An intelligent detection method for industrial robot terminal jitter

Abstract

The present invention relates to an intelligent detection method for terminal jitter of an industrial robot. Firstly, an acceleration sensor is installed at the terminal of the industrial robot to determine the test motion trajectory of the industrial robot. Secondly, the acceleration sensor collects the vibration signal of the terminal of the industrial robot and divides it into start, run and stop stages. Then, the jitter position is searched through a root mean square threshold and the initial search phase is changed to repeatedly obtain jitter samples, and a continuous wavelet transform is used to calculate a time-frequency diagram of the terminal jitter samples. Finally, an optimal convolutional neural network structure is adaptively determined through a small number of training samples, and the optimal network is used to perform deep feature extraction and classification on the time-frequency diagram to realize the recognition of the terminal jitter state of the industrial robot. The present invention combines sensor detection technology with artificial intelligence algorithm, realizes accurate intelligent detection of terminal jitter of the industrial robot through a small number of sensors and adaptive convolutional neural networks, and has the advantages of low cost, simple operation and high accuracy.

Description

Technical Field

[0001] The invention relates to the technical field of intelligent detection of industrial robots, and specifically discloses an intelligent detection method for terminal jitter of an industrial robot. Background Art

[0002] Industrial robots are an important part of intelligent manufacturing and are widely used in mechanical processing, welding, assembly and other manufacturing. Industrial robots can also realize multifunctional applications such as palletizing, handling, spraying, etc. in the automotive, 3C electronics, home appliances, photovoltaic and other industries due to their advantages such as multiple degrees of freedom, large workspace, repeatable programming and automatic control. However, in high-precision industry applications, such as laser welding, gluing, high-precision handling, etc., higher requirements are placed on industrial robots, such as robots cannot shake. If abnormal shaking occurs, it will seriously affect production quality and efficiency, so it is very important to detect the shaking state of the industrial robot end.

[0003] Existing industrial robot terminal jitter analysis is usually performed by manual detection and modal analysis: manual detection determines whether the terminal generates jitter through experience or related indicators, and cannot achieve qualitative and quantitative analysis of the degree of jitter, while modal analysis is difficult to achieve complex joint robot structure modeling and is greatly affected by specific environments and working conditions. Therefore, a detection method that combines sensor technology and artificial intelligence algorithms is urgently needed to identify industrial robot terminal jitter. Summary of the invention

[0004] In view of this, the present invention provides an intelligent detection method for the terminal jitter of an industrial robot in order to solve the problem that the degree of terminal jitter of an industrial robot in actual engineering is difficult to effectively distinguish through simple indicators, and the identification process is subject to large external interference and uneven distribution of jitter positions, resulting in inaccurate detection results.

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

[0006] An intelligent detection method for industrial robot terminal jitter includes the following steps:

[0007] S1. Install an acceleration sensor at the end of the industrial robot to determine the test motion trajectory of the industrial robot;

[0008] S2, collect the terminal vibration signal through the acceleration sensor, and use the root mean square value Define the degree of jitter of the industrial robot end under the motion trajectory, repeat sampling, and store the data and labels in the training database;

[0009] S3, dividing the training database data of the complete motion cycle of step S2 into three stages: start, run, and stop;

[0010] S4, searching the jitter position by using the RMS threshold and changing the initial phase of the search to repeatedly obtain jitter samples;

[0011] S5. The time-frequency diagram of the terminal jitter samples is obtained by continuous wavelet transform calculation. The continuous wavelet transform expression is:

[0012]

[0013] where h a，b (t) is the wavelet basis function, a is the time shift factor, b is the scaling factor, and Morlet wavelet is used as the basis function; the jitter samples include jitter training samples and jitter test samples;

[0014] S6. Build convolutional neural networks with various structures. The convolutional neural network structure must include convolutional layers and fully connected layers. You can add structures such as pooling layers and batch normalization.

[0015] S7, randomly selecting a small number of training time-frequency samples in step S5 and inputting them into the various networks in step S6, and performing multiple iterations to adaptively and quickly find the network structure with the minimum classification loss;

[0016] S8, inputting the training time-frequency samples obtained in step S5 into the optimal network structure in step S7, and performing multiple iterations to fully learn and obtain a training model;

[0017] S9. Collect the vibration signal of the industrial robot terminal according to the trajectory of step S1, process the data through steps S4 and S5, input the test sample into the adaptive convolutional neural network of step S8, and output the degree of jitter of each movement stage of the terminal.

[0018] Furthermore, in step S1, the acceleration sensor is installed at a position close to the wrist of the industrial robot to prevent the flexibility of the end effector of the industrial robot from attenuating the jitter signal.

[0019] Furthermore, in step S1, the motion trajectory of the industrial robot test is selected from the motion trajectory with the most obvious jitter in actual production, and it is recommended to use an arc motion trajectory.

[0020] Furthermore, in step S3, when the end of the industrial robot changes its position greatly during start-up and stop and during operation, its vibration signal will show severe impact and oscillation. The first group of instantaneous impact and oscillation in the complete motion cycle is divided into the start-up phase, the last group is divided into the stop phase, and the rest is the operation phase.

[0021] Further, the specific steps of obtaining the jitter sample in step S4 are:

[0022] S41, set jitter sample length n, initial phase Taking the RMS threshold δ as the screening index, when searching for the RMS value of sample x(n) When , it is recorded as a jitter sample, otherwise it is recorded as a non-jitter sample, and n is used as the search step length to screen the samples of the entire motion stage; the root mean square threshold δ can accurately screen out the jitter samples, and the root mean square value of the complete motion cycle of the minimum jitter state obtained by experimental measurement is the threshold δ;

[0023] S42, set the initial phase of the change Repeat the screening of the vibration signal to obtain more jitter samples.

[0024] Furthermore, in step S41, δ plays a role in preliminary screening of the shaking position, and can be fine-tuned according to the actual shaking degree of the industrial robot end; in step S42, to ensure that the sample is not repeated, it must meet permissible conditions.

[0025] Furthermore, the specific steps of step S6 to build a convolutional neural network are:

[0026] S61, the convolution layer performs inner product operations with multiple convolution kernels and the input matrix, and performs nonlinear mapping through activation functions to achieve weighted evaluation of input features; the mathematical expression of the convolution process is:

[0027]

[0028] in H is the kth convolution output feature of the jth layer; j is the i-th convolution region of the k-1 layer feature map; H j The convolution element of ; is the weight matrix; is the bias; * is the convolution operation; f(·) is the activation function; the ReLU activation function is used here, and its expression is:

[0029]

[0030] S62, batch normalization reduces the data distribution changes caused by convolution and nonlinear operations by optimizing the variance size and mean position of each intermediate layer, making the extracted features closer to the real data. Batch normalization is added after each convolution layer and fully connected layer; its calculation formula is:

[0031]

[0032] where y i is the output; x i is the input; γ and β are training parameters; ε is a constant added for numerical stability;

[0033] S63. The pooling layer is usually located after the convolution layer. It samples the features obtained in the previous convolution layer to reduce the dimension, reduce the parameters of the neural network and prevent overfitting. However, the pooling layer reduces the dimension of the features, which may cause the loss of some useful information. If the pooling area size is n×n, the side length of the output feature map becomes 1 / n of the original, and the number of feature maps remains unchanged.

[0034] S64, the fully connected layer combines and connects the features extracted by the convolutional layer, and realizes output classification through softmax logistic regression.

[0035] Further, the specific steps of step S7 to obtain the optimal network structure are:

[0036] In order to measure the pros and cons of various structural convolutional neural networks in the current jitter state data, the network structure is selected through cross entropy loss; the average cross entropy loss formula is:

[0037]

[0038] Where: N is the total number of data, log represents the natural logarithm with e as the base, y nk is the output of the convolutional neural network, t nk Indicates whether the label is solved correctly, the correct label is represented by 1, and other labels are represented by 0; when the average cross entropy loss E is closer to 0, it means that the classification effect of the network structure is better.

[0039] The working principle and beneficial effects of this solution are:

[0040] 1. The intelligent detection method for the terminal jitter of an industrial robot disclosed in the present invention, first, an acceleration sensor is installed at the terminal of the industrial robot, and a motion path that is commonly used in actual production and has obvious jitter is used as a test motion trajectory; secondly, the vibration signal of the terminal of the industrial robot is collected by the acceleration sensor, and the start, run, and stop stages are divided; then, sufficient terminal jitter data that can show the motion state of the industrial robot is obtained by searching the jitter position through the root mean square threshold and changing the search initial phase data enhancement method, and the terminal jitter data is decomposed by continuous wavelet transform to obtain a time-frequency diagram that can fully retain the terminal jitter impact and oscillation characteristics; finally, through a small number of training samples, the optimal convolutional neural network (CNN) structure is adaptively determined based on the principle of minimum classification loss, and the optimal network is used to perform deep feature extraction and classification recognition on the time-frequency diagram, thereby realizing the recognition of the terminal jitter state of the industrial robot. The present invention combines sensor detection technology with artificial intelligence algorithm, and realizes accurate intelligent detection of terminal jitter of industrial robots through a small number of sensors and adaptive convolutional neural networks, which has the advantages of low cost, simple operation and high accuracy.

[0041] 2. The intelligent detection method for the terminal jitter of an industrial robot disclosed in the present invention defines the degree of terminal jitter of an industrial robot by the root mean square value representing the total energy of the signal, constructs a jitter sample database, and divides the motion position of the sample, so as to realize the detection of the real-time degree and position of the operation data of the industrial robot. The root mean square threshold and phase-changing data are used to enhance the jitter sample construction method, and the time-frequency diagram obtained by continuous wavelet transform decomposition is used for detection. The sample can locate the jitter position through the impact characteristics of the time domain, and retain the oscillation characteristics of the frequency domain to reduce the impact of the working condition change, which can fully reflect the terminal jitter characteristics of the industrial robot. Convolutional neural networks of various network structures are trained by a small number of samples, and the optimal network structure is adaptively selected as the training model of the current jitter sample database with the help of classification loss. This method can adaptively select the optimal network structure for a sample library that increases or decreases the jitter type at any time, and has the advantages of fast network determination speed and strong adaptability to corresponding data.

[0042] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flow chart of the intelligent detection method for the end jitter of an industrial robot according to the present invention;

[0044] Figure 2 For the present invention Figure 1 Flowchart of medium jitter sample construction and phase-variable data enhancement;

[0045] Figure 3 The results of intelligent detection of industrial robot terminal jitter with different network structures are shown in the figure;

[0046] Figure 4 The following are the results of intelligent detection of industrial robot end-user jitter for different data types. DETAILED DESCRIPTION

[0047] The following is further described in detail through specific implementation methods:

[0048] like Figures 1-2 The method for intelligently detecting the end jitter of an industrial robot shown in the figure comprises the following steps:

[0049] S1. Install an acceleration sensor at the end of the industrial robot to determine the test motion trajectory of the industrial robot. The acceleration sensor is installed close to the wrist of the industrial robot to prevent the flexibility of the end effector of the industrial robot from attenuating the jitter signal. The test motion trajectory of the industrial robot selects the motion trajectory with the most obvious jitter in actual production. It is recommended to use an arc motion trajectory.

[0050] S2, collect the terminal vibration signal through the acceleration sensor, and use the root mean square value Define the degree of jitter of the industrial robot end under the motion trajectory, repeat sampling, and store the data and labels in the training database.

[0051] S3. Divide the training database data of the complete motion cycle of step S2 into three stages: start, run, and stop; the vibration signal of the industrial robot terminal will show severe impact and oscillation when the start and stop and running angle changes greatly. The first group of instantaneous impact and oscillation in the complete motion cycle is divided into the start stage, the last group is divided into the stop stage, and the rest is the running stage.

[0052] S4, searching the jitter position by using the RMS threshold and changing the initial phase of the search to repeatedly obtain jitter samples, the specific steps are as follows:

[0053] S41, set jitter sample length n, initial phase Taking the RMS threshold δ as the screening index, when searching for the RMS value of sample x(n) When , it is recorded as a jitter sample, otherwise it is recorded as a non-jitter sample, and n is used as the search step length to screen the samples of the entire motion stage; the root mean square threshold δ can accurately screen out the jitter samples, and the root mean square value of the complete motion cycle of the minimum jitter state obtained by experimental measurement is the threshold δ; δ plays the role of preliminary screening of the jitter position, and can be fine-tuned according to the actual jitter degree of the industrial robot end;

[0054] S42, set the initial phase of the change Repeatedly screen the vibration signal to obtain more jitter samples; to ensure that the samples are not repeated, the permissible conditions.

[0055] S5. The time-frequency diagram of the terminal jitter samples is obtained by continuous wavelet transform calculation. The continuous wavelet transform expression is:

[0056]

[0057] where h a，b (t) is the wavelet basis function, a is the time shift factor, b is the scaling factor, and Morlet wavelet is used as the basis function; the jitter samples include jitter training samples and jitter test samples;

[0058] S6. Build convolutional neural networks with various structures. The convolutional neural network structure must include convolutional layers and fully connected layers. You can add structures such as pooling layers and batch normalization. The specific steps are:

[0059] S61, the convolution layer performs inner product operations with multiple convolution kernels and the input matrix, and performs nonlinear mapping through activation functions to achieve weighted evaluation of input features; the mathematical expression of the convolution process is:

[0060]

[0061] in H is the kth convolution output feature of the jth layer; j is the i-th convolution region of the k-1 layer feature map; H j The convolution element of ; is the weight matrix; is the bias; * is the convolution operation; f(·) is the activation function; the ReLU activation function is used here, and its expression is:

[0062]

[0063] S62, batch normalization reduces the data distribution changes caused by convolution and nonlinear operations by optimizing the variance size and mean position of each intermediate layer, making the extracted features closer to the real data. Batch normalization is added after each convolution layer and fully connected layer; its calculation formula is:

[0064]

[0065] where y i is the output; x i is the input; γ and β are training parameters; ε is a constant added for numerical stability;

[0066] S63. The pooling layer is usually located after the convolution layer. It samples the features obtained in the previous convolution layer to reduce the dimension, reduce the parameters of the neural network and prevent overfitting. However, the pooling layer reduces the dimension of the features, which may cause the loss of some useful information. If the pooling area size is n×n, the side length of the output feature map becomes 1 / n of the original, and the number of feature maps remains unchanged.

[0067] S64, the fully connected layer combines and connects the features extracted by the convolutional layer, and realizes output classification through softmax logistic regression.

[0068] S7, randomly selecting a small number of training time-frequency samples in step S5 and inputting them into the various networks in step S6, and performing multiple iterations to adaptively and quickly find the network structure with the minimum classification loss; the specific steps are:

[0069] In order to measure the pros and cons of various structural convolutional neural networks in the current jitter state data, the network structure is selected through cross entropy loss; the average cross entropy loss formula is:

[0070]

[0071] Where: N is the total number of data, log represents the natural logarithm with e as the base, y nk is the output of the convolutional neural network, t nk Indicates whether the label is solved correctly, the correct label is represented by 1, and other labels are represented by 0; when the average cross entropy loss E is closer to 0, it means that the classification effect of the network structure is better.

[0072] S8, inputting the training time-frequency samples obtained in step S5 into the optimal network structure in step S7, and performing multiple iterations to fully learn and obtain a training model;

[0073] S9. Collect the vibration signal of the industrial robot terminal according to the trajectory of step S1, process the data through steps S4 and S5, input the test sample into the adaptive convolutional neural network of step S8, and output the degree of jitter of each movement stage of the terminal.

[0074] Among them: 1) Install the acceleration sensor and determine the end jitter test motion trajectory, collect vibration signals of different jitter degrees at the end of the industrial robot, organize part of it as training data, and organize part of it as test data. The training data needs to be divided into three stages: start, run, and stop. 2) Set the root mean square threshold δ, sample length n, and multiple initial phases The training and test data are searched and intercepted repeatedly to obtain the jitter samples that fully reflect each motion condition. Then, the terminal jitter time-frequency diagram is obtained by continuous wavelet transform calculation using Morlet wavelet as wavelet basis function.

[0075] 3) Build CNN models with various network structures and set network hyperparameters. Then input a small amount of training samples into each CNN network, obtain the classification loss of the training data through iterative learning, and determine the optimal network structure based on the principle of minimizing the classification loss. Finally, input all training samples into the optimal CNN network, perform multiple iterations to fully learn and obtain the training model.

[0076] 4) Intelligent detection of industrial robot terminal jitter. Input the test sample into the trained CNN model to obtain the terminal jitter state recognition result.

[0077] Example

[0078] The following example illustrates this embodiment:

[0079] Step 1: Signal acquisition and data division. The sensor is installed at the end of the industrial robot, with an arc as the test trajectory. By adjusting the control parameters of the industrial robot, vibration signals of four motion states, normal, small jitter, medium jitter, and large jitter, are collected. The experimental sampling frequency is 10240Hz, the sampling length is 160k, and each motion state signal is sampled 10 times. Two groups of each motion state are randomly selected as test sets, and the rest are training sets. The first group of instantaneous impacts and shocks in the complete motion cycle of the training data are divided into the startup stage, the last group is divided into the stop stage, and the rest are the running stage.

[0080] Step 2: Jitter sample construction. Set the sample length n = 2048, the initial phase The root mean square value of the small vibration signal in the training set is set as the threshold δ = 0.9. Figure 2 The method is used to obtain samples of each jitter state, and the root mean square threshold is set to 0 to intercept the samples of normal motion state in sequence. Then the terminal jitter time-frequency samples are calculated by continuous wavelet transform. The 10 types of jitter states obtained in this example are: a) small jitter-starting state; b) small jitter-running state; c) small jitter-stopping state; d) medium jitter-starting state; e) medium jitter-running state; f) medium jitter-stopping state; g) large jitter-starting state; h) large jitter-running state; i) large jitter-stopping state; j) normal state. A total of 5653 jitter training samples and 1435 test samples were obtained.

[0081] Step 3: Adaptive CNN model construction and training. The constructed CNN structure is shown in Table 1. The parameters of the three convolutional layers in the network structure are: 64 convolution kernels in the first layer, 128 convolution kernels in the second layer, and 256 convolution kernels in the third layer, and the convolution kernel size is 5×5. The pooling layer selects the maximum value pooling with a size of 2×2. Batch normalization is added after each convolution layer and fully connected layer. The optimizer selects the stochastic gradient descent algorithm (SDG), the learning rate is 0.002, the momentum is set to 0.9, the batch size is 256, and the number of iterations k=50. 200 training samples are randomly selected in proportion and input into each CNN network structure for training, and the optimal CNN structure is adaptively selected. Finally, all the training samples in step 2 are input into the CNN network, and multiple iterations are performed to fully learn, and a model that can be used for intelligent detection of the degree of terminal jitter of industrial robots is obtained.

[0082] Table 1 Different CNN network structures

[0083]

[0084] Step 4: Intelligent detection of industrial robot terminal vibration. Input the test sample in step 2 and obtain the test result.

[0085] Table 2 shows the optimization and detection results of different network structures. The optimization data are 200 randomly selected time-frequency training samples, and the detection data are the time-frequency sample data set obtained in step 2. The average loss and average accuracy of the last 5 rounds are used as the test results. Figure 3 Comparison chart of test iterations for different network structures. Figure 3 As shown in Table 2, the CNN structure with the lowest training loss is network model 1, and its corresponding average test loss is also the lowest at 1.5×10 -4 The average test accuracy is 95.9%, indicating that the optimization method of the present invention is reliable and effective. The optimization time of the four network models consumes a total of 1.956 minutes, which is much lower than the normal training time of each network structure, indicating that the method of the present invention has the advantage of high efficiency.

[0086] Table 2 Detection results of different network structures

[0087]

[0088] Figure 4 The detection results of the data type of the present invention and the enhanced time domain signal, the unenhanced time-frequency graph and the non-jittered sample time-frequency graph (samples are intercepted in sequence) in the optimal CNN, and the average loss and average accuracy of the last 5 rounds of the test set are shown in Table 1. Figure 4 As shown in Table 3, the proposed method can converge quickly. In the sixth round, its accuracy is 93.4%, and after multiple iterations, the accuracy can reach 95.9%. The test loss is 1.5×10 -4 , and its results are significantly better than other types of data, proving that the time-frequency diagram obtained by continuous wavelet transform can accurately characterize the impact and oscillation characteristics generated by the jitter process, which is conducive to the deep feature extraction of the network, indicating that the data type of the present invention is effective for industrial robot terminal detection. The average accuracy of the enhanced time domain signal method is only 74.7%, and large fluctuations occur during the iteration process. This is because the time domain signal only shows time-related information and lacks oscillation characteristics that can alleviate the impact of variable working conditions, thereby reducing the stability of terminal jitter state recognition. The unenhanced time-frequency diagram cannot fully reflect the jitter state information of the industrial robot terminal due to the small amount of data, so its convergence speed and recognition accuracy are not as good as the proposed method. Since the non-jitter sample time-frequency diagram does not screen the terminal motion data, the samples of each jitter degree are mixed with non-jitter state information, which is easy to be confused with normal samples, so its test accuracy is only the lowest 66.5%.

[0089] Table 3 Recognition results of different data types

[0090]

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of protection of the claims of the present invention.

Claims

1. An intelligent detection method for industrial robot end jitter, characterized by: The following steps are involved: S1. Install an acceleration sensor at the end of the industrial robot to determine the test motion trajectory of the industrial robot; S2, collect the terminal vibration signal through the acceleration sensor, and use the root mean square value Define the degree of jitter of the industrial robot end under the motion trajectory, repeat sampling, and store the data and labels in the training database; S3, dividing the training database data of the complete motion cycle of step S2 into three stages: start, run, and stop; S4, searching the jitter position by using the RMS threshold and changing the initial phase of the search to repeatedly obtain jitter samples; S5. The time-frequency diagram of the terminal jitter samples is obtained by continuous wavelet transform calculation. The continuous wavelet transform expression is: where h a，b (t) is the wavelet basis function, a is the time shift factor, b is the scaling factor, and Morlet wavelet is used as the basis function; the jitter samples include jitter training samples and jitter test samples; S6. Build convolutional neural networks with various structures, where the convolutional neural network structure includes convolutional layer, fully connected layer, and the added structure is pooling layer and batch normalization; S7, randomly selecting a small number of training time-frequency samples in step S5 and inputting them into the various networks in step S6, and performing multiple iterations to adaptively and quickly find the network structure with the minimum classification loss; S8, inputting the training time-frequency samples obtained in step S5 into the optimal network structure in step S7, and performing multiple iterations to fully learn and obtain a training model; S9. Collect the vibration signal of the industrial robot terminal according to the trajectory of step S1, process the data through steps S4 and S5, input the test sample into the adaptive convolutional neural network of step S8, and output the degree of jitter of each movement stage of the terminal.

2. The intelligent detection method for industrial robot end jitter according to claim 1 is characterized in that: In step S1, the acceleration sensor is installed at the wrist of the industrial robot to prevent the flexibility of the end effector of the industrial robot from attenuating the jitter signal.

3. The intelligent detection method for industrial robot end jitter according to claim 2 is characterized in that: In step S1, the test motion trajectory of the industrial robot selects the arc motion trajectory with the most obvious jitter in actual production.

4. The intelligent detection method for industrial robot end jitter according to claim 1 is characterized in that: In step S3, the end of the industrial robot changes its position during start-up and stop and during operation, and its vibration signal will experience severe shocks and oscillations. The first group of instantaneous shocks and oscillations in the complete motion cycle is divided into the start-up phase, the last group is divided into the stop phase, and the rest is the operation phase.

5. The intelligent detection method for industrial robot end jitter according to claim 1 is characterized in that: The specific steps of step S4 to obtain the jitter sample are: S41, set jitter sample length n, initial phase Taking the RMS threshold δ as the screening index, when searching for the RMS value of sample x(n) When , it is recorded as a jitter sample, otherwise it is recorded as a non-jitter sample, and n is used as the search step length to screen the samples of the entire motion stage; the root mean square threshold δ can accurately screen out the jitter samples, and the root mean square value of the complete motion cycle of the minimum jitter state obtained by experimental measurement is the threshold δ; S42, set the initial phase of the change Repeat the screening of the vibration signal to obtain more jitter samples.

6. The intelligent detection method for industrial robot end jitter according to claim 5 is characterized in that: In step S41, δ plays a role in preliminary screening of the shaking position, and can be fine-tuned according to the actual shaking degree of the industrial robot end; in step S42, to ensure that the sample is not repeated, it must meet permissible conditions.

7. The intelligent detection method for industrial robot end jitter according to claim 1 is characterized in that: The specific steps of step S6 to build a convolutional neural network are: S61, the convolution layer performs inner product operations on multiple convolution kernels and the input matrix, and performs nonlinear mapping through activation functions to achieve weighted evaluation of input features; the mathematical expression of the convolution process is: in H is the kth convolution output feature of the jth layer; j is the i-th convolution region of the k-1 layer feature map; H j The convolution element of ; is the weight matrix; is the bias; * is the convolution operation; f(·) is the activation function; the ReLU activation function is used here, and its expression is: S62, batch normalization reduces the data distribution changes caused by convolution and nonlinear operations by optimizing the variance size and mean position of each intermediate layer, making the extracted features closer to the real data. Batch normalization is added after each convolution layer and fully connected layer; its calculation formula is: where y i is the output; x i is the input; γ and β are training parameters; ε is a constant added for numerical stability; S63, the pooling layer is usually located after the convolution layer. It samples the features obtained in the previous convolution layer to reduce the dimension, reduce the parameters of the neural network and prevent overfitting. After the pooling operation with a pooling area size of n×n, the side length of the output feature map becomes 1 / n of the original, and the number of feature maps remains unchanged; S64, the fully connected layer combines and connects the features extracted by the convolutional layer, and realizes output classification through softmax logistic regression.

8. The intelligent detection method for industrial robot end jitter according to claim 1 is characterized in that: The specific steps of step S7 to obtain the optimal network structure are: In order to measure the pros and cons of various structural convolutional neural networks in the current jitter state data, the network structure is selected through cross entropy loss; the average cross entropy loss formula is: Where: N is the total number of data, log represents the natural logarithm with e as the base, y nk is the output of the convolutional neural network, t nk Indicates whether the label is solved correctly, the correct label is represented by 1, and other labels are represented by 0; when the average cross entropy loss E is closer to 0, it means that the classification effect of the network structure is better.

Images