Intelligent detection method for abnormal working state of paint spraying mechanical arm based on time series diagram neural network technology

Abstract

The invention relates to the technical field of industrial manufacturing, in particular to a paint spraying mechanical arm working state abnormity intelligent detection method based on a time sequence graph neural network technology, which comprises the following steps of S1, modeling of multi-sensor fused time sequence data, S2, abnormity detection and time sequence learning based on a graph neural network, S3, multi-layer embedding and feature representation, and S4, multi-layer embedding and feature representation. S4, a real-time monitoring and abnormity early warning system, S5, data analysis and calculation, and S6, result and output: through processing of a graph neural network, the system can learn abnormal features in time sequence signals, and an output layer generates a detection result according to embedded features and time sequence information. The multi-dimensional data of the mechanical arm are collected through the multi-sensor fusion technology, and the time sequence data of the sensors are processed in combination with the graph neural network model, so that the dependency relationship and the time sequence characteristics among the sensors can be effectively captured, the working state of the mechanical arm is monitored in real time, and anomaly detection and fault early warning are realized.

Description

Technical Field

[0001] The present invention relates to the field of industrial manufacturing technology, and in particular to a method for intelligently detecting abnormal working status of a paint spraying robot arm based on time series graph neural network technology. Background Art

[0002] In the modern industrial manufacturing process, paint spraying robots are widely used in the automated spraying process of the automotive, aviation, home appliance and other industries. The quality of paint spraying directly affects the appearance, durability and protective effect of the product. Therefore, it is very important to ensure the stability of the robot's spraying process and the spraying quality. However, due to the complex production environment, the robot may have unstable state, equipment aging or environmental changes during long-term operation. These will lead to abnormal problems in the spraying process, such as uneven spraying, particle contamination, inconsistent coating thickness, etc. Therefore, it is particularly necessary to develop a real-time and accurate anomaly detection method to monitor the state and predict faults of the spraying robot.

[0003] Current paint spraying robot arm status detection mostly relies on regular maintenance, manual inspection or data analysis based on a single sensor. However, these methods have many shortcomings:

[0004] 1. Limitations of manual inspection: Manual inspection has problems such as strong subjectivity and low efficiency. Especially in large-scale automated production, manual inspection is difficult to meet the needs of real-time monitoring.

[0005] 2. Limitations of single sensor detection: Single sensor data can only capture a certain aspect of the robot arm's operation (such as pressure, speed, etc.), while abnormal conditions in the paint spraying process are often multi-dimensional and complex phenomena caused by the interaction of multiple factors. Therefore, data based on a single sensor cannot fully reflect the working status of the robot arm.

[0006] 3. Ignoring the timing dependencies: Traditional methods often ignore the timing dependencies existing in time series signals, such as how changes in force and torque affect the vibration and stability of the robotic arm, and how temperature and pressure jointly affect the spraying effect during the spraying process.

[0007] In order to overcome the above shortcomings, we proposed an intelligent detection method for abnormal working status of paint spraying robot arm based on time series graph neural network technology to solve the above problems. Summary of the invention

[0008] In view of the shortcomings of the prior art, the purpose of the present invention is to propose an intelligent detection method for the abnormal working state of a paint spraying robot arm based on time series graph neural network technology, which has the advantages of effectively capturing the dependencies and timing characteristics between sensors, monitoring the working state of the robot arm in real time, and realizing abnormality detection and fault warning, thereby solving the problems raised in the above-mentioned background technology.

[0009] To achieve the above object, the present invention provides the following technical solutions:

[0010] An intelligent detection method for abnormal working state of a paint spraying robot arm based on time series graph neural network technology comprises the following steps:

[0011] S1. Multi-sensor fusion time series data modeling: By fusing the data of multiple sensors, such as position sensor, 6-axis force / torque sensor, speed sensor, spray gun pressure sensor, electrostatic voltage sensor, etc., the time series signals collected by these sensors are multi-dimensionally modeled;

[0012] S2. Anomaly detection and time series learning based on graph neural network: Introduce time series graph neural network (GNN) into robot arm state monitoring, and use the powerful spatiotemporal dependency modeling ability of graph neural network to jointly learn the time series dependency and spatial relationship of signals from different sensors;

[0013] S3. Multi-layer embedding and feature representation: Using feature embedding technology, multi-dimensional sensor time series signals are embedded into high-dimensional space so that they can be processed by graph neural networks;

[0014] S4. Real-time monitoring and abnormal warning system: Through real-time acquisition and processing of robot arm sensor signals, real-time monitoring of the working status is achieved. Through the abnormal detection module of the time series graph neural network, the system can quickly determine abnormal phenomena in the working status (such as excessive torque, excessive vibration, abnormal temperature, etc.). If an abnormal situation occurs, an early warning signal will be issued, where the abnormal situation includes abnormal warning data, which can prevent equipment failure or spraying quality problems in time;

[0015] S5. Data analysis and calculation: The central processor is used to control the data analysis and calculation module to analyze and detect the monitored abnormal warning data, and analyze the causes of abnormal working status and abnormal data points, as follows:

[0016] (1) Execute the Windows menu command, Start | All Applications Anaconda3 (64-bit) | JupyterNotebook, start the local server, and open the homepage in the default browser;

[0017] (2) Create a new Notebook file, click the "New" drop-down menu, select Python 3, and create a new Notebook named Untitled;

[0018] (3) Enter Markdown text. In the first cell input box, enter "Use IPython Notebook for scientific computing" and set the cell type to Markdown.

[0019] (4) Enter and execute Python code. Use the menu command Insert|Insert Cell Below or the shortcut key B to insert a cell below, enter the code, and press the shortcut key Ctrl+Enter to execute and display the results;

[0020] S6. Results and Output: Through the processing of graph neural networks, the system can learn the complex relationships between various sensors, especially the abnormal features in the timing signals. Finally, the output layer generates detection results based on the embedded features and timing information, and predicts the state of the robotic arm.

[0021] As a preferred technical solution, the time series data modeling of multi-sensor fusion protects the technology of constructing a graph structure through the time series data of multiple sensors, especially the fusion modeling of real-time multi-dimensional data of the robotic arm, which reflects the dynamic changes and status of the robotic arm during the spraying process through time series analysis and graph structure construction.

[0022] As a preferred technical solution, the anomaly detection and time series learning based on graph neural network protects the algorithm for anomaly detection of the working state of the robotic arm based on the time series graph neural network (GNN), especially the learning model that aggregates the features of multiple sensor nodes and time series dependencies through graph convolution. By performing multi-layer graph convolution operations on the features of sensor nodes, the system can gradually aggregate and learn the working state characteristics of the robotic arm, thereby realizing the detection of abnormal states.

[0023] As a preferred technical solution, the embedding layer can extract high-level features in the time series, retain key information of the robotic arm during dynamic work, further improve the accuracy of anomaly detection, and protect the multi-layer embedding method of sensor data, especially in the feature encoding, normalization, and embedding of time series data into high-dimensional space, ensuring the data's expressiveness and subsequent learning effects.

[0024] As a preferred technical solution, the method protects the real-time monitoring and abnormal warning system, especially a technical solution for capturing abnormal changes in sensor data and responding to them during the dynamic operation of the robot arm.

[0025] As a preferred technical solution, the sensor input signal and variable number are specifically as follows:

[0026] (1) Position sensor: monitors the rotation angle (P1) and displacement (P2) of the robot arm to ensure accurate positioning of each joint of the robot arm;

[0027] (2) 6-axis force / torque sensor: It can measure force (Fx, Fy, Fz) and torque (Tx, Ty, Tz) on three spatial axes, helping to monitor force feedback during the spraying process and ensure the flexibility and stability of the robot arm movement;

[0028] (3) Speed ​​sensor: integrated with the position sensor to measure the speed (V) of the robot arm or spray head in real time;

[0029] (4) Spray gun pressure sensor: monitors the pressure (PS) of gas or liquid during spraying to ensure the consistency of paint spraying speed and amount;

[0030] (5). Gas flow sensor: used to monitor the flow rate (FS) of compressed air to ensure the stability of gas-assisted spraying process;

[0031] (6) Paint pipeline temperature sensor: monitors the temperature of the paint in the pipeline (TS) to prevent excessively high or low temperatures from affecting the fluidity of the paint;

[0032] (7) Ultrasonic sensor: Use ultrasonic waves to detect the distance (DS) between the spray head and the workpiece to ensure the optimal spray distance in a dynamic environment;

[0033] (8) Vibration sensor: used to monitor the stability of the robot arm to ensure that there is no abnormal vibration during the movement process (VS);

[0034] (9) Electrostatic voltage sensor: During the electrostatic spraying process, it is used to monitor the voltage (EV) of the workpiece and the spray gun to ensure that the electrostatic field strength is sufficient to ensure that the paint can be evenly adhered to the workpiece surface.

[0035] As a preferred technical solution, the encoding method of the time signal is as follows:

[0036] The time series signals from multiple sensors are first processed through the encoding layer to extract the key information of the time series data and convert it into features suitable for graphical model processing. The sensor signal X = {x1(t), x2(t), ...xn(t)}, where each sensor xi(t) is a time series function;

[0037] (1) Sliding window and feature extraction;

[0038] For each time series signal xi(t), the sliding window technique is used to segment the signal. The size of the sliding window is set to w, and the signal in each window can be expressed as:

[0039]

[0040] in, Represents the signal data in the kth window;

[0041] Within each sliding window, we can extract statistical features from the time series, which may include:

[0042] Mean:

[0043] Standard Deviation:

[0044] Maximum value and minimum value: reflect the extreme value information within the window;

[0045] Trend coefficient: Calculate the trend of the time series through linear regression analysis;

[0046] All these features are combined to form a vector As the feature representation of the signal xi(t) in window k:

[0047]

[0048] (2) Normalization and standardization;

[0049] In order to ensure that the features of different sensors have the same dimension and range, we need to normalize or standardize the extracted features. The normalization formula is:

[0050]

[0051] in, is the normalized eigenvalue, is the eigenvalue, min(fi) and max(fi) are the minimum and maximum values ​​of the feature fi respectively; after feature extraction and normalization, each sensor signal xi(t) can be converted into a feature vector Used to indicate the state at each moment.

[0052] As a preferred technical solution, the embedding method of the time signal is as follows: the embedding method of the time signal is used to map the encoded feature vector into a high-dimensional space, so that the graph neural network can perform feature aggregation and learning in this space;

[0053] (1) Embedding Layer;

[0054] Encoded features Will be mapped to a high-dimensional embedding space; the role of the embedding layer is to project the low-dimensional feature vector to the high-dimensional feature space through a learnable weight matrix We; the mathematical expression of the embedding process is:

[0055]

[0056] in:

[0057] We is the weight matrix of the embedding layer, which represents the linear transformation of the features.

[0058] be is a bias term used to adjust the embedded features.

[0059] is the high-dimensional feature vector embedded in window k by sensor i;

[0060] (2) Temporal dependency of time series embedding;

[0061] In order to preserve the temporal dependencies in time series signals, a time series embedding method based on recurrent neural networks is used; this method embeds the features It is further input into a time recursive network to extract time-dependent high-level features;

[0062] The hidden state update formula of LSTM is as follows:

[0063]

[0064] in:

[0065] It is the hidden state at the kth moment, which contains the information of the current window k and the information of the previous moment k-1. The LSTM unit retains the key information in the past window through the memory gating mechanism and learns the time-dependent features;

[0066] (3) Embedding processing of graph neural networks;

[0067] For each time window k, after processing through the encoding and embedding layers, all sensor features can form a feature matrix Z, where each row represents the embedded features of a sensor in a specific time window. Combined with the adjacency matrix A, a graph structure G = (A, Z) is generated as the input of the graph neural network.

[0068] As a preferred technical solution, the application of overall embedding and graph neural network, after the above steps, the encoded time series signal will be converted into a high-dimensional feature representation and input into the graph neural network for further processing; the graph neural network (GNN) will learn the relationship between sensor nodes and time series features through multi-layer graph convolution operations to detect abnormal situations. The formula of the graph convolution layer is as follows:

[0069]

[0070] in:

[0071] H(l) represents the node features of the lth layer,

[0072] is the adjacency matrix of the graph, which is used to represent the dependency relationship between sensor nodes.

[0073] W(l) is the weight matrix of layer l;

[0074] σ(·) is the activation function.

[0075] As a preferred technical solution, the abnormal detection results can be classified by the following formula:

[0076] y i =softmax(W o Z i +b o );

[0077] in:

[0078] Zi is the embedding feature of node i,

[0079] Wo and bo are the weights and biases of the output layer, used for classification;

[0080] yi is the abnormal state prediction of node i.

[0081] Compared with the prior art, the present invention has the following beneficial effects:

[0082] 1. The present invention collects multi-dimensional data of the robotic arm through multi-sensor fusion technology, and processes the time series data of the sensor in combination with the graph neural network model. It can effectively capture the dependency and time series characteristics between sensors, monitor the working status of the robotic arm in real time, and realize anomaly detection and fault warning.

[0083] 2. The present invention integrates the real-time data of multiple sensors, constructs a graph structure between time series signals, and utilizes the spatiotemporal convolution capability of the graph neural network model to effectively capture the dependencies between sensors. It can also timely detect abnormal conditions of the robotic arm during operation by aggregating and learning signal features. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0085] Figure 1 It is the technical roadmap of the method of the present invention. DETAILED DESCRIPTION

[0086] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0087] According to the attached Figure 1 As shown, the embodiment of the present invention provides an intelligent detection method for abnormal working state of a paint spraying robot arm based on time series graph neural network technology, comprising the following steps:

[0088] S1. Multi-sensor fusion time series data modeling: By fusing the data of multiple sensors, such as position sensor, 6-axis force / torque sensor, speed sensor, spray gun pressure sensor, electrostatic voltage sensor, etc., the time series signals collected by these sensors are multi-dimensionally modeled;

[0089] S2. Anomaly detection and time series learning based on graph neural network: Introduce time series graph neural network (GNN) into robot arm state monitoring, and use the powerful spatiotemporal dependency modeling ability of graph neural network to jointly learn the time series dependency and spatial relationship of signals from different sensors;

[0090] S3. Multi-layer embedding and feature representation: Using feature embedding technology, multi-dimensional sensor time series signals are embedded into high-dimensional space so that they can be processed by graph neural networks;

[0091] S4. Real-time monitoring and abnormal warning system: Through real-time acquisition and processing of robot arm sensor signals, real-time monitoring of the working status is achieved. Through the abnormal detection module of the time series graph neural network, the system can quickly determine abnormal phenomena in the working status (such as excessive torque, excessive vibration, abnormal temperature, etc.), and issue a warning signal to prevent equipment failure or spraying quality problems in time;

[0092] S5. Data analysis and calculation: The central processor is used to control the data analysis and calculation module to analyze and detect the monitored abnormal warning data, and analyze the causes of abnormal working status and abnormal data points, as follows:

[0093] (1) Execute the Windows menu command, Start | All Applications Anaconda3 (64-bit) | JupyterNotebook, start the local server, and open the homepage in the default browser;

[0094] (2) Create a new Notebook file, click the "New" drop-down menu, select Python 3, and create a new Notebook named Untitled;

[0095] (3) Enter Markdown text. In the first cell input box, enter "Use IPython Notebook for scientific computing" and set the cell type to Markdown.

[0096] (4) Enter and execute Python code. Use the menu command Insert|Insert Cell Below or the shortcut key B to insert a cell below, enter the code, and press the shortcut key Ctrl+Enter to execute and display the results;

[0097] In the data analysis and calculation formula, the following steps are required:

[0098] Examples of creating vectors and matrices (arrays)

[0099] In [2]: %pylab inline

[0100] A=np.array([[1,2,3],[4,5,6]])

[0101] print(A)

[0102] ```

[0103] Populating the interactive namespace from numpy and matplotlib [

[123] [

[456] ]

[0106] In[3]: I = np.eye(3)

[0107] print(I) [[1.0.0.] [0.1.0.] [0.0.1]]

[0111] Matrix operations

[0112] In[4]: C=np.array([[1,2,3],[4,5,6],[7,8,9]])

[0113] D=np.array([[5,4,3],[2,7,8],[0,1,2]])

[0114] C+D

[0115] Out[4]: array([[6, 6, 6],

[0116] [6, 12, 14],

[0117] [7, 9, 11])

[0118] In[5]: C+DOut[5]: array([[6, 6, 6]

[0119] [6, 12, 14]

[0120] [7, 9, 11])

[0121] In[6]: np.dot(C, D) #matrix multiplication

[0122] Out[6]: array([[9, 21, 25],

[0123] [30, 57.64]

[0124] [51, 93, 103]])

[0125] HIn[7]:Transpose of CT##C

[0126] Out[7]: array([[1, 4, 7],

[0127] [2, 5, 8],

[0128] [3, 6, 9]])

[0129] Solving a linear equation system Given a linear equation system AX=b, then X=A -1 b. For example:

[0130] x·+·3y·+·5z=-10

[0131] 2x·1·5y·r=-8

[0132] 2x·+·3y·+·8z·=3.

[0133] but:

[0134] The solution process is as follows:

[0135] In

[10] : A=[[1,3,5],[2,5,1],[2,3,8]]

[0136] b = [

[10] , [8], [3]]

[0137] np.dot(np.linalg.inv(A), b) #inverse of coefficient matrix multiplied by constant matrix

[0138] Out

[10] : array([[-9.28],

[0139] [5.16], [0.76]])

[0141] In[11:np,linalg.solve(A,b)#directly use np.linalg.solve0 function to solve

[0142] Out

[11] :array([[-9.28], [5.16]. [0.76]])

[0145] S6. Results and Output: Through the processing of graph neural networks, the system can learn the complex relationships between various sensors, especially the abnormal features in the timing signals. Finally, the output layer generates detection results based on the embedded features and timing information, and predicts the state of the robotic arm.

[0146] As a preferred embodiment of this invention, time series data modeling of multi-sensor fusion protects the technology of constructing graph structure through time series data of multiple sensors, especially the fusion modeling of real-time multi-dimensional data of the robotic arm. These signals reflect the dynamic changes and status of the robotic arm during the spraying process through time series analysis and graph structure construction.

[0147] As a preferred embodiment of this invention, anomaly detection and time series learning based on graph neural network is used to protect the algorithm for anomaly detection of the working state of the robotic arm based on time series graph neural network (GNN), especially the learning model for aggregating features of multiple sensor nodes and time series dependency through graph convolution. By performing multi-layer graph convolution operations on the features of sensor nodes, the system can gradually aggregate and learn the working state features of the robotic arm, thereby realizing the detection of abnormal states.

[0148] As a preferred embodiment of this invention, the embedding layer can extract high-level features in the time series, retain key information of the robot arm in dynamic work, further improve the accuracy of anomaly detection, and protect the multi-layer embedding method of sensor data, especially in the feature encoding, normalization and embedding of time series data into high-dimensional space, ensuring the data's expressiveness and subsequent learning effect.

[0149] As a preferred embodiment of the present invention, the method protects the real-time monitoring and abnormal warning system, especially the technical solution that can timely capture abnormal changes in sensor data and respond to them during the dynamic operation of the robot arm.

[0150] As a preferred embodiment of this invention, the sensor input signal and the variable number are as follows:

[0151] (1) Position sensor: monitors the rotation angle (P1) and displacement (P2) of the robot arm to ensure accurate positioning of each joint of the robot arm;

[0152] (2) 6-axis force / torque sensor: It can measure force (Fx, Fy, Fz) and torque (Tx, Ty, Tz) on three spatial axes, helping to monitor force feedback during the spraying process and ensure the flexibility and stability of the robot arm movement;

[0153] (3) Speed ​​sensor: integrated with the position sensor to measure the speed (V) of the robot arm or spray head in real time;

[0154] (4) Spray gun pressure sensor: monitors the pressure (PS) of gas or liquid during spraying to ensure the consistency of paint spraying speed and amount;

[0155] (5). Gas flow sensor: used to monitor the flow rate (FS) of compressed air to ensure the stability of gas-assisted spraying process;

[0156] (6) Paint pipeline temperature sensor: monitors the temperature of the paint in the pipeline (TS) to prevent excessively high or low temperatures from affecting the fluidity of the paint;

[0157] (7) Ultrasonic sensor: Use ultrasonic waves to detect the distance (DS) between the spray head and the workpiece to ensure the optimal spray distance in a dynamic environment;

[0158] (8) Vibration sensor: used to monitor the stability of the robot arm to ensure that there is no abnormal vibration during the movement process (VS);

[0159] (9) Electrostatic voltage sensor: During the electrostatic spraying process, it is used to monitor the voltage (EV) of the workpiece and the spray gun to ensure that the electrostatic field strength is sufficient to ensure that the paint can be evenly adhered to the workpiece surface.

[0160] As a preferred embodiment of this embodiment, the encoding method of the time signal is as follows:

[0161] The time series signals from multiple sensors are first processed through the encoding layer to extract the key information of the time series data and convert it into features suitable for graphical model processing. The sensor signal X = {x1(t), x2(t), ...xn(t)}, where each sensor xi(t) is a time series function;

[0162] (1) Sliding Window and Feature Extraction

[0163] For each time series signal xi(t), we can use the sliding window technique to segment the signal;

[0164] The size of the sliding window is set to w, and the signal in each window can be expressed as:

[0165]

[0166] in, Represents the signal data in the kth window;

[0167] Within each sliding window, we can extract statistical features from the time series, which may include:

[0168] Mean:

[0169]

[0170] Maximum value and minimum value: reflect the extreme value information within the window;

[0171] Trend coefficient: Calculate the trend of the time series through linear regression analysis;

[0172] All these features are combined to form a vector As the feature representation of the signal xi(t) in window k:

[0173]

[0174] (2) Normalization and standardization;

[0175] In order to ensure that the features of different sensors have the same dimension and range, we need to normalize or standardize the extracted features; the normalization formula is:

[0176]

[0177] in, is the normalized eigenvalue, is the eigenvalue, min(fi) and max(fi) are the minimum and maximum values ​​of the feature fi respectively; after feature extraction and normalization, each sensor signal xi(t) can be converted into a feature vector Used to indicate the state at each moment.

[0178] As a preferred embodiment of this embodiment, the embedding method of the time signal is as follows: the embedding method of the time signal is used to map the encoded feature vector to a high-dimensional space, so that the graph neural network can perform feature aggregation and learning in this space;

[0179] (1) Embedding Layer;

[0180] Encoded features Will be mapped to a high-dimensional embedding space; the role of the embedding layer is to project the low-dimensional feature vector to the high-dimensional feature space through a learnable weight matrix We; the mathematical expression of the embedding process is:

[0181]

[0182] in:

[0183] We is the weight matrix of the embedding layer, which represents the linear transformation of the features.

[0184] be is a bias term used to adjust the embedded features.

[0185] is the high-dimensional feature vector embedded in window k by sensor i;

[0186] (2) Temporal dependency of time series embedding;

[0187] In order to preserve the temporal dependencies in time series signals, a time series embedding method based on recurrent neural networks is used; this method embeds the features It is further input into a time recursive network to extract time-dependent high-level features;

[0188] The hidden state update formula of LSTM is as follows:

[0189]

[0190] in:

[0191] is the hidden state at the kth moment, which contains the information of the current window k and the information of the previous moment k-1.

[0192] The LSTM unit retains key information in the past window through the memory gating mechanism and learns the temporal dependency features;

[0193] (3) Embedding processing of graph neural networks;

[0194] For each time window k, after processing through the encoding and embedding layers, all sensor features can form a feature matrix Z, where each row represents the embedded features of a sensor in a specific time window. Combined with the adjacency matrix A, a graph structure G = (A, Z) is generated as the input of the graph neural network.

[0195] As a preferred embodiment of this invention, the application of overall embedding and graph neural network, after the above steps, the encoded time series signal will be converted into a high-dimensional feature representation and input into the graph neural network for further processing; the graph neural network (GNN) will learn the relationship between sensor nodes and the timing characteristics through multi-layer graph convolution operations, so as to detect abnormal situations. The formula of the graph convolution layer is as follows:

[0196]

[0197] in:

[0198] H(l) represents the node features of the lth layer,

[0199] is the adjacency matrix of the graph, which is used to represent the dependency relationship between sensor nodes.

[0200] W(l) is the weight matrix of layer l;

[0201] σ(·) is the activation function.

[0202] As a preferred embodiment of this embodiment, the abnormal detection results can be classified by the following formula:

[0203] y i =softmax(W o Z i +b o );

[0204] in:

[0205] Zi is the embedding feature of node i,

[0206] Wo and bo are the weights and biases of the output layer, used for classification;

[0207] yi is the abnormal state prediction of node i.

[0208] It is worth mentioning that the key points and protection points of the present invention are concentrated on the integrated application of multi-sensor time series data fusion, time-dependent learning of graph neural network models, feature embedding representation, and real-time monitoring and early warning systems. Through the combination of these technologies, the system can effectively monitor the status of paint spraying robot arms in complex production environments, promptly detect potential abnormal phenomena, and ensure efficient and stable operation of the production line.

[0209] The intelligent detection method for abnormal working state of a paint spraying robot arm based on time series graph neural network technology of the present invention, when in use, collects multiple sensor data during the working process of the robot arm in real time, constructs a time series graph neural network model, detects abnormal states during work (such as excessive vibration, uneven spraying, abnormal torque, etc.), and warns of potential faults.

[0210] In the description of this specification, the description with reference to the terms "embodiment", "one embodiment", "certain embodiments", "illustrative embodiment", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0211] The basic principles, main features and advantages of the present invention are described above in conjunction with specific embodiments. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. An intelligent detection method for abnormal working state of a paint spraying robot arm based on time series graph neural network technology, characterized in that: The steps include: S1. Time series data modeling of multi-sensor fusion: Fusion of data from multiple sensors to perform multi-dimensional modeling of time series signals collected by sensors; S2. Anomaly detection and time series learning based on graph neural network: Introduce time series graph neural network (GNN) into robot arm state monitoring, and jointly learn the time series dependency and spatial relationship of signals from different sensors; S3. Multi-layer embedding and feature representation: embedding multi-dimensional sensor time series signals into high-dimensional space; S4. Real-time monitoring and abnormal warning system: collect and process the robot arm sensor signals in real time, and issue a warning signal if an abnormal situation occurs, where the abnormal situation includes abnormal warning data; S5. Data analysis and calculation: The central processor is used to control the data analysis and calculation module to analyze and detect the abnormal warning data monitored, and analyze the causes of abnormal working status and abnormal data points; S6. Results and output: Through the processing of the graph neural network, the output layer finally generates the detection results based on the embedded features and timing information.

2. According to claim 1, a method for detecting abnormal working state of a paint spraying robot arm based on time series graph neural network technology is characterized in that: The time series data modeling of multi-sensor fusion constructs a graph structure through the time series data of multiple sensors, and then the signal is analyzed through time series and the graph structure is constructed.

3. According to claim 1, a method for detecting abnormal working state of a paint spraying robot arm based on time series graph neural network technology is characterized in that: The graph neural network-based anomaly detection and time series learning, especially the learning model of multi-sensor node feature aggregation and time series dependency through graph convolution.

4. According to claim 1, a method for detecting abnormal working state of a paint spraying robot arm based on time series graph neural network technology is characterized in that: The embedding layer extracts high-level features in the time series, retains key information of the robot arm during dynamic work, and protects the multi-layer embedding method of sensor data, especially in the processing steps of feature encoding, normalization and embedding of time series data into high-dimensional space.

5. The method for detecting abnormal working state of a paint spraying robot arm based on time series graph neural network technology according to claim 1 is characterized in that: The method protects the real-time monitoring and abnormal warning system, and is a technical solution for capturing abnormal changes in sensor data and responding to them during the dynamic operation of the robot arm.

6. The method for detecting abnormal working state of a paint spraying robot arm based on time series graph neural network technology according to claim 1 is characterized in that: The sensor input signal and variable number are as follows: (1) Position sensor: monitors the rotation angle (P1) and displacement (P2) of the robot arm; (2) 6-axis force / torque sensor: capable of measuring force (Fx, Fy, Fz) and torque (Tx, Ty, Tz) on three spatial axes, helping to monitor force feedback during the spraying process; (3) Speed ​​sensor: integrated with the position sensor to measure the speed (V) of the robot arm or spray head in real time; (4). Spray gun pressure sensor: monitors the pressure of gas or liquid during spraying (PS); (5). Gas flow sensor: monitors the flow rate (FS) of compressed air; (6). Paint pipeline temperature sensor: monitors the temperature of paint in the pipeline (TS); (7) Ultrasonic sensor: Use ultrasonic waves to detect the distance between the spray head and the workpiece (DS); (8) Vibration sensor: The eddy current sensor uses a relative non-contact sensor with a wide frequency range (0-10 kHz). It is used for measuring static displacement, vibration displacement, and vibration measurement of rotating shafts in rotating machinery. (9). Electrostatic voltage sensor: Model: AAC-EVS-I, threshold setting: ±100V, alarm range: ±50-±5000V, detection angle: <15 degrees, working voltage: 12V power supply, communication distance: <500MM, alarm indication: sound and light alarm.

7. The method for detecting abnormal working state of a paint spraying robot arm based on time series graph neural network technology according to claim 1 is characterized in that: The encoding method of the time signal is as follows: The time series signals from multiple sensors are first processed through the encoding layer, sensor signal X = {x1(t), x2(t), ...xn(t)}, where each sensor xi(t) is a time series function; (1) Sliding window and feature extraction; For each time series signal xi(t), the sliding window technique is used to segment the signal; the size of the sliding window is set to w, and the signal in each window can be expressed as: in, Represents the signal data in the kth window; Within each sliding window, statistical features are extracted from the time series. These features include: Mean: Standard Deviation: Maximum value and minimum value: reflect the extreme value information within the window; Trend coefficient: Calculate the trend of the time series through linear regression analysis; All these features are combined to form a vector As the feature representation of the signal xi(t) in window k: (2) Normalization and standardization; In order to ensure that the features of different sensors have the same dimension and range, we need to normalize or standardize the extracted features; the normalization formula is: in, is the normalized eigenvalue, is the eigenvalue, min(fi) and max(fi) are the minimum and maximum values ​​of the feature fi respectively; after feature extraction and normalization, each sensor signal xi(t) can be converted into a feature vector Used to indicate the state at each moment.

8. The method for detecting abnormal working state of a paint spraying robot arm based on time series graph neural network technology according to claim 7 is characterized in that: The embedding method of the time signal is as follows: The embedding method of the time signal is used to map the encoded feature vector to a high-dimensional space, so that the graph neural network can perform feature aggregation and learning in this space; (1) Embedding Layer; Encoded features Will be mapped to a high-dimensional embedding space; the role of the embedding layer is to project the low-dimensional feature vector to the high-dimensional feature space through a learnable weight matrix We; the mathematical expression of the embedding process is: in: We is the weight matrix of the embedding layer, which represents the linear transformation of the features. be is a bias term used to adjust the embedded features. is the high-dimensional feature vector embedded in window k by sensor i; (2) Temporal dependency of time series embedding; In order to preserve the temporal dependencies in time series signals, a time series embedding method based on recurrent neural networks is used; this method embeds the features It is further input into a time recursive network to extract time-dependent high-level features; The hidden state update formula of LSTM is as follows: in: is the hidden state at the kth moment, which contains the information of the current window k and the information of the previous moment k-1. The LSTM unit retains key information in the past window through the memory gating mechanism and learns the temporal dependency features; (3) Embedding processing of graph neural networks; For each time window k, after processing through the encoding and embedding layers, all sensor features can form a feature matrix Z, where each row represents the embedded features of a sensor in a specific time window. Combined with the adjacency matrix A, a graph structure G = (A, Z) is generated as the input of the graph neural network.

9. The method for detecting abnormal working state of a paint spraying robot arm based on time series graph neural network technology according to claim 1 is characterized in that: Application of overall embedding and graph neural network. After the above steps, the encoded time series signal will be converted into a high-dimensional feature representation and input into the graph neural network for further processing. The graph neural network (GNN) will learn the relationship between sensor nodes and the timing characteristics through multi-layer graph convolution operations to detect abnormal situations. The formula of the graph convolution layer is as follows: in: H(l) represents the node features of the lth layer, is the adjacency matrix of the graph, which is used to represent the dependency relationship between sensor nodes. W(l) is the weight matrix of layer l; σ(·) is the activation function.

10. The method for detecting abnormal working state of a paint spraying robot arm based on time series graph neural network technology according to claim 1, characterized in that: The anomaly detection results can be classified by the following formula: y i =softmax(W o Z i +b o ): in: Zi is the embedding feature of node i, Wo and bo are the weights and biases of the output layer, used for classification; yi is the abnormal state prediction of node i.

Images