of running robots, abnormal detection ratio and production efficiency, etc.). Figure 3 The sub-interface supports viewing the robot operation details in a specific area, including statistical analysis (such as trend strength, seasonality strength, etc.) and in-depth analysis results (such as anomaly detection, predictive maintenance warning), and provides maintenance options to support the arrangement of maintenance tasks, such as Figure 4 The intuitive interactive features of the front-end display layer help managers make efficient decisions and significantly improve the convenience of system monitoring and maintenance.

[0157] Under the control of the cloud-based master system, each subsystem uploads data to the cloud when it detects an anomaly. After engineers confirm the nature of the anomaly, the cloud-based master system further identifies similar areas based on the uploaded similar area information and transmits the anomaly data to these similar areas, triggering collaborative learning and model optimization.

[0158] 1.2. Subsystem Architecture Design

[0159] The subsystem of each robot area includes the following five modules: data preprocessing module, which ensures the integrity and consistency of the data and adopts a comprehensive filling method of Lagrange interpolation, neighboring data filling method and GraphSAGE filling to solve the problem of missing data; multi-source data fusion and alignment module, which processes data from different devices through time series alignment and feature normalization technology; anomaly monitoring module, which combines the improved ETS model and GAT-GRU joint model for anomaly detection; robot area collaboration and optimization module, which is used to realize collaborative learning between different areas of the robot cluster. When an anomaly occurs in a certain area and is confirmed, the system selects similar areas based on cosine similarity to share anomaly data and trigger collaborative learning; and predictive maintenance module, which uses temporal convolutional network (TCN) to predict future equipment status and establish an early warning system. Subsystem architecture and function: Each subsystem consists of a three-layer structure, including regional communication layer, regional logic layer and regional storage layer, such as Figure 5 shown.

[0160] The regional communication layer facilitates data transmission within subsystems, between subsystems, and between a subsystem and the main system. This ensures data synchronization within the region and, upon detecting an anomaly, uploads the anomaly data, similarity threshold, and similarity region information to the cloud-based main system in real time. Inter-subsystem data transmission can receive anomaly data from other subsystems through the communication layer and transmit it to the logic layer for parameter optimization.

[0161] The regional logic layer includes a data preprocessing module, a multi-source data fusion and alignment module, a regional anomaly monitoring module, a robot regional collaboration and optimization module, and a predictive maintenance module. The data processing module fills missing values ​​to ensure data integrity and performs data cleaning, outlier processing, and data normalization on the filled data. The multi-source data fusion and alignment module aligns and normalizes data from different sources. The regional anomaly monitoring module captures abnormal behavior. The robot regional collaboration and optimization module identifies the most similar regions after detecting an anomaly, shares anomaly data, and enables cross-regional collaborative learning and parameter optimization. The predictive maintenance module predicts future equipment status to provide early warnings. In this embodiment, the regional logic layer is responsible for data preprocessing, fusion, monitoring, and optimization. When an anomaly is detected in a region, the logic layer analyzes the anomaly, triggers the predictive maintenance module, and uploads the anomaly data, similarity threshold, and similar region information to the main system via the communication layer. After receiving anomaly data from other subsystems, the system updates parameters based on similarity and the current model state to improve the system's recognition accuracy for similar anomalies.

[0162] The regional storage layer stores state data from the robot's various sensors, facilitating subsequent model optimization and anomaly analysis. This data is periodically updated to ensure continuous model learning and accurate predictions.

[0163] 1.3 Cluster mutual learning strategy

[0164] This strategy significantly improves the response speed and fault prevention capabilities of the robot cluster in abnormal monitoring through collaborative learning, dynamic updating, and adaptive similarity adjustment between regions, thereby improving production efficiency, reducing failure rates, and effectively lowering maintenance costs. The overall process of the mutual learning strategy is as follows: Figure 6 shown.

[0165] 1.3.1. Initialization of the fusion model

[0166] In this embodiment, the robot cluster in each subsystem area is equipped with a fusion anomaly monitoring model that combines the improved ETS model (exponential smoothing state space model) and the GAT-GRU model (graph attention network-gated recurrent unit model). The model is initially trained using periodic data of normal operating parameters in the area. The initial model training takes into account the possible differences in robot models in the area, so that the model adapts to the different operating characteristics of the robots in the group, thereby enhancing the synergy of the robots within the subsystem in anomaly monitoring. The fusion model combines the advantages of ETS in time series smoothing, the accuracy of GAT in spatial relationship modeling, and the powerful ability of GRU in long-term dependency modeling, ensuring that the monitoring model is sensitive to periodic anomalies and can identify nonlinear features in complex scenarios.

[0167] 1.3.2. Upload and confirm abnormal data of subsystem

[0168] During actual operation, the sub-models within each subsystem monitor the operating status of the robots in the area in real time and upload the monitoring data to the main cloud system. When a subsystem detects an abnormal event, the abnormal data is uploaded to the cloud through the communication layer, where engineering personnel verify the authenticity of the abnormal data. If the abnormal event is confirmed to be a real fault, the abnormal data and the corresponding subsystem ID are stored and recorded to facilitate accurate tracking and comparison in subsequent analysis of similar areas. This confirmation mechanism ensures that the system is based on reliable abnormal information during collaborative learning, preventing false positives from propagating between clusters and reducing the false alarm rate of the entire system.

[0169] 1.3.3 Periodic Similarity Detection

[0170] In order to improve the effectiveness of abnormal data transmission between regions, the system of this embodiment is designed with a regular similarity detection mechanism between the cloud main system and each subsystem. Each subsystem periodically aligns the key parameters of the robots in the area during the monitoring period to generate a function distribution of the parameters. On this basis, each subsystem calculates the similarity between itself and the models of other subsystems, thereby quantifying the differences in predicted data between different subsystems. This similarity value will be uploaded to the cloud main system along with the detected abnormal data. By receiving these similarity data and abnormal information, the cloud main system can more effectively analyze the applicability and reference value of abnormal data in different regions. When conducting collaborative learning, the system will give priority to subsystems with high similarity to the fault area for abnormal data sharing based on these similarity results, so as to ensure the effectiveness and accuracy of abnormal transmission, reduce interference in information transmission and improve the overall collaborative learning effect.

[0171] 1.3.4. Similar Area Data Broadcast and Dynamic Threshold Adjustment

[0172] Based on the similarity detection results, the cloud-based main system broadcasts anomaly data from other subsystems with similar characteristics to each subsystem. Based on a set threshold condition, the system triggers the anomaly data sharing mechanism only when the similarity between two regions exceeds the set threshold. This ensures that data sharing occurs only between the most relevant subsystems, improving learning efficiency. To further optimize the system's adaptability, this embodiment uses a PID (proportional-integral-differential) control algorithm to dynamically adjust the similarity threshold, enabling the system to automatically adapt to the appropriate collaborative learning range under different abnormal conditions and operating environments. In actual operation, when a similar region receives anomaly data shared by other subsystems, it triggers a parameter update of the region model to enhance the model's detection accuracy for similar anomalies. Through dynamic threshold control, the system can adjust the coverage of collaborative learning in a timely manner, making collaborative learning between subsystems more flexible and efficient. The updated monitoring results of the similar region model are further uploaded to the main system for reference and optimization by other subsystems, forming a closed-loop learning mechanism.

[0173] 2. Data preprocessing module

[0174] In this embodiment, the data preprocessing module is used to process monitoring data from robot clusters in industrial production environments. This module first fills in missing data points, then performs data cleaning, outlier processing, and data normalization to ensure data quality and provide high-quality input for subsequent anomaly monitoring and predictive maintenance analysis. Before conducting research on multi-source data fusion technology for robot clusters, comprehensive preprocessing of the multi-dimensional data collected by the robot cluster is required. This is specifically divided into the following steps.

[0175] 2.1 Missing Data Filling

[0176] Missing data is a common problem in robot swarm operation data, which can be caused by equipment failure, data transmission interruption, or other technical issues. To ensure data integrity, we use a combination of the following two methods to perform comprehensive data filling.

[0177] 2.1.1. Neighboring Data Filling

[0178] Using data from adjacent time points to fill missing values ​​is suitable for situations where there is less missing data and the time series is continuous. mk , the total length of the missing area is s, and data segments of length s / 2 are extracted before and after the missing area respectively without any calculation and directly used as the filling value of the missing area. In this embodiment, the adjacent data filling method is:

[0179]

[0180] in: It is t mk The supplementary value obtained by the adjacent data filling strategy at the time point; t c is the central position time point in the missing area data segment; s is the length of the extracted data segment.

[0181] 2.1.2 Lagrangian filling

[0182] When dealing with missing values ​​in time series, the Lagrange interpolation method is a classic polynomial interpolation method that accurately estimates the value of the unknown point by constructing a polynomial using the known data points adjacent to the missing point. The general formula of Lagrange interpolation is defined as:

[0183]

[0184] Where: L(t) is the interpolation polynomial used to predict any time point t; y i corresponds to the time point t i The known data value of i (t) is the i-th basis polynomial, and the specific calculation is This product traverses all indexes j except i to construct a polynomial. In this embodiment, for the time t at which the data is missing, md , select the first three time points t of the missing region md 3. t md ―2 and t md ―1, the corresponding data values ​​are and By substituting the data of these three points into the Lagrange interpolation polynomial, we can get t md The predicted data value at the moment is . In this embodiment, the Lagrange interpolation method is:

[0185]

[0186] in: At time point t md Supplementary values ​​using Lagrangian filling strategy; and The first three time points t of the selected missing region are md 3. t md ―2 and t md -The data value corresponding to 1.

[0187] 2.1.3 GraphSAGE filling

[0188] In the GraphSAGE model, a one-way propagation relationship between nodes is defined, and a network structure containing P layers is constructed. This setting enables the model to iteratively update the node features of each layer, and finally effectively estimate and fill t through the prediction function of Pred. md The value at the time point. In this embodiment, the GraphSAGE filling method is:

[0189]

[0190] in: is predicted by Pred mk The supplementary value of the moment; is t of the pth layer mk Feature representation at a time point.

[0191] 2.1.4 Comprehensive gap filling

[0192] Combining the three methods of neighboring data filling, Lagrangian filling and GraphSAGE filling, the weight coefficients ω1 and ω2 are set to satisfy ω1+ω2=1. Figure 7 As shown, the method for filling missing values ​​in the data preprocessing module of this embodiment is:

[0193]

[0194] Where: X filled is the missing value to be filled; X LagI is the filling value obtained by Lagrange interpolation method; X nearest is the effective data point filling value obtained by the adjacent data filling method; X Grap hSAGE is the effective data point filling value obtained by the GraphSAGE filling method; ω1 and ω2 are weights, and: ω1+ω2=1.

[0195] By comprehensively using these three filling methods, the integrity and consistency of the data can be effectively guaranteed. This not only helps to avoid possible errors in subsequent analysis, but also ensures the accuracy and reliability of the analysis results. The integrity of the data is the basis for ensuring accurate prediction and anomaly detection of the model, so this comprehensive data filling strategy is crucial.

[0196] 2.2. Outlier Handling (IQR)

[0197] In this embodiment, the method for the data preprocessing module to handle outliers in the data after missing value filling is as follows: Calculate the first quartile (Q1) and the third quartile (Q3) of the data, that is:

[0198] IQR = Q3 - Q1

[0199] Based on IQR, calculate the boundaries of outliers: x < Q1 - 1.5×IQR or x > Q3 + 1.5×IQR; all points falling outside this range are considered outliers; if the outliers are considered to be caused by errors or accidental large deviations and the number is small, these outliers are directly deleted from the dataset; if there is a certain number of outliers, the average value of the overall data is used to replace the outliers.

[0200] 2.3. Data Standardization

[0201] In this embodiment, the method for the data preprocessing module to perform data standardization on the data after missing value filling is as follows: Convert the data into a form with zero mean and unit variance to eliminate the influence between data of different magnitudes. The formula is as follows:

[0202]

[0203] Where: Z is the data obtained by data standardization processing; X is the data after missing value filling; μ and σ are the mean and standard deviation of the data.

[0204] The data preprocessing module outputs high-quality data, providing reliable input for subsequent anomaly detection and predictive maintenance. The comprehensive preprocessing method adopted by the data preprocessing module improves the accuracy and efficiency of the entire monitoring system.

[0205] 3. Multi-source Data Fusion and Alignment Module

[0206] The multi-source data fusion and alignment module in the subsystem aims to process the data of different robots in the area to solve the problems of differences in data dimensions, data ranges, data collection frequencies, and data upload formats. In this embodiment, the multi-source data fusion and alignment method (DTM) is adopted to ensure the consistency and comparability of the data, thereby improving the accuracy of subsequent analysis.

[0207] DTW (Dynamic Time Warping) is an algorithm that can measure the similarity between two different time series data and adjust their lengths, such as Figure 8 Compared to the Euclidean distance, which can more accurately measure the similarity of sequence data, the DTW method is used to align the data collected by the robot, which can ensure that the morphology of different time series data is as consistent as possible, so as to facilitate subsequent data fusion.

[0208] Specifically, in this embodiment, the multi-source data fusion alignment module uses a dynamic time warping method to perform time series alignment on the multi-source data to eliminate the dimension, frequency and format differences of different robot data in the region. The method steps are as follows.

[0209] S11: Construct time series data: Organize the preprocessed data into the form of time series. Each time series contains a series of data points, and each data point corresponds to a timestamp.

[0210] S12: Define reference template and test template: Select one time series as the reference template and other time series as the test templates, and align them with the reference template.

[0211] S13: Calculate the distance matrix: Construct an n*m matrix grid, where n and m are the lengths of the reference template and the test template respectively; each element (i, j) of the matrix represents the distance d between the i-th point in the reference template and the j-th point in the test template i,j :

[0212]

[0213] Where: x i is the horizontal coordinate of the i-th point in the reference template; y j is the ordinate of the j-th point in the test template.

[0214] S14: DTM method: Given two time series P and Q of length LP and LQ respectively:

[0215] P=(P1,…,P i ,…,P LP )

[0216] Q=(q1,…,q i ,…,q LQ )

[0217] Where: P i represents the value of the i-th data point in the time series P; q i Represents the value of the i-th data point in the time series Q;

[0218] The alignment between P and Q is represented by the matching path W, which is expressed as:

[0219] W=(w1,…,w l ,…,w k );w l =(α l ,β l ),k≥max(LP,LQ),1≤α l ≤LP,1≤β l ≤LQ

[0220] Where: (α l ,β l ) represents the αth l The value matches the βth value in Q l value; w l Indicates the path matched in step l.

[0221] The goal of the DTW method is to find the optimal matching path W so that the cumulative distance Minimum is achieved under the following three constraints:

[0222] (1) Boundary constraint: This constraint restricts the path to start at (1,1) and end at (LP,LQ) when aligning sequences, that is:

[0223] (α1,β1)=(1,1)

[0224] (α l ,β l )=(LP,LQ)

[0225] (2) Continuity constraint: This constraint restricts the temporally changing curved paths when aligning sequences. Under this condition, the path must move forward one step at a time, i.e.: l ―α l+1 ≤1 and β l ―β l+1 ≤1.

[0226] (3) Monotonicity constraint: This condition preserves the temporal order of the points and restricts the path to move forward without decreasing, i.e.: l ≥α l+1 And β l ≥β l+1 .

[0227] Assume D∈R LP×LQ Represents the cumulative distance, then the global minimum cumulative distance can be obtained by the following dynamic programming:

[0228] D i,j =C i,j +min(D i―1,j ,D i―1,j―1 ,Di,j―1 )

[0229] Where: Initial value D 1,1 =C 1,1 ; The DTW distance of two sequences P and Q is expressed as DTW(P,Q), which is equivalent to D LP,LQ ,The optimal alignment path can be obtained by backtracking in the dynamic ,programming process.

[0230] Through this multi-source data fusion and alignment module, the system can effectively process diverse data from different robots within a region, providing a consistent and accurate data foundation for anomaly monitoring and equipment maintenance. This approach significantly improves the flexibility of data processing and the effectiveness of subsequent analysis, and is one of the key technologies of this invention.

[0231] 4. Regional anomaly monitoring module

[0232] The anomaly monitoring module uses a joint deep learning approach combining an improved ETS model (exponential smoothing state space model) and a GAT-GRU model (graph attention network-gated recurrent unit model) to detect anomalies in preprocessed data from industrial robot clusters. Specifically, the ETS model first models time series data. By decomposing the trend, seasonality, and residual components in the time series, it smooths fluctuations in the data and extracts periodic behavior characteristics, providing a stable baseline for anomaly monitoring while effectively reducing the impact of noise interference and ensuring the robustness of the initial modeling. In the deep learning component, a joint GAT and GRU architecture is utilized to deeply model the spatial and temporal characteristics of the data. By integrating the time series stability of the ETS model with the spatial and temporal characteristics of the GAT-GRU model, the anomaly monitoring module can comprehensively analyze the operating status of the industrial robot cluster, accurately detect and predict abnormal behavior, and, by optimizing the comprehensive loss function, ensure efficient identification and early warning of potential faults, providing reliable support for maintenance decisions and stable production line operation.

[0233] The regional anomaly monitoring module of this embodiment combines the ETS model and the GAT-GRU model to comprehensively analyze the operating status of the industrial robot cluster, accurately detect and predict abnormal behavior, and use an optimized fusion loss function to efficiently identify and warn of potential faults, providing reliable support for maintenance decisions and the stable operation of the production line. The GAT-GRU model includes a graph attention network (GAT) and a gated recurrent unit (GRU); the graph attention network (GAT) dynamically calculates the importance weights between robot features and time through an attention mechanism, and combines the dual-channel design of feature GAT and time GAT to capture the complex spatial and temporal dependencies of industrial robot cluster data; the gated recurrent unit (GRU) takes the graph embedding vector generated by the graph attention network (GAT) as input. The gated recurrent unit (GRU) network dynamically adjusts the hidden state through update gates and reset gates based on the time series modeling process.

[0234] 4.1. ETS Model (Exponential Smoothing State Space Model)

[0235] The ETS model is a time series forecasting model that is suitable for processing data with trends and seasonality, such as Figure 9 The model predicts future data points using exponential smoothing. The principle is:

[0236] Smoothing equation:

[0237]

[0238] in: is the predicted value for the next time step; y t is the current observation value; is the current predicted value; T t is the trend term; α is the smoothing coefficient, and 0<α<1, which adjusts the emphasis on historical data.

[0239] Trend Adjustment:

[0240]

[0241] Where: T t+1 is the trend component of the next time step; β is the trend smoothing parameter;

[0242] Seasonal Adjustment:

[0243]

[0244] Where: S t+m is the seasonal component after m steps; γ is the seasonal smoothing parameter; and m is the seasonal period.

[0245] 4.2 GAT-GRU Model

[0246] GAT dynamically calculates the importance weights between robot features through the attention mechanism, captures the complex spatial dependencies and interaction characteristics within the robot cluster, and embeds each robot feature into a graph representation with a global context. Subsequently, the graph embedding vector generated by GAT is input into the GRU model to further analyze the dynamic changes of the time series. GRU effectively captures long-term temporal dependencies through the update gate and reset gate mechanism, while suppressing irrelevant information, thereby accurately identifying abnormal patterns in time series, such as Figure 10 shown.

[0247] 4.2.1 Graph Attention Network (GAT)

[0248] GAT uses an attention mechanism to dynamically calculate the importance weights between robot features and time. Combining a dual-channel design of feature GAT and time GAT, it captures the complex spatial and temporal dependencies of industrial robot cluster data. Specifically, in this embodiment, the graph attention network (GAT) captures the complex spatial and temporal dependencies of industrial robot cluster data as follows.

[0249] (1) Based on feature GAT: The number of key parameter features n inside the robot is taken as a node, and the length of the time series intercepted by the sliding window function is taken as the number of features in each node. The attention weight is calculated:

[0250]

[0251] in, is the attention coefficient between feature GAT node i and adjacent node j; LeakyReLU is a nonlinear activation function; W Fe is the weight matrix based on the feature GAT; α is the learned weight vector; || represents the connection operation; h i and h j They represent the feature vectors of node i and node j in the feature GAT, respectively, representing all key parameter features of node i and node j.

[0252] Use the softmax function to normalize the attention weights of the features within the group:

[0253]

[0254] in: is the normalized attention weight between node i and its neighboring node j; N(i) is the set of neighboring nodes of node i.

[0255] Use attention weights to perform weighted averaging on features and update feature representation:

[0256]

[0257] in: It is based on the overall influence of other nodes on node i in the feature GAT.

[0258] (2) Based on time GAT: The length l intercepted by each sliding window function is used as a node, the key parameter feature number n is used as the feature in each node, and the learning parameter is adjusted to γ ​​and the weight matrix is ​​adjusted to W Ti ; Calculate weight normalization through the softmax function and use attention weights to update node representations:

[0259]

[0260] in: It is the overall influence of other nodes on node i in the temporal GAT; is the normalized attention weight between node i and its neighboring node j in the time-based GAT; is the feature vector of node j in the time-based GAT, representing all key parameter features of node j.

[0261] Combining the outputs of parallel feature-based GAT and the output of the time-based GAT Fusion is the output of the graph attention network GAT

[0262] 4.2.2 Gated Recurrent Unit (GRU)

[0263] Taking the graph embedding vector generated by GAT as input, the GRU network dynamically adjusts the hidden state through the update gate and reset gate based on the time series modeling process. Specifically, Figure 11 As shown, the principle of the gated recurrent unit GRU is:

[0264] Update Gate:

[0265] z t =σ(W z ·[h t―1 ,x t ]+b z )

[0266] Where: z t is the update gate at time step t, controlling the previous hidden state; σ is the sigmoid function; W z is the weight matrix of the update gate; h t―1 is the hidden state of the previous time step; x t is the input of the current time step; b z is the bias term.

[0267] Reset the gate:

[0268] r t =σ(W r ·[h t―1 ,x t ]+b r )

[0269] Where: r t It is the reset gate that determines how much information of the previous state is forgotten.

[0270] Candidate hidden states:

[0271]

[0272] in: is a candidate hidden state; W h is the weight matrix of the candidate hidden state; b h is the bias term; * denotes element-wise multiplication, meaning that the output of the reset gate affects which parts of the hidden state should be retained.

[0273] Final hidden state:

[0274]

[0275] Where: h t is the hidden state at the current time step t, which is used for prediction or further processing.

[0276] Prediction output: GRU’s final hidden state h t It will be converted into the final output prediction through a fully connected layer

[0277]

[0278] Where: W o and b o are the weights and biases of the output layer.

[0279] 4.3 Weighted Loss Function

[0280] In order to optimize the prediction performance of the ETS model and the GAT-GRU model at the same time, a weighted loss function can be defined. This function combines the prediction errors of the two models and balances the contribution of the two models to the final performance by adjusting the weights.

[0281] 4.3.1 Fusion Loss Function

[0282] Set the prediction output of the ETS model to The predicted output of the GAT-GRU model is The true observation value is y t , then the fusion loss function is defined as:

[0283] L=ε·L ETS +(1―ε)·L GRU

[0284] Where: ε is a weight parameter (0<ε<1), which is used to control the contribution ratio of the ETS model and the GAT-GRU model to the total loss; L ETS and L GRU are the prediction errors of the ETS model and the GAT-GRU model, respectively, which are usually calculated using the mean square error (MSE) and expressed as:

[0285]

[0286] Where: y t is the actual value; is the predicted value calculated using the statistical model ETS; N is the total number of samples.

[0287]

[0288] in: is the predicted value obtained using the GAT-GRU model.

[0289] 4.3.2 Parameter Adjustment and Optimization

[0290] To find the optimal ε, cross-validation or other model selection techniques can be used to evaluate the model performance under different ε values. In addition, gradient descent or other optimization algorithms can be used to minimize the loss function L to adjust the parameters of the ETS and GAT-GRU models.

[0291] Through this weighted loss function, we can effectively combine the ETS model's ability to capture trends and seasonality in time series data with the GAT-GRU model's advantages in handling complex long-term dependencies, achieving more accurate and robust anomaly monitoring performance. This not only improves the accuracy of predictions, but also brings greater flexibility and efficiency to the practical application of the system.

[0292] 5. Robot regional collaboration and optimization module

[0293] The core goal of the robot regional collaboration and optimization module is to achieve efficient collaborative learning and dynamic optimization between regions within the robot cluster to enhance the accuracy and response speed of overall abnormality monitoring. In this embodiment, the robot regional collaboration and optimization module consists of the following key steps, including similarity calculation, collaborative learning mechanism, application of multi-objective optimization algorithm, dynamic parameter update and adaptive feedback control mechanism. The specific process is as follows: Figure 12In this embodiment, the robot regional collaboration and optimization module identifies the most similar region after detecting an anomaly, shares anomaly data, and implements cross-region collaborative learning and parameter optimization.

[0294] S21: Similarity calculation:

[0295] When a robot in a certain area detects an anomaly, the anomaly data will be uploaded to the cloud and confirmed by the equipment engineer as a valid anomaly. The confirmed anomaly data will be used to identify other similar areas through similarity calculation to trigger the collaborative learning process. The similarity uses the cosine similarity method to quantify the feature similarity of different areas. Specifically, suppose area A contains n robots, denoted as A1, A2, ...A n , area B contains m robots, denoted as B1, B2, ... B m ; then robot A i and Robot B j The cosine similarity is:

[0296]

[0297] Where: cos(θ ij ) is the i-th robot A i With the jth robot B i Similarity; A i,k and B j,k Robot A i and B i The value of the kth feature; d is the dimension of the feature vector. In order to obtain the overall similarity between region A and region B, the similarities of all robot pairs are averaged to obtain the overall similarity between region A and region B:

[0298]

[0299] Here, sim(A,B) is the average similarity between regions A and B; n and m represent the number of robots in regions A and B, respectively. When the average similarity sim(A,B) between regions is above a certain threshold, the system can assume that the two regions have a high degree of similarity in their operating characteristics.

[0300] S22: Collaborative learning mechanism: Determine whether the average similarity sim(A,B) between region A and region B is higher than the set similarity threshold: If so, there is similarity between region A and region B, and the confirmed abnormal data is broadcast from region A to region B to trigger collaborative learning in region B; if not, there is no similarity between region A and region B;

[0301] To further optimize the effects of collaborative learning and parameter updates, this embodiment employs an adaptive feedback control mechanism. This mechanism dynamically adjusts the collaborative learning similarity threshold by monitoring the actual performance of each regional model, thereby adaptively adjusting the collaborative learning range. Specifically, this embodiment employs PID control to update the similarity threshold. Specifically, the dynamic update method for the similarity threshold is as follows:

[0302]

[0303] Where: t is the similarity threshold at time step t; e t is the error of the current time step, that is, the deviation between the current false alarm rate or missed alarm rate and the set value; K p is the proportional coefficient, which is used to adjust the immediate impact of the error; K i is the integral coefficient, which is used to adjust the cumulative effect of the error; K d is the differential coefficient, which is used to adjust the rate of change of the error; is the cumulative error from the initial time step to the current time step.

[0304] By adaptively adjusting the similarity threshold δ, the system can dynamically adapt to different anomaly types and regional characteristics, making collaborative learning and anomaly monitoring more flexible and efficient.

[0305] S23: Multi-objective optimization: Based on regional collaborative learning, in order to strike a balance between monitoring accuracy and response time, the NSGA-II multi-objective optimization algorithm is applied to dynamically optimize the model parameters of each region. The objective function of the multi-objective optimization is:

[0306] min f(θ)=(f1(θ),f2(θ))

[0307] Where θ represents the set of model parameters to be optimized; f1(θ) is the accuracy target of the anomaly monitoring model, which is defined as the accuracy score, and the optimization goal is to maximize the accuracy; f2(θ) is the response time target, which is defined as the average detection time of the model, and the optimization goal is to minimize the response time.

[0308] The NSGA-II algorithm constructs a Pareto optimal solution set by comparing non-dominated sorting and congestion to achieve a balance between multiple objectives; the optimal parameter combination θ is selected from the Pareto optimal solution set. * Used to update the model and improve the overall performance of anomaly detection in the region.

[0309] S24: Dynamic parameter update: abnormal data obtained through collaborative learning and optimal parameters θ calculated by multi-objective optimization *, the anomaly monitoring model in each region will dynamically update its parameters. The loss function L(θ) is used to measure the prediction error of the model and is combined with the regularization term to prevent the model from overfitting:

[0310]

[0311] Where: L(θ) is the loss function used to evaluate the prediction accuracy of the model; N is the number of samples, that is, the number of training data points used; y i is the actual observation value of the i-th sample; is the predicted value of the model for the i-th sample under the parameter θ; λ is the regularization coefficient, which is used to control the impact of the regularization term on the loss and avoid overfitting of the model; Ω(θ) is the regularization term, which is used to penalize excessively large parameter values; in the process of minimizing the loss function L(θ), the regional anomaly monitoring module in each area will update the parameters to enhance the monitoring accuracy while maintaining the stability of the model.

[0312] In summary, the robot regional collaboration and optimization module of this embodiment achieves efficient collaborative learning and model optimization across regions through similarity calculation, collaborative learning, multi-objective optimization, and an adaptive feedback control mechanism. Cosine similarity is used to determine the similarity of regional features and determine the broadcast range for collaborative learning. The NSGA-II optimization algorithm ensures a balance between model accuracy and response time. The PID control mechanism dynamically adjusts the similarity threshold to enhance system adaptability. This module significantly enhances the collaborative capabilities and system reliability of the robot cluster while improving anomaly detection accuracy and response speed.

[0313] 6. Predictive maintenance module

[0314] The predictive maintenance module of this embodiment adopts a temporal convolutional network (TCN) model. The predictive maintenance module is used to predict the future operating status of the industrial robot cluster and issue fault warnings in advance based on the prediction results to optimize the maintenance plan. Figure 13 This module is specially designed with a primary and secondary alarm system to respond to predicted potential problems with different levels of urgency.

[0315] Specifically, the method for predicting the future operating status of the industrial robot cluster in the predictive maintenance module is as follows.

[0316] 6.1. Building a TCN Model

[0317] The TCN model is used to predict the future operational status of industrial robot clusters, generating a sequence of key parameters that characterize the robot states over a period of time. Specifically, the TCN consists of multiple layers of one-dimensional convolutional layers, each of which uses dilated convolutions to process temporally dependent sequence data. Residual connections are used to mitigate the vanishing gradient problem in deep networks, thereby improving the model's stability and accuracy in long-term series prediction. Multi-source data from robots in each region is fed into the TCN network as input features. The network automatically learns the temporal dynamics of the data, capturing long-term dependencies and providing precise support for predicting the future state of the robot cluster.

[0318] 6.2 Predictive Maintenance and Early Warning

[0319] 6.2.1 Fault Prediction

[0320] Prediction goal: Use the trained TCN model to predict the operating status of the equipment in the next few time steps, including the trend changes of key parameters such as temperature, vibration, pressure, etc.

[0321] Forecast results: The output forecast results include a series of equipment status parameters for a period of time in the future. These predicted values ​​are compared with the actual equipment status in operation to identify potential abnormal trends.

[0322] 6.2.2 Early Warning Mechanism

[0323] Level 1 Warning (Warning): When the predicted value is greater than or equal to the level 1 warning threshold and less than the level 2 warning threshold, the system issues a level 1 warning signal to remind maintenance personnel to pay attention to the equipment status, but no immediate action is required. Level 2 Warning (Critical Alert): When the predicted value is greater than or equal to the level 2 warning threshold and the trend is clearly deteriorating, the system issues a level 2 warning signal, requiring immediate inspection or maintenance operations. That is, the generation of the warning signal is based on the comparison between the predicted value and the threshold. is the predicted value, T1 and T2 are the thresholds of the first-level warning and the second-level warning respectively. The warning condition is: if It is a first-level warning; if the actual observation value ≥T2, it is a level 2 warning.

[0324] Multivariate early warning comprehensive analysis: When processing multi-source equipment data, it may involve early warnings for multiple monitoring parameters. This embodiment can also use a comprehensive scoring method to combine the early warning signals of each parameter into a global early warning signal, and define the global early warning index as:

[0325]

[0326] Where: iis the importance weight of each parameter; is the predicted value of the i-th parameter; It is the set second-level warning threshold; when the global warning index CWI exceeds the set global warning threshold, a global warning signal is generated.

[0327] By applying TCN, factories can achieve precise predictive maintenance in complex production environments with robot clusters, providing more robust assurance for industrial production. This approach not only improves system security and stability, but also enhances the factory's ability to cope with complex and changing production environments, effectively reducing unplanned downtime and the resulting economic losses.

[0328] The intelligent monitoring and predictive maintenance system for the operation status of industrial robot clusters in this embodiment includes multiple subsystems and a cloud-based main system. Through collaborative work, it realizes real-time monitoring, collaborative learning, and predictive maintenance of cluster equipment. In the system architecture design, the subsystem is responsible for the collection, monitoring, and processing of robot data within the area. The subsystem contains five core modules: data preprocessing, multi-source data fusion and alignment, regional anomaly monitoring, robot regional collaboration and optimization, and predictive maintenance module. The five module processes of the subsystem are as follows: Figure 14 As shown in the figure, the synergy of these modules ensures a smooth flow of data from collection and processing to anomaly detection and predictive maintenance. This subsystem integrates advanced data processing technologies, deep learning models, and time series analysis methods to achieve efficient equipment status monitoring and potential fault prediction. The robot cluster intelligent monitoring and maintenance system of this embodiment not only improves the timeliness and accuracy of maintenance, but also makes maintenance decisions more scientific and systematic through an intelligent early warning mechanism, significantly reducing the risk of unplanned downtime, improving production efficiency, and enhancing the operational stability of industrial robot clusters and the productivity of enterprises.

[0329] The above-described embodiments are merely preferred embodiments for the purpose of fully illustrating the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are within the scope of protection of the present invention. The scope of protection of the present invention shall be subject to the claims.

Claims

1. An intelligent monitoring and predictive maintenance system for the operating status of an industrial robot cluster, characterized by: It includes a cloud main system and multiple subsystems, wherein the subsystems are set up corresponding to the regional robot clusters; The cloud main system includes a core platform layer, a communication layer, a logic layer and a storage layer; The core platform layer is used to process computing tasks, coordinate data analysis and collaborative learning of each subsystem, and support efficient computing and task scheduling of the system; The communication layer is used to realize data transmission between the cloud main system and subsystems, ensuring real-time synchronization and reliable transmission of data; The logic layer is used to receive abnormal data, similarity thresholds, and similar area information from the subsystems, confirm the abnormal data, and further confirm similar areas based on the similarity thresholds. It is also responsible for coordinating and initiating cross-region collaborative learning to promote collaborative optimization between similar areas. The storage layer is used to store abnormal data, historical monitoring data, similarity thresholds, similar area information and confirmed abnormal events from each subsystem for subsequent query, analysis and model update; The subsystem includes a regional communication layer, a regional logic layer and a regional storage layer. The regional communication layer is used to realize data transmission within the subsystem, between subsystems, and between the subsystem and the cloud host system; The regional logic layer is equipped with a data preprocessing module, a multi-source data fusion and alignment module, a regional anomaly monitoring module, a robot regional collaboration and optimization module, and a predictive maintenance module; the data preprocessing module is used to fill missing values ​​to ensure data integrity, and perform data cleaning, outlier processing, and data normalization on the missing data; the multi-source data fusion and alignment module is used to align and normalize data from different sources; and the regional anomaly monitoring module is used to capture abnormal behavior. The robot regional collaboration and optimization module is used to identify the most similar areas after detecting an anomaly, share anomaly data, and achieve cross-region collaborative learning and parameter optimization; the predictive maintenance module is used to predict future equipment status to issue early warnings; The regional storage layer is used to store robot sensor data; The method for filling missing values ​​in the data preprocessing module is: in: is the missing value to be filled; is the filling value obtained by Lagrange interpolation method; The filling value of the valid data point obtained by the adjacent data filling method; is the filling value of the valid data point obtained by the GraphSAGE filling method; and is the weight, and: ; The Lagrange interpolation method is: in: For the time point Supplementary values ​​using Lagrangian filling strategy; 、 and The first three time points of the selected missing region are 、 and The corresponding data value; The adjacent data filling method is: in: yes The supplementary value obtained by the adjacent data filling strategy at the time point; is the central position time point in the missing area data segment; Indicates the length of the extracted data segment; The GraphSAGE filling method is: in: is through predict The supplementary value of the moment; It is layer Feature representation of time points; The method for the data preprocessing module to process outliers in the data after missing filling is to calculate the first quartile of the data. and the third quartile ,Right now: Calculate the outlier bounds based on the IQR: or All points that fall outside this range are considered outliers. If the outliers are considered to be caused by errors or accidental large deviations and are not numerous, they are directly deleted from the dataset. If there are a certain number of outliers, the outliers are replaced by the average value of the entire data. The data preprocessing module performs data standardization on the data after missing fill by converting the data into a form with zero mean and unit variance to eliminate the influence between data of different magnitudes. The formula is as follows: in: The data obtained by data standardization; is the data after missing fill; and are the mean and standard deviation of the data respectively.

2. The intelligent monitoring and predictive maintenance system for the operating status of an industrial robot cluster according to claim 1 is characterized by: The multi-source data fusion and alignment module uses a dynamic time warping method to align the time series of multi-source data to eliminate the dimension, frequency and format differences of different robot data in the region. The method steps are as follows: S11: Construct time series data: Organize the preprocessed data into a time series format. Each time series contains a series of data points, and each data point corresponds to a timestamp. S12: Define reference template and test template: select one time series as the reference template and other time series as the test templates, and align them with the reference template; S13: Calculate the distance matrix: Construct a A matrix grid, where and are the lengths of the reference template and the test template respectively; each element of the matrix Indicates the reference template point and the test template The distance between points : in: For the reference template The horizontal coordinate of the point; For the test template The vertical coordinate of a point; S14: DTM method: Given two lengths and Time series and : in: Representing time series Middle The value of the data point; Representing time series Middle The value of the data point; Using Match Path To express and Alignment between Expressed as: in: represent The Values ​​match The values; Indicates the Step matching path; The goal of the DTW method is to find the optimal matching path Make the cumulative distance Minimum is achieved under the following three constraints: (1) Boundary constraint: This constraint restricts the path to start at (1,1) and end at (1,1) when aligning the sequence. ,Right now: (2) Continuity constraint: This constraint restricts the temporally changing curved paths when aligning sequences. Under this condition, the path must move forward one step at a time, i.e.: and ; (3) Monotonicity constraint: This condition preserves the temporal order of the points and restricts the path to move forward without decreasing, i.e.: and ; Assumptions Represents the cumulative distance, then the global minimum cumulative distance can be obtained through the following dynamic programming: Among them: initial value ; Two sequences and The DTW distance is expressed as , Equivalent to ,The optimal alignment path can be obtained by backtracking in the dynamic ,programming process.

3. The intelligent monitoring and predictive maintenance system for the operating status of an industrial robot cluster according to claim 1 is characterized by: The regional anomaly monitoring module combines the ETS model and the GAT-GRU model to comprehensively analyze the operating status of the industrial robot cluster, accurately detect and predict abnormal behaviors, and use an optimized fusion loss function to efficiently identify and warn of potential faults, providing reliable support for maintenance decisions and the stable operation of the production line. The GAT-GRU model includes a graph attention network (GAT) and a gated recurrent unit (GRU). The GAT dynamically calculates the importance weights between robot features and time through an attention mechanism. Combining a dual-channel design of feature GAT and time GAT, it captures the complex spatial and temporal dependencies of industrial robot cluster data. The gated recurrent unit GRU takes the graph embedding vector generated by the graph attention network GAT as input, and the gated recurrent unit GRU network dynamically adjusts the hidden state by updating the gate and resetting the gate based on the time series modeling process.

4. The intelligent monitoring and predictive maintenance system for the operating status of an industrial robot cluster according to claim 3 is characterized by: The ETS model predicts future data points using exponential smoothing. The principle is: Smoothing equation: in: is the predicted value for the next time step; is the current observation value; is the current forecast value; is the trend item; is the smoothing coefficient, and 0< <1; Trend Adjustment: in: is the trend component of the next time step; is the trend smoothing parameter; Seasonal Adjustment: in: yes seasonal component after the step; is the seasonal smoothing parameter; It's the seasonal cycle.

5. The intelligent monitoring and predictive maintenance system for the operating status of an industrial robot cluster according to claim 3 is characterized by: The method used by the graph attention network GAT to capture the complex spatial and temporal dependencies of industrial robot cluster data is: Based on feature GAT: the key parameter feature numbers of the robot are As a node, the length of the time series intercepted by the sliding window function As the number of features in each node, calculate the attention weight: in, For feature-based GAT nodes and adjacent nodes The attention coefficient between is a nonlinear activation function; is the weight matrix based on feature GAT; is the learned weight vector; Indicates a connection operation; and Respectively represent the nodes in the feature GAT and nodes The feature vector of the node and nodes All key parameter characteristics; use The function normalizes the attention weights of features within a group: in: It is based on the node in the feature GAT Its adjacent nodes Normalized attention weights between ; It is based on the node in the feature GAT The set of adjacent nodes of Use attention weights to perform weighted averaging on features and update feature representation: ) in: It is based on other nodes in the feature GAT. the overall impact of Time-based GAT: The length of each sliding window function intercepted As a node, the key parameter characteristic number as features in each node and adjust the learning parameters to , the weight matrix is ​​adjusted to ;pass The function performs weight normalization calculations and uses attention weights to update node representations: ) in: It is based on the other nodes in the time GAT to the node the overall impact of For nodes in time-based GAT Its adjacent nodes Normalized attention weights between ; For nodes in time-based GAT The feature vector of the node All key parameter characteristics; Combining the outputs of parallel feature-based GAT and the output of the time-based GAT , fused into the output of the graph attention network GAT .

6. The intelligent monitoring and predictive maintenance system for the operating status of an industrial robot cluster according to claim 3, characterized in that: The principle of the gated recurrent unit GRU is: Update Gate: in: is in the time step The update gate controls the previous hidden state; is the sigmoid function; is the weight matrix of the update gate; is the hidden state at the previous time step; is the input of the current time step; is the bias term; Reset the gate: in: It is the reset gate that determines how much information of the previous state is forgotten; Candidate hidden states: in: is a candidate hidden state; is the weight matrix of the candidate hidden state; is the bias term; represents the element-wise product, meaning that the output of the reset gate affects which parts of the hidden state should be retained; Final hidden state: in: is the current time step The hidden state of is used for prediction or further processing; Prediction output: final hidden state of GRU It will be converted into the final output prediction through a fully connected layer : in: and are the weights and biases of the output layer.

7. The intelligent monitoring and predictive maintenance system for the operating status of an industrial robot cluster according to claim 3, characterized in that: Set the prediction output of the ETS model to , the predicted output of the GAT-GRU model is , the true observation value is , then the fusion loss function is defined as: in: is a weight parameter, and 0< <1, used to control the contribution ratio of the ETS model and the GAT-GRU model to the total loss; and are the prediction errors of the ETS model and the GAT-GRU model, respectively, expressed as: in: is the actual value; is the predicted value calculated using the statistical model ETS; is the total number of samples; in: is the predicted value obtained using the GAT-GRU model.

8. The intelligent monitoring and predictive maintenance system for the operating status of an industrial robot cluster according to claim 1, characterized in that: The robot regional collaboration and optimization module identifies the most similar region after detecting an anomaly, shares anomaly data, and implements cross-region collaborative learning and parameter optimization in the following steps: S21: Similarity Calculation: Hypothesis Region Include robots, respectively ,area Include robots, respectively ; then the robot and robots The cosine similarity is: in: For the robots With the robots similarity; and Robots and In the The value of a feature; is the dimension of the feature vector; Average the similarities of all robot pairs to get the region and region The overall similarity of: in, For the region and region The average similarity between and Represents the area and region The number of robots in S22: Collaborative learning mechanism: judgment area and region The average similarity between Is it higher than the set similarity threshold: If so, then the region and region There is similarity between them, and the confirmed abnormal data is removed from the region Broadcast to Region , to trigger the area collaborative learning; if not, the region and region There is no similarity between them; The dynamic update method of the similarity threshold is: in: is the time step The similarity threshold when ; is the error of the current time step, that is, the deviation between the current false alarm rate or missed alarm rate and the set value; is the proportional coefficient, which is used to adjust the immediate impact of the error; is the integral coefficient, which is used to adjust the cumulative effect of the error; is the differential coefficient, which is used to adjust the rate of change of the error; is the cumulative error from the initial time step to the current time step; S23: Multi-objective optimization: Apply the NSGA-II multi-objective optimization algorithm to dynamically optimize the model parameters of each region. The objective function of the multi-objective optimization is: in, Represents the set of model parameters to be optimized; is the accuracy target of the anomaly monitoring model, and the optimization goal is to maximize the accuracy; For the response time objective, the optimization goal is to minimize the response time; Construct a Pareto optimal solution set to achieve a balance between multiple objectives; select the optimal parameter combination in the Pareto optimal solution set Used to update the model; S24: Dynamic Parameter Update: Loss Function Used to measure the prediction error of the model and combined with the regularization term to prevent the model from overfitting: in: is the loss function, which is used to evaluate the prediction accuracy of the model; is the number of samples, that is, the number of training data points used; For the The actual observed value of the sample; For the model in parameters Next pair The predicted value of samples; is the regularization coefficient, which is used to control the impact of the regularization term on the loss and avoid overfitting of the model; is a regular term used to penalize excessively large parameter values; in the loss function During the minimization process, the regional anomaly monitoring module in each region will update parameters to enhance monitoring accuracy while maintaining the stability of the model.

9. The intelligent monitoring and predictive maintenance system for the operating status of an industrial robot cluster according to claim 1, characterized in that: The method for predicting the future operating status of the industrial robot cluster in the predictive maintenance module is: Construct a TCN model and use it to predict the future operating status of an industrial robot cluster, obtaining a sequence of key parameters that can characterize the robot status over a period of time in the future. Based on historical data and expert experience, the warning threshold range for each key parameter is set, and the warning threshold is divided into two levels: the first warning threshold and the second warning threshold; If the predicted value of the key parameter is greater than the set first-level warning threshold, a first-level warning is triggered; if the predicted value of the key parameter is between the first-level warning threshold and the second-level warning threshold, a second-level warning is triggered; or, Using the comprehensive scoring method, the warning signals of each parameter are integrated into a global warning signal, and the global warning index is defined as: in: is the importance weight of each parameter; For the The predicted value of the parameter; The second level warning threshold is set; When the global warning index CWI exceeds the set global warning threshold, a global warning signal is generated.

Images