Model and method for early abnormality monitoring of industrial robots

Abstract

The present invention discloses an early abnormality monitoring model for an industrial robot, comprising an encoder module, an aggregator and a conditional decoder; the encoder module comprises a generation path and a potential path; a first encoder is arranged on the generation path, the first encoder learns the mapping relationship between known context data and the ordinal number of a target data point, and obtains a representation vector of an industrial robot operation cycle signal in a high-dimensional space; a second encoder is arranged on the potential path, the second encoder learns the internal features of the known context data and simulates Gaussian process reasoning, and obtains a representation vector of an industrial robot operation cycle signal in a high-dimensional space; the aggregator is used to aggregate two representation vectors to obtain a global representation parameter; the conditional decoder decodes the global latent variable to obtain a signal function, thereby obtaining a corresponding target data point prediction signal value after inputting the target data point; the present invention also discloses an early abnormality monitoring method for an industrial robot.

Description

Technical Field

[0001] The present invention belongs to the technical field of equipment operation monitoring and management, and specifically relates to an early abnormality monitoring model and method for an industrial robot. Background Art

[0002] With the development of economy and technology and the improvement of people's living standards, the production and sales of automobiles are increasing day by day, and the automobile industry has entered a golden period of development. As a key part of automobile components, the body in white is a collection of all automobile parts. The quality of the body in white welding manufacturing basically determines whether the final quality of the car can meet the acceptance requirements, and its manufacturing cost can almost account for half of the investment in vehicle manufacturing, so the manufacturing process level of the body in white basically determines the final production quality, efficiency and safety performance of the model product. As the core equipment on the body in white welding production line, if there is a sudden abnormal failure, the industrial robot will cause delays in the production plan, affect the production progress of the entire production line, increase production costs, and lead to a decline in welding quality, thereby affecting the quality and safety of the product. Therefore, manufacturers have high requirements for the safety, reliability and stability of the body in white industrial robot. The body in white industrial robot has the advantages of high efficiency, accuracy, stability, consistency and safety. It can improve the production efficiency, product quality and overall quality level of the welding production line. It is an indispensable and important equipment on the automobile body in white welding production line. However, due to the complex and ever-changing production environment, industrial robots often have sudden abnormal failures. However, most current automobile manufacturers use traditional regular manual inspection and maintenance methods to prevent these abnormal failures. This method is time-consuming and labor-intensive, reduces the production efficiency and quality level of the production line, and is also unable to detect and predict abnormal failures in a timely manner. Therefore, it is very necessary to conduct real-time online early abnormal monitoring of industrial robots, adopt more intelligent monitoring and diagnosis technologies, conduct real-time monitoring and early warning of industrial robots, and timely detect and solve potential abnormal failures, which is crucial to ensure the safety, reliability and production efficiency of industrial robots.

[0003] The abnormality monitoring system of mechanical equipment monitors the dynamic physical signals generated during the operation of the equipment, comprehensively evaluates the operation status of the equipment, and determines whether the equipment is currently abnormal. If the judgment result is abnormal, the monitoring system will mark the abnormal equipment and issue an alarm message. Traditional equipment abnormality monitoring relies heavily on mature mathematical, physical and empirical models. These methods require accurate data collection, processing and analysis, and have high requirements for the quality of signal data. In practical applications, signal data is often subject to various interferences, such as noise and signal instability, which poses a challenge to the accuracy of traditional methods. In recent years, with the rapid development of AI technology, big data technology and high-performance computing clusters, machine learning-based methods rely on the powerful computing power of computers to simulate the way humans analyze and deal with problems, learn relevant knowledge from massive data information without relying entirely on human experience and knowledge, and promote abnormality monitoring technology into the intelligent development stage. Because of its good applicability in various scenarios, it has been increasingly used in the industrial field. However, the research on abnormality monitoring of industrial robots faces problems such as lack of abnormal signal data and extremely unbalanced normal and abnormal signal data, which makes it difficult to construct an industrial robot abnormal signal data set with accurate classification labels, and it is impossible to use supervised deep learning methods to achieve online abnormality monitoring of industrial robots. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide an industrial robot early abnormality monitoring model and method, which can perform early abnormality monitoring of industrial robots in real time online to prevent sudden and unexpected failure abnormalities and ensure the safety, reliability and production efficiency of industrial robots.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] The present invention firstly proposes an early abnormality monitoring model for industrial robots, including an encoder module, an aggregator and a conditional decoder;

[0007] The encoder module includes a generative path and a latent path;

[0008] The generation path is provided with a first encoder, and the first encoder is (x C ,y C ,x t ) as input to learn known context data (x C ,y C ) and the target data point ordinal number x t The mapping relationship between them is obtained, and the representation vector R of the industrial robot operation cycle signal in the high-dimensional space is obtained. T ;

[0009] A second encoder is provided on the potential path, and the second encoder uses known context data (x C ,y C ) as input to learn known context data (x C ,y C ) internal features and simulate Gaussian process reasoning to obtain the representation vector R of the industrial robot operation cycle signal in high-dimensional space C ;

[0010] The aggregator is used to aggregate the representation vector R T and the representation vector R C , to obtain the global characterization parameters;

[0011] The conditional decoder receives the global latent variable z r and the target data point ordinal number x t And for the global hidden variable z r Decode and get the determined function f r (x), so that when input x t Then obtain the corresponding target data point prediction signal value

[0012] Where: x C Represents the ordinal number of the context data point; y C represents the signal value of the context data point; x t Indicates the ordinal number of the target data point; Represents the predicted signal value of the target data point; z r Represents the global latent variable, which is obtained by sampling the global representation parameters.

[0013] Furthermore, both the first encoder and the second encoder adopt a channel cross-attention module; the channel cross-attention module includes a channel attention unit and a cross-attention unit connected in series.

[0014] Furthermore, the channel attention unit first uses global maximum pooling operation and global average pooling operation to input context signal value y C The features of are aggregated to obtain the aggregate feature map, and then one-dimensional convolution is used to extract the channel features in the aggregate feature map to achieve local cross-channel interaction and capture the connection between channels. Finally, the sigmoid activation function is used to scale the learned attention coefficient to the range of [0,1] to obtain the signal value y of the additional channel weight channel .

[0015] Furthermore, the cross attention unit adopts the Query-Key-Value mode, with the signal value y output by the channel attention unit channel The corresponding context data point signal value y CConcatenate to obtain weighted signal matrix As value, the ordinal number of the context data point As key, with the ordinal number of the target data point As the query, the input value, key and query are linearly mapped to three different spaces to obtain the query vector q i , key vector k i Sum value vector v i :

[0016] Q=W q x t

[0017] K=W k x c

[0018] V=W v y channel

[0019] For each query vector q i ∈Q can get the output vector h i :

[0020] h i =att((K,V),q i )

[0021] Where: W q , W k , W v are the linear mapping parameter matrices respectively; d r and d x The feature dimensions representing the weighted signal and the ordinal number of the context data point respectively; and Respectively represent each element as d r and d x dimensional vector, and each component in the vector is a real number; Q = [q 1 ,q 2 ,…,q n ], K = [k 1 ,k 2 ,…,k n ] and V = [v 1 ,v 2 ,…,v n ] represent the matrices consisting of query vector, key vector and value vector respectively; i=1,2,…,n.

[0022] Furthermore, the cross attention unit adopts a multi-head cross attention module; the multi-head cross attention module dynamically generates an attention weight matrix by learning the association between value, key and query to deeply mine context data points (x C ,y C ) and the target data point ordinal number x t The latent dependencies between them capture the interactive information in multiple different projection spaces.

[0023] Furthermore, the multi-head cross-attention module is used to project the input data into multiple spaces; each space uses an independent attention head to perform linear transformation on the data, and each attention head obtains Q, K, and V after linear transformation; wherein: Q is the query vector and K is the key vector, which are used to calculate the attention weight, and the attention weight matrix is ​​normalized using the softmax function; V is the value vector, which is used to calculate the output features; the features output by all the attention heads are concatenated to obtain the output features of the multi-head cross-attention module.

[0024] Furthermore, a multi-layer perceptron is provided in the conditional decoder, and the multi-layer perceptron adopts a ReLU activation function and fuses the outputs of the cross-attention units to obtain a representation vector of the industrial robot operation signal data in a high-dimensional space.

[0025] The present invention also proposes an industrial robot early abnormality monitoring method, comprising the following steps:

[0026] Step 1: Collect the operation signal data of the industrial robot on the welding production line; divide the collected industrial robot operation signal data into known context data and target data;

[0027] Step 2: Construct the industrial robot early abnormality monitoring model as described above;

[0028] Step 3: Using known context data and target data to train the industrial robot early abnormality monitoring model to learn the functional distribution characteristics of the normal operation cycle signal and achieve accurate prediction of the normal operation signal value of the industrial robot;

[0029] Step 4: Counting the error distribution range of the normal operation signal value predicted by the early abnormality monitoring model of the industrial robot, and dividing the prediction error range of the normal operation signal value by using the 3σ principle;

[0030] Step 5: Collect the operation signal data of the industrial robot on the welding production line in real time, and use the early abnormality monitoring model of the industrial robot to predict the operation signal value of the industrial robot in real time;

[0031] Step 6: Determine whether the operation signal value falls within the prediction error range of the normal operation signal value: if so, the industrial robot operates normally; if not, the industrial robot operates abnormally.

[0032] The beneficial effects of the present invention are:

[0033] The industrial robot early abnormality monitoring method of the present invention constructs an industrial robot early abnormality monitoring model (CANP) and uses the generated path to learn the known context data (x C ,y C ) and the target data point ordinal number x t The mapping relationship between them is used to learn the known context data (x C ,y C ) internal features, the representation vectors R obtained by aggregating the generated path and the potential path respectively through the aggregator T and the representation vector R C To obtain the global characterization parameters, the global latent variables are obtained by sampling the global characterization parameters and then decoded using the conditional decoder to obtain the signal function to accurately predict the normal operation signal value of the industrial robot; finally, the error distribution range of the CANP model predicting the normal signal value is calculated and statistically analyzed, and the 3σ principle is used to divide the normal signal prediction error range to obtain the abnormal discrimination threshold for abnormal judgment, thereby realizing real-time online abnormal monitoring of industrial robots to prevent sudden and unexpected failures and abnormal situations, and ensure the safety, reliability and production efficiency of industrial robots. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to make the purpose, technical solution and beneficial effects of the present invention clearer, the present invention provides the following drawings for illustration:

[0035] Figure 1 It is the structural diagram of NP model;

[0036] Figure 2 Process diagram for predicting periodic signals of industrial robots based on neural processes;

[0037] Figure 3 This is the structural diagram of the early abnormality monitoring model for industrial robots;

[0038] Figure 4 This is the structural diagram of the channel cross attention module;

[0039] Figure 5 This is the structural diagram of the multi-head cross attention module;

[0040] Figure 6 It is the distribution histogram of the prediction error value of the normal signal;

[0041] Figure 7Scatter plot of prediction error values ​​for normal signals. DETAILED DESCRIPTION

[0042] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0043] The industrial robot early abnormality monitoring method of this embodiment includes the following steps:

[0044] Step 1: Collect the operating signal data of the industrial robot on the welding production line.

[0045] Collect the operating signal data of industrial robots on the welding production line, and divide the collected operating signal data of industrial robots into known context data and target data. Specifically, before or during data collection, analyze the form and characteristics of industrial robot anomalies, and analyze the problems existing in the data collection process and the operating signal data of industrial robots on the welding production line. According to the problems and difficulties that need to be solved in the current research, propose an overall plan for early monitoring of industrial robot anomalies.

[0046] Step 2: Build an early abnormality monitoring model for industrial robots.

[0047] (1) Principles of the early abnormality monitoring model for industrial robots

[0048] The industrial robot periodic signal data is analyzed from the perspective of probability distribution. The function distribution space of normal periodic signals is learned using the neural process (NP) model. Then, a large number of normal periodic signals are accurately predicted. The error distribution range between all predicted values ​​and actual values ​​is calculated and counted. Since the function distribution of abnormal periodic signals is far away from the function distribution space, the corresponding model prediction error will also be outside the error range of normal periodic signals, thereby achieving abnormal identification through function prediction.

[0049] Specifically, the neural process (NP) model is a family of regression function models that can map input to output. It uses the advantages of neural networks in processing high-dimensional features and parallel computing, imitates the inference process of Gaussian process regression (GPR), and can infer the probability distribution of the target function through a small amount of observation data, so that it has some basic properties of GPR, that is, modeling by learning the probability distribution of functions, which is very different from the parameterized learning idea of ​​deep neural networks for single deterministic functions. At the same time, it uses the powerful function fitting ability of neural networks to directly learn implicit kernel functions from data, overcoming the limitation of the Gaussian process that requires the specification of appropriate and fixed kernel functions for specific scenarios in advance. In addition, since the NP model inherits the computational advantages of deep neural networks, it makes up for the high computational complexity of GPR. During the training process, the NP model models multiple tasks of random processes, constructs an encoder network for probability inference based on flexible and variable observation data, obtains the distribution function of the random process through the decoder network, and finally realizes the prediction of the target function.

[0050] like Figure 1 The following is the structure diagram of the NP model. C ,y C ) is the known context observation data, (x t ,y t ) is the target data, z is the global latent variable, is the predicted value. The NP model assumes that the context data is known and predicts the target input data x t The corresponding target output data, the overall model mainly consists of three parts: encoder, aggregator and conditional decoder. The encoder h receives the known context data (x C ,y C ), generate the representation vector r of context data in high-dimensional space through neural network C ; Aggregator a performs aggregation operation on the representation vector to obtain the global representation vector r, which parameterizes the distribution of the global latent variable z. The aggregator performs the mean operation; the conditional decoder g uses the sampled values ​​of the z distribution and the target input x t As input, by sampling the hidden variable z value, we get the deterministic function f(x), so that when input x t Then get the corresponding predicted value

[0051] Based on the deep neural process model, the distribution characteristics of the normal operation cycle signal function of the industrial robot are learned to achieve accurate prediction of the robot's periodic signal. The data set is the normal cycle data of the current, speed and operation phase signal of the industrial robot joint (6 axes). Each operation cycle of the industrial robot lasts about 26 seconds, and the sampling frequency is 250Hz, that is, each operation cycle data is 20 channels and 6500 data points of two-dimensional data, and the signal data shape is 6500×20.

[0052] The robot operation cycle signal data is analyzed from the perspective of probability distribution. It is assumed that the entire signal cycle data set is defined as the set R, and each cycle data is represented as r (r∈R). Assuming that the signal cycle data is generated by a random process F, the ordinal number of each sampling point in a single cycle is regarded as the input, and the corresponding signal value is regarded as the output random variable, then each signal cycle can be regarded as a set of random variables whose signal values ​​change with the signal sampling ordinal number. In other words, the normal robot periodic signal data is analyzed from the perspective of probability distribution. It is assumed that the distribution function of each normal periodic signal data is f r (r∈R), then the distribution functions of all normal periodic signal data constitute a function set space F R , that is, the signal function is distributed in F R The periodic signal in the space is judged as a normal periodic signal, and the signal function is distributed in F R The period outside the space or at a large distance is determined as an abnormal periodic signal. In general, the core idea of ​​this embodiment is to use the training NP model to learn the function distribution space F of normal periodic signals. R , and then use the learned F R The model accurately predicts a large number of normal periodic signals, calculates and counts the error distribution range between all predicted values ​​and actual values, and then the error range is the normal signal function distribution space F R Mapping in the error space. Since the distribution of the abnormal periodic signal function is far away from F R , so the corresponding model prediction error will also be outside the normal periodic signal error range, thereby achieving abnormal discrimination through function prediction.

[0053] Each periodic signal data of an industrial robot consists of 6500 data point pairs. The data point contains an ordinal number x and a corresponding signal value vector y. The ordinal number represents the position of the data point in the entire periodic signal data. Define set A as the set of all ordinals. Then any ordinal number in each period is a i =i∈A, where 1≤i≤6500. Assuming that the signal values ​​of different ordinal points are independent of each other and that there is random noise in the data itself, then the ordinal a i The signal value at is y i =f(a i)+e i , where e i ~N(0,σ 2 ) is the noise value. The conditional probability distribution of the signal value vector is:

[0054]

[0055] A deep neural network is used to approximate the deterministic function g to make the neural process model the random process F. Assuming z is a high-dimensional vector to represent F, then F(A) = g(A,z). The random properties of the random process F are transferred to the hidden variable z, and then the corresponding function f(a) can be obtained by sampling the hidden variable z. i )=g(a i ,z), where function g is a deterministic function approximated by a deep neural network. By approximating the signal value distribution through the deterministic function g, the corresponding target signal value output can be predicted by combining the hidden variable z with the target ordinal input. Therefore, we can get:

[0056]

[0057] The above derivation process is carried out when the hidden variable z has been obtained. In the actual training model process, each normal signal cycle r has its own hidden variable z r As the representation of the cycle in high-dimensional space, the deterministic function g is obtained by training the entire data set through a deep neural network and is applicable to all signal cycle data in the entire set R. Therefore, the probability distribution function f of each signal cycle r is r Use the same function g to approximate:

[0058] f r (a i )≈g(a i ,z r ),r∈R,i∈r

[0059] The function space F formed by the probability distribution function of normal periodic signals R It can be represented by the deterministic function g and the hidden variable {z r ,r∈R}. Obtain the latent variable z r The step is crucial. For a certain three signal cycle data, the first two-thirds of the data points are intercepted as the observed known context data C, where C is the context data point (x C ,y C ) (It should be noted that the context data x C Same as the previous article i , are all subsets of the ordinal set A, x C Specifically refers to the ordinal number in the known context set, a irepresents any ordinal number in set A, y C For x C The corresponding signal value vector). The hidden variable z of this data r Assume z as a random variable in the latent space of the function, so that it obeys the posterior probability distribution of the known context data C. r Subject to a mean μ r , the covariance is ∑ r Gaussian distribution, and μ r ,∑ r Depends on different periodic signal data r. Obtain hidden variable z r The task is now transformed into finding a set of Gaussian distribution parameters ((μ r ,∑ r ):r∈R), similar to the g function, NP uses a deep neural network to approximate a deterministic function l approximate distribution function space F R Every function f in r The posterior distribution parameters of are:

[0060] l((x C ,y C ))=(μ r ,∑ r ),r∈R

[0061] Figure 2 The deep neural process model is used to predict the periodic signal of industrial robots based on neural processes. C ,y C ) to obtain the appropriate kernel function and hidden variables, and the model finally predicts the target sequence x through the input t Find the corresponding signal value vector y t The distribution function f r , to achieve accurate prediction of the target signal.

[0062] (2) Early abnormality monitoring model for industrial robots

[0063] The structure of the industrial robot early abnormality monitoring model constructed in this embodiment is as follows: Figure 3 As shown, the industrial robot early abnormality monitoring model of this embodiment includes an encoder module, an aggregator and a conditional decoder.

[0064] In this embodiment, the encoder module includes a generation path and a potential path, which correspond to the solid line and the dotted line in the figure respectively, and the two paths have different input data. Among them, a first encoder is provided on the generation path, and the first encoder is (x C ,y C ,x t ) as input to learn known context data (x C ,yC ) and the target data point ordinal number x t The mapping relationship between them is obtained, and the representation vector R of the industrial robot operation cycle signal in the high-dimensional space is obtained. T The second encoder is set on the potential path, and the second encoder takes the known context data (x C ,y C ) as input to learn known context data (x C ,y C ) internal features and simulate Gaussian process reasoning to obtain the representation vector R of the industrial robot operation cycle signal in high-dimensional space C . Where: x C Represents the ordinal number of the context data point; y C represents the signal value of the context data point; x t Indicates the ordinal number of the target data point; Represents the predicted signal value of the target data point; z r Represents the global latent variable, which is obtained by sampling the global representation parameters.

[0065] In this embodiment, the aggregator is used to aggregate the representation vector R T and the representation vector R C To obtain the global representation parameters. Specifically, the aggregator performs a mean operation to aggregate two representation vectors R containing different information. T and R C , and obtain the global characterization parameters.

[0066] In this embodiment, the conditional decoder receives the global latent variable z r and the target data point ordinal number x t And for the global hidden variable z r Decode and get the signal function f r (x), so that when input x t Then obtain the corresponding target data point prediction signal value In this embodiment, the conditional decoder is a multi-layer perceptron MLP, which uses the ReLU activation function. The conditional decoder receives the global latent variable z r With the target data ordinal x t , and for z r Decode and get the determined function f r (x), so that when input x t Then obtain the corresponding target signal prediction value

[0067] In this embodiment, both the first encoder and the second encoder use a channel cross attention module. Figure 4As shown in Figure 1, the channel cross attention module includes a channel attention unit and a cross attention unit connected in series. The channel attention unit first uses the global maximum pooling operation (GMP) and the global average pooling (GAP) operation to the input context signal value y C The features of are aggregated to obtain the aggregate feature map, and then one-dimensional convolution is used to extract the channel features in the aggregate feature map to achieve local cross-channel interaction and capture the connection between channels. Finally, the sigmoid activation function is used to scale the learned attention coefficient to the range of [0,1] to obtain the signal value y of the additional channel weight channel That is, the principle of the channel attention unit of this embodiment is: first, the context signal value y C As input, the global features of the signal are aggregated through global maximum pooling (GMP) and global average pooling (GAP) operations. Then 1DCNN is used to extract channel features, realize effective local cross-channel interaction, and replace unnecessary global channel interaction to capture the connection between channels. Then the learned attention coefficient is scaled to the range of [0,1] through the sigmoid activation function, and the signal value y with additional channel weight is obtained through aggregation channel . The industrial robot signal data contains 20 channels of four types of signals, including the motor current, speed, joint rotation angle, and robot operation stage of the robot's six joints. There are complex connections between the channels, and industrial noise exists in the signal data. The channel attention unit of this embodiment can mine the connections between channels, and through the learned attention weights, it can highlight important signal channels and weaken channels with low feature density, thereby achieving the effect of suppressing industrial noise and highlighting key channel information.

[0068] The cross attention unit in this embodiment is structurally the same as the self-attention model, adopting the Query-Key-Value mode, with the signal value y output by the channel attention unit channel The corresponding context data point signal value y C Concatenate to obtain weighted signal matrix As value, the ordinal number of the context data point As key, with the ordinal number of the target data point As the query, the input value, key and query are linearly mapped to three different spaces to obtain the query vector q i , key vector k i Sum value vector v i :

[0069] Q=W q x t

[0070] K=W k x c

[0071] V=W v y channel

[0072] For each query vector q i ∈Q can get the output vector h i :

[0073] h i =att((K,V),q i )

[0074] Where: W q , W k , W v are the linear mapping parameter matrices respectively; d r and d x The feature dimensions representing the weighted signal and the ordinal number of the context data point respectively; and Respectively represent each element as d r and d x dimensional vector, and each component in the vector is a real number; Q = [q 1 ,q 2 ,…,q n ], K = [k 1 ,k 2 ,…,k n ] and V = [v 1 ,v 2 ,…,v n ] represent the matrices consisting of query vector, key vector and value vector respectively; i=1,2,…,n.

[0075] In this embodiment, the cross attention unit adopts a multi-head cross attention module, which dynamically generates an attention weight matrix by learning the association between value, key and query to deeply mine the context data point (x C ,y C ) and the target data point ordinal number x t The potential dependencies between them can capture interactive information in multiple different projection spaces. Figure 5 As shown, the multi-head cross-attention module is used to project the input data into multiple spaces; each space uses an independent attention head to perform linear transformation on the data, and each attention head obtains Q, K, and V after linear transformation; where: Q is the query vector and K is the key vector, which are used to calculate the attention weight, and the attention weight matrix is ​​normalized using the softmax function; V is the value vector, which is used to calculate the output features; the features output by all the attention heads are concatenated to obtain the output features of the multi-head cross-attention module. Figure 5 middle, represents the transpose of the i-th head key vector k; represents the transpose of the i-th header value vector v; q T represents the transpose of the query vector Q; K T represents the transpose of the key vector K; d k represents the dimension of the key vector k; Represents the weight corresponding to the query vector Q; Represents the weight corresponding to the key vector K; Represents the weight corresponding to the value vector V; head h represents the output features of the h-th attention head; h = 1, 2, …, H; i = 1, 2, …, n.

[0076] Correspondingly, a multi-layer perceptron is provided in the conditional decoder of this embodiment. When the cross-attention unit adopts a multi-head cross-attention module, the multi-layer perceptron adopts a ReLU activation function and fuses the outputs of the cross-attention unit to obtain a representation vector of the industrial robot operation signal data in a high-dimensional space.

[0077] Step 3: Training the industrial robot early anomaly detection model

[0078] The early abnormality monitoring model of industrial robots is trained using known context data and target data to learn the distribution characteristics of signal functions and achieve accurate prediction of the normal operation signal value of the industrial robot.

[0079] During the training process, the CANP model of this embodiment inputs the signal data of the normal operation cycle of the industrial robot for three consecutive cycles, and randomly extracts signal data points at a ratio of 70%, and uses the data points extracted from the first two cycles as the known context data points (x C ,y C ), extracting data points from all three cycles as the predicted target signal data points (x t ,y t ).

[0080] After being trained with a large amount of normal signal data, the CANP model learns the distribution characteristics of industrial robot signal functions, achieves accurate prediction of signal data, and then calculates and counts the prediction error distribution range of normal signals.

[0081] Step 4: Divide the prediction error range of the normal operation signal value

[0082] The error distribution range of the normal operation signal value predicted by the early abnormality monitoring model of industrial robots is statistically analyzed, and the prediction error range of the normal operation signal value is divided using the 3σ principle.

[0083] For the problem of abnormal monitoring, the selection of abnormal discrimination threshold is particularly important. Prediction-based anomaly detection generally discriminates abnormalities based on prediction errors. When the prediction error is greater than the selected threshold, it is considered abnormal, otherwise it is considered normal. After completing the construction and training of the CANP model, the network weights of the encoder and decoder inside the model have converged, and the model prediction results are good. This shows that the model has learned the functional distribution characteristics of normal signals, so it can realize abnormal judgment based on the size of the model prediction error value. In order to obtain a suitable and effective abnormal discrimination threshold, 2728 normal operating cycle signals of industrial robots are randomly selected from the training set and input into the trained CANP model to calculate the predicted signal value each time. The actual signal value y i The mean absolute error (MAE) of Figure 6 As shown, it is a distribution histogram of normal signal prediction error values.

[0084]

[0085] Here, n represents the data sample size.

[0086] It should be noted that the KS test shows that the prediction error value of the normal signal obeys the normal distribution, so the classic 3σ principle can be used to obtain the abnormal discrimination threshold, that is, the industrial robot operation cycle signal with the prediction error value of the CANP model distributed outside (μ-3σ,μ+3σ) is judged as an abnormal signal. Figure 7 The figure shows a scatter plot of the normal signal prediction error value. It can be seen that the 3σ principle can effectively divide the normal periodic signal prediction error interval of the industrial robot, thereby realizing real-time online abnormal monitoring of the industrial robot and completing early abnormality identification.

[0087] Step 5: Real-time online data collection

[0088] Collect the operation signal data of industrial robots on the welding production line in real time, and use the early abnormality monitoring model of industrial robots to predict the operation signal values ​​of industrial robots in real time.

[0089] Step 6: Abnormality judgment

[0090] Determine whether the operation signal value falls within the prediction error range of the normal operation signal value: if so, the industrial robot operates normally; if not, the industrial robot operates abnormally.

[0091] The above-described embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or changes made by those skilled in the art based on the present invention are within the protection scope of the present invention. The protection scope of the present invention shall be subject to the claims.

Claims

1. A method for early abnormality monitoring of industrial robots, Features: The steps include: Step 1: Collect the operation signal data of the industrial robot on the welding production line; divide the collected industrial robot operation signal data into known context data and target data; Step 2: Build an early abnormality monitoring model for industrial robots; Step 3: Using known context data and target data to train the industrial robot early abnormality monitoring model to learn the functional distribution characteristics of the normal operation cycle signal and achieve accurate prediction of the normal operation signal value of the industrial robot; Step 4: Counting the error distribution range of the normal operation signal value predicted by the early abnormality monitoring model of the industrial robot, and dividing the prediction error range of the normal operation signal value by using the 3σ principle; Step 5: Collect the operation signal data of the industrial robot on the welding production line in real time, and use the early abnormality monitoring model of the industrial robot to predict the operation signal value of the industrial robot in real time; Step 6: Determine whether the operation signal value falls within the prediction error range of the normal operation signal value: if so, the industrial robot operates normally; if not, the industrial robot operates abnormally; The industrial robot early anomaly monitoring model includes an encoder module, an aggregator and a conditional decoder; The encoder module includes a generative path and a latent path; The generation path is provided with a first encoder, and the first encoder is (x C ,y C ,x t ) as input to learn known context data (x C ,y C ) and the target data point ordinal number x t The mapping relationship between them is obtained, and the representation vector R of the industrial robot operation cycle signal in the high-dimensional space is obtained. T ; The first encoder and the second encoder both use a channel cross attention module; the channel cross attention module includes a channel attention unit and a cross attention unit connected in series; A second encoder is provided on the potential path, and the second encoder uses known context data (x C ,y C ) as input to learn known context data (x C ,y C ) internal features and simulate Gaussian process reasoning to obtain the representation vector R of the industrial robot operation cycle signal in high-dimensional space C ; The aggregator is used to aggregate the representation vector R T and the representation vector R C , to obtain the global characterization parameters; The conditional decoder receives the global latent variable z r and the target data point ordinal number x t And for the global hidden variable z r Decode and get the determined function f r (x), so that when input x t Then obtain the corresponding target data point prediction signal value A multi-layer perceptron is provided in the conditional decoder, and the multi-layer perceptron adopts a ReLU activation function and fuses the outputs of the cross attention units to obtain a representation vector of the industrial robot operation signal data in a high-dimensional space; Where: x C Represents the ordinal number of the context data point; y C represents the signal value of the context data point; x t Indicates the ordinal number of the target data point; Represents the predicted signal value of the target data point; z r Represents the global latent variable, which is obtained by sampling the global representation parameters.

2. The method for early abnormality monitoring of industrial robots according to claim 1, Features: The channel attention unit first uses global maximum pooling operation and global average pooling operation to input context signal value y C The features of are aggregated to obtain the aggregate feature map, and then one-dimensional convolution is used to extract the channel features in the aggregate feature map to achieve local cross-channel interaction and capture the connection between channels. Finally, the sigmoid activation function is used to scale the learned attention coefficient to the range of [0,1] to obtain the signal value y of the additional channel weight channel .

3. The method for early abnormality monitoring of industrial robots according to claim 1, Features: The cross attention unit adopts the Query-Key-Value mode, with the signal value y output by the channel attention unit channel The corresponding context data point signal value y C Concatenate to obtain weighted signal matrix As value, with the ordinal number of the context data point As key, with the ordinal number of the target data point As the query, the input value, key and query are linearly mapped to three different spaces to obtain the query vector q i , key vector k i Sum value vector v i : Q=W q x t K=W k x c V=W v y channel For each query vector q i ∈Q can get the output vector h i : h i =that((K,V),q i ) Where: W q , W k , W v are the linear mapping parameter matrices respectively; d r and d x The feature dimensions representing the weighted signal and the ordinal number of the context data point respectively; and Respectively represent each element as d r and d x dimensional vector, and each component in the vector is a real number; Q = [q 1 ,q 2 ,L,q n ], K = [k 1 ,k 2 ,L,k n ] and V = [v 1 ,v 2 ,L,v n ] represent the matrices composed of query vector, key vector and value vector respectively; i=1,2,L,n.

4. The method for early abnormality monitoring of an industrial robot according to claim 3, Features: The cross attention unit adopts a multi-head cross attention module; the multi-head cross attention module dynamically generates an attention weight matrix by learning the association between value, key and query to deeply mine context data points (x C ,y C ) and the target data point ordinal number x t The latent dependencies between them capture the interactive information in multiple different projection spaces.

5. The method for early abnormality monitoring of industrial robots according to claim 4, Features: The multi-head cross-attention module is used to project the input data into multiple spaces; each space uses an independent attention head to perform linear transformation on the data, and each attention head obtains Q, K, and V after linear transformation; wherein: Q is the query vector and K is the key vector, which are used to calculate the attention weight, and the attention weight matrix is ​​normalized using the softmax function; V is the value vector, which is used to calculate the output features; the features output by all the attention heads are concatenated to obtain the output features of the multi-head cross-attention module.

Images