A deep learning-based robot arm fault diagnosis method in a nuclear radiation environment

Abstract

The application provides a kind of nuclear radiation environment based on deep learning mechanical arm fault diagnosis method, belong to mechanical arm technical field, this nuclear radiation environment based on deep learning mechanical arm fault diagnosis method obtains the running data of mechanical arm normal time and the running data of each component after irradiation damage by mechanical arm fault data acquisition module, constructs training sample data;Training sample data is used to train data generation model;Real training sample data and pseudo data sample are used to train preset fault diagnosis model together, wherein real training sample data is the data collected by fault data acquisition module, pseudo data sample is the sample generated by data generation model, and the input of fault diagnosis model is running data, and the output is running condition;The real-time running data of the mechanical arm system to be fault diagnosed is collected, and the collected real-time running data is input into the trained fault diagnosis model, to realize fault diagnosis.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of mechanical arms, and particularly relates to a nuclear radiation environment-based mechanical arm fault diagnosis method based on deep learning. BACKGROUND

[0002] In recent years, with the rapid development of industries in various countries and the rapid growth of the world population, the consumption of traditional energy is accelerating, and clean and renewable energy such as nuclear energy, wind energy and solar energy is increasingly valued. However, due to the limitations of other energy, nuclear energy has become the only energy that can achieve industrial application and replace fossil fuels on a large scale. Many countries have invested huge financial and human resources in the development of nuclear power, and nuclear power is entering a period of rapid development. However, the widespread use of nuclear power has also brought nuclear safety problems. The service environment and operation tasks of radiation-resistant robots are extremely complex: especially in the core area and surrounding area of nuclear power plant accidents, it is a strong radiation, strong corrosion (boric acid erosion), high temperature and high humidity environment. Various charged particles and high-energy rays in the strong radiation environment can cause ionization damage or failure of electronic components and circuit boards in semiconductor devices, integrated circuits, computer systems, information acquisition and transmission systems in the internal equipment of radiation-resistant robots, cause harm to insulation in power supply systems and connection, sealing and lubrication in transmission mechanisms, thereby causing degradation and damage to the performance of radiation-resistant robot systems. Therefore, radiation-resistant robots must have safety protection to ensure long-term operation in a nuclear environment and complete various tasks.

[0003] In recent years, the tasks that radiation-resistant robots need to perform have also become increasingly complex, such as welding, grinding or residue cleaning of equipment, and mechanical arms with various end actuators and high flexibility are undoubtedly the best choice. For radiation-resistant mechanical arms, the development cost is much higher than that of other ordinary mechanical arms, so it is necessary to ensure that the radiation-resistant mechanical arm can work normally after entering a high radiation environment, and to ensure that the robot can retreat as much as possible to reduce losses while providing valuable research data for the next generation of radiation-resistant robots. However, according to different work tasks, radiation-resistant mechanical arms need to carry different types of sensors and end actuators for corresponding work, so the radiation-resistant mechanical arm is a highly integrated mechatronic product. However, in a high radiation environment, every part of the mechatronic system may have different forms of failure and damage. It is extremely difficult to take radiation reinforcement measures for every part of the mechanical arm in terms of cost and mechanical arm structure design. Moreover, due to the complexity of the mechanical arm, traditional fault diagnosis methods are more difficult to accurately diagnose the fault mode and locate the fault damage, so it is necessary to research a fault diagnosis method suitable for this special scenario, find the weak position of the mechanical arm in the radiation environment, take local reinforcement measures, and monitor the running status of the mechanical arm system in real time. SUMMARY

[0004] The application aims to provide a nuclear radiation environment-based mechanical arm fault diagnosis method based on deep learning, aiming to solve the technical problems of the lack of existing mechanical arm fault diagnosis methods in a nuclear radiation environment.

[0005] To achieve the above-mentioned purpose, the application provides the following technical scheme: a nuclear radiation environment-based mechanical arm fault diagnosis method based on deep learning, which comprises the following steps: S1: obtaining a mechanical arm fault data set, wherein the fault data set comprises data collected by each sensor when the mechanical arm is normally running and data when each part of the mechanical arm appears a fault after being irradiated;

[0006] S2: using the above-mentioned data as fault data samples and the running conditions of each component of the corresponding mechanical arm as sample labels to construct a training sample set; wherein the running conditions include normal running and fault types;

[0007] S3: using the training sample set to train a data generation model; wherein the data generation model is a multi-discriminator auxiliary classification generative adversarial network model, the data generation model takes real data samples as input and approximates real data samples as output;

[0008] S4: using the real fault data samples and the pseudo data samples generated by the data generation model as the final training sample set to train a preset fault diagnosis model; wherein the fault diagnosis model is a deep convolutional neural network model, the fault diagnosis model takes mechanical arm running data as input and running conditions as output;

[0009] S5: collecting real-time running data of a mechanical arm to be fault diagnosed and inputting the collected real-time running data into a deployed fault diagnosis model to realize the fault diagnosis model; wherein the deployed fault model is a trained model deployed to a web page by flask.

[0010] The fault types include driver irradiation damage faults, joint motor irradiation damage faults, encoder damage faults, and combinations of any one or more thereof.

[0011] The data generation model comprises three discriminator modules, one generator module, and one classifier module:

[0012] The generator generates labeled data by embedding label information in random noise, inputs the generated data and real data into three discriminators to determine whether the data is real or fake, calculates a loss function, and updates network parameters using an optimizer; the generated data and real data are input into a classifier module for classification, the classifier loss is calculated, and the network parameters are updated; the discriminators and the classifier are trained, the combined structure is trained, the weights of the discriminators and the classifier are kept unchanged, the generator loss is calculated, and the parameters are updated; after one training, the generator and the discriminator model converge to a Nash equilibrium state after sufficient iterations, and the generated data is a pseudo sample approximating the real data.

[0013] The fault diagnosis model comprises a plurality of convolution modules, a global average pooling layer, and a fully connected layer.

[0014] The fault data is input into the plurality of convolution modules to extract features, and the output is sent to the global average pooling layer for pooling processing; the output of the global average pooling layer is sent to the fully connected layer to realize operation condition classification.

[0015] Further, the convolution module comprises a plurality of convolution kernels, a batch normalization layer, and an activation layer.

[0016] Further, the size of the convolution kernel is 3x3, the number is 64, 128, 256, and 512, and the activation function of the activation layer is a ReLu function.

[0017] Further, in the fault diagnosis model, the number of convolution kernels is doubled once after passing through two convolution modules.

[0018] The nuclear radiation environment-based mechanical arm fault diagnosis method also provides a mechanical arm fault diagnosis system, which comprises:

[0019] A mechanical arm fault data acquisition module is configured to acquire operation data of each component of the mechanical arm; wherein the operation data comprises operation data of each component of the mechanical arm in a normal state and operation data of the mechanical arm after each component is damaged by irradiation;

[0020] A fault data sample set construction module is configured to construct a fault data sample set by taking the operation data acquired by the mechanical arm fault data acquisition module as sample data and taking the operation condition of each component of the mechanical arm as sample labels; wherein the operation condition comprises normal and fault types of each component damage.

[0021] The data generation model training module is configured to train a data generation model by using the training sample set constructed by the fault data sample set module; wherein the data generation model is a multi-discriminator generative adversarial network model, and the data generation model takes the operation data as input and outputs pseudo data samples approximating real samples;

[0022] The fault diagnosis model training module is configured to train a preset fault diagnosis model by using the training sample set constructed by the fault data sample set module and the pseudo sample data set generated by the data generation model; wherein the fault diagnosis model is a deep convolutional neural network model, and the fault diagnosis model takes the operation data of the robot arm as input and outputs the operation condition.

[0023] The fault diagnosis module is configured to collect real-time operation data of a robot arm to be diagnosed, and input the collected real-time operation data into the fault diagnosis model trained by the fault diagnosis model training module, so as to realize fault diagnosis.

[0024] The nuclear radiation environment-based robot arm fault diagnosis method based on deep learning further provides a robot arm joint module test platform configured to constitute a robot arm fault data acquisition module; wherein the robot arm joint module test platform comprises an encoder, a torque and speed sensor, a reducer, a torque motor and the like.

[0025] Compared with the prior art, the method has the following advantages:

[0026] 1. The nuclear radiation environment-based robot arm fault diagnosis method based on deep learning acquires operation data of a robot arm in a normal state and operation data of each component after radiation damage by using a robot arm fault data acquisition module, and constructs training sample data; a data generation model is trained by using the training sample data, wherein the data generation model is a multi-discriminator generative adversarial network model, the input of which is operation data and the output of which is pseudo data samples approximating real samples; a preset fault diagnosis model is trained by using real training sample data and pseudo data samples, wherein the real training sample data is collected by the fault data acquisition module, the pseudo data samples are generated by the data generation model, and the fault diagnosis model is a deep convolutional neural network model, the input of which is operation data and the output of which is operation condition; real-time operation data of a robot arm system to be diagnosed is collected, and the collected real-time operation data is input into the trained fault diagnosis model, so as to realize fault diagnosis.

[0027] 2.The nuclear radiation environment-based mechanical arm fault diagnosis method based on deep learning is an end-to-end fault diagnosis method, which avoids excessive data processing operations while ensuring diagnosis accuracy, and reduces the complexity of manual operation.The data generation model can expand the data set and solve the problem of difficult data acquisition in a radiation environment.The fault diagnosis method does not depend on the mathematical model of the mechanical arm, and can identify the fault categories of the mechanical arm by training the collected fault data.The application provides a new evaluation method for the fault diagnosis of the mechanical arm in the radiation environment, and provides a basis for the anti-radiation reinforcement design and emergency maintenance of the radiation-resistant mechanical arm. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. The drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0029] Figure 1 The principle diagram of the nuclear radiation environment-based mechanical arm fault diagnosis method based on deep learning provided by the embodiment of the present application is shown in the figure.

[0030] Figure 2 The flowchart of the nuclear radiation environment-based mechanical arm fault diagnosis method based on deep learning provided by the embodiment of the present application is shown in the figure.

[0031] Figure 3 The structure diagram of the multi-discriminator auxiliary classification generative adversarial network is shown in the figure.

[0032] Figure 4 The structure diagram of the deep convolutional neural network is shown in the figure.

[0033] Figure 5 The composition principle diagram of the mechanical arm joint module test platform is shown in the figure. DETAILED DESCRIPTION

[0034] The technical solutions in the embodiments of the present application will be described clearly and completely with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0035] Embodiment 1

[0036] Please refer to Figures 1-4In the irradiation environment, each part of the robot arm can be damaged, and due to the complexity of the task, the structure is highly integrated. Based on the traditional fault diagnosis method, the model-based method not only needs an accurate mathematical model, but also can only determine which joint is faulty and is difficult to determine the specific components of the faulty joint. The example is based on the deep learning method, and a deep convolutional neural network is designed for robot arm fault diagnosis. The diagnosis principle of the method is as shown in Figure 1 The implementation process is as shown in Figure 2 The implementation process is as shown in

[0037] Step one: obtain the fault data set of the robot arm, wherein the fault data set includes the data collected by the sensor when the robot arm is running normally and the running data of each component of the robot arm after irradiation;

[0038] Step two: based on the above data as the fault data sample, the running condition of each component of the robot arm as the sample label, a training sample set is constructed; wherein the running condition includes normal operation and fault type;

[0039] Step three: generate a model using the training sample set, wherein the data generation model is a multi-discriminator auxiliary classification generative adversarial network model, the real training sample data is input, and the data sample approximating the real data sample is output;

[0040] Step four: use the real training sample set and the pseudo sample data generated by the data generation model as the final training sample set to train the preset fault diagnosis model; wherein the fault diagnosis model is a deep convolutional neural network, and the running data of the robot arm is input, and the running condition is output;

[0041] Step five: collect the real-time running data of the robot arm with fault diagnosis, and input the collected real-time running data into the deployed fault diagnosis model to realize the fault diagnosis model; wherein the deployed fault diagnosis model is a trained model deployed to the web end by flask.

[0042] It should be noted that the most vulnerable part of the robot arm in the irradiation environment is the component containing electronic devices, so the robot arm fault types obtained in the example include driver irradiation damage fault, joint motor irradiation damage fault, encoder irradiation damage fault, and any one or more combinations thereof.

[0043] Based on the above data set, the data is expanded once by the data generation model, the pre-constructed fault diagnosis model is trained by using the mixed training data set, and after the trained fault diagnosis model is deployed, the real-time diagnosis of whether the current mechanical arm fails and the fault type can be realized, and the fault positioning is realized. Among them, the data generation model is a multi-discriminator auxiliary classification generative adversarial network model, which includes three modules of discriminator, generator and classifier, and the model structure diagram is as shown in Figure 3 The discriminator mainly judges whether the data source is real data or false data generated by the generator, the generator extracts data features to fit the distribution and generates pseudo data samples close to the real sample, and the classifier classifies the data, so that the generator can generate multiple types of fault data. Among them, three discriminators are set to avoid the model from collapsing due to the generator fitting to the special conditions of cheating discriminators; at the same time, the Wasserstein distance and Lipschitz penaty are introduced into the loss function to construct a new loss function to stabilize the training of the model; the efficient channel attention mechanism (ECA) is introduced to further improve the feature extraction capability of the model. The multi-discriminator auxiliary classification generative adversarial network will be further introduced below.

[0044] The original generative adversarial neural network (GAN) is a kind of deep learning method composed of two adversarial training models, one is the generator (Generator), and the other is the discriminator (Discriminator). The purpose of the generator is to maximize the fitting of the data distribution of the real data, that is, to generate data similar to the training sample; the goal of the discriminator is to distinguish whether the data is from the real sample or the pseudo data sample generated by the generator, and the two are trained by each other to reach a Nash equilibrium state. The objective function of GAN is as follows:

[0045] (1)

[0046] In the formula: represents the data distribution of the real sample, represents the distribution of the noise sample, E represents the expectation function, is the output result of the discriminator, is the output value of the generator.

[0047] The auxiliary classification generative adversarial neural network (ACGAN) is a variant of GAN, which can output the class to which the generated data belongs when generating data meeting certain conditions. The input noise of the generator embeds label information, so it can generate labeled data; the discriminator can distinguish real and false data sources and distinguish data class information. Since the label information is introduced, the objective function of ACGAN contains two parts, which are as follows:

[0048] (2)

[0049] (3)

[0050] The multi-discriminator assisted classification generative adversarial network (MDACGAN) used in this example is a series of improvements and optimizations based on the ACGAN network. A new loss function is designed to stabilize the ACGAN network under small sample training. To solve the problem of model collapse, a multi-discriminator and single generator adversarial training is constructed, which solves the problem. The design of separating the classifier reduces the redundancy of the model, and also increases the scalability of the model. The objective function of MDACGAN is as follows:

[0051] (4)

[0052] (5)

[0053] (6)

[0054] (7)

[0055] (8)

[0056] In the formula: , and are the loss functions of the discriminator, the generator and the classifier, respectively, , and have the same expression form, and the above formula only gives the expression of , the expressions of and can be similarly obtained; is a proportionality factor; and represent the cross-entropy loss functions of the classification real samples and the generated samples, respectively.

[0057] Wherein, the fault diagnosis model is a deep convolutional neural network model, through multiple convolution layers (Convolution layer), maximum pooling layer (Maxpool), batch normalization (Batch normalization), activation layer (Active layer), discard layer (Dropout layer) for data feature extraction, dimension reduction, etc. The extracted feature map is reduced in dimension through the global average pooling layer (Goabal average pooling), which greatly reduces the network structure parameters, and finally outputs the diagnosis result through the fully connected layer (Fully connected layer). Each layer will be introduced as follows:

[0058] Convolution layer: Convolution layer is the core of convolutional neural network, which learns the features of input data through convolution calculation. According to a certain step, the input data is calculated by sliding, realizing the feature extraction of original data, so as to obtain the feature map. The main characteristics are local connection and weight sharing. The forward propagation of convolutional neural network from layer to layer can be represented by the following formula:

[0059] (9)

[0060] In the formula, represents the output value of layer convolution, represents the selected feature map, represents the output value of layer, represents the weight value of layer, represents the bias value of layer.

[0061] Max pooling layer: the role of max pooling layer is to reduce the dimension of the feature output by convolution layer, so as to reduce the parameters and calculation amount in network. It divides the input area into several rectangular regions, and the output of each sub region is the maximum value, which can keep the feature invariability and improve the generalization ability of model. Its formula is as follows:

[0062] (10)

[0063] In the formula, represents the output value of the th neuron of the th feature map of layer, , is the width of the pooling region, is the value of the corresponding neuron in layer after pooling.

[0064] Batch Normalization Layer: The purpose of the batch normalization layer is to make the feature map meet the distribution rule of mean 0 and variance 1, because the network is more sensitive to data near zero, but as the number of network layers increases, the data will deviate, so using the batch normalization layer can return the data to near zero, thereby accelerating network convergence and stabilizing training. The mathematical expression is as follows:

[0065] (11)

[0066] wherein, represents the i data feature, and the mean and standard deviation of the data in each batch, to prevent the introduction of a very small amount (negligible) of zero, and denote the scaling factor and the offset factor, respectively; is the output of the data feature after batch standardization.

[0067] Activation Layer: The activation layer is to use an activation function to perform a nonlinear transformation on the output value of the convolution calculation, thereby improving the feature learning ability of the network. The activation function selected in this example is the ReLU (Rectified Linear Unit) function, and its calculation formula is:

[0068] (12)

[0069] Dropout Layer: The dropout layer is a training trick in training deep learning networks. By temporarily discarding a certain proportion of neurons in the hidden layer, the synergistic effect between features is reduced, thereby alleviating overfitting and improving the generalization ability of the network.

[0070] Global Average Pooling Layer: The global average pooling layer only needs to perform a convolution operation on multiple convolution kernels after multiple convolution layers and pooling layers, and then perform global average pooling on the formed feature matrix, that is, each column vector is unitized to form a one-dimensional vector.

[0071] Fully Connected Layer: The fully connected layer is equivalent to a classifier, and the input is the output of the global average pooling layer. After processing by the softmax function, classification can be achieved. The output value of the layer fully connected can be calculated by the following formula:

[0072] (13)

[0073] wherein, is output value of a layer, is a weight matrix, is a bias value matrix.

[0074] Suppose the output of the full connection layer is In total n classes, the probability distribution requirement can be met by the softmax function:

[0075] (14)

[0076] In summary, the present example constructs a fault data set by obtaining data of the mechanical arm when it is running normally and data after irradiation damage, adopts a data generation model to further expand the data set, trains a preset fault diagnosis model using the fused data, and then deploys the trained data; collects real-time running data of the mechanical arm, and inputs the collected data into the deployed model, thereby realizing real-time fault diagnosis. This method is an end-to-end fault diagnosis method, which can ensure the diagnosis accuracy while avoiding excessive data processing operations and reducing manual operations. The data generation model of the present application can expand the data set and solve the problem of difficult data acquisition in a radiation environment. The fault diagnosis method of the present example does not rely on the mathematical model of the mechanical arm, but can recognize the fault categories of the mechanical arm by training using the collected fault data. The present application provides a new method for fault diagnosis of a mechanical arm in a radiation environment, provides a basis for the design and reinforcement of a radiation-resistant mechanical arm, and provides state detection for the operation of the mechanical arm in a radiation environment, thereby guiding the subsequent operation and maintenance of the mechanical arm.

[0077] Example 2

[0078] The present example provides a mechanical arm fault diagnosis system in a radiation environment, which includes the following modules:

[0079] A mechanical arm fault data acquisition module is configured to acquire running data of each component of the mechanical arm; wherein the running data includes running data of each component of the mechanical arm when it is normal and running data of the mechanical arm after each component is damaged by irradiation;

[0080] A fault data sample set construction module is configured to construct a fault data sample set using the running data acquired by the mechanical arm fault data acquisition module as sample data and the running conditions of each component of the mechanical arm as sample labels; wherein the running conditions include normal and fault types of each component damage.

[0081] The data generation model training module is configured to train a data generation model by using the training sample set constructed by the fault data sample set module; the data generation model is a multi-discriminator generative adversarial network model, and the data generation model takes the operation data as input and outputs pseudo data samples approximating real samples;

[0082] The fault diagnosis model training module is configured to train a preset fault diagnosis model by using the training sample set constructed by the fault data sample set module and the pseudo sample data set generated by the data generation model; the fault diagnosis model is a deep convolutional neural network model, and the fault diagnosis model takes the operation data of the robot arm as input and outputs the operation state;

[0083] The fault diagnosis module is configured to collect real-time operation data of a robot arm to be diagnosed, and input the collected real-time operation data into the fault diagnosis model trained by the fault diagnosis model training module, so as to realize fault diagnosis.

[0084] The robot arm fault diagnosis system of the present embodiment corresponds to the robot arm fault diagnosis method of the first embodiment; the functions of each functional module of the robot arm fault diagnosis system of the present embodiment correspond to each process step of the robot arm fault diagnosis method of the first embodiment, and thus will not be described here.

[0085] Embodiment 3

[0086] The present embodiment provides a robot arm joint module test platform system for constructing a robot arm fault data acquisition module, and a schematic diagram of the composition of the robot arm joint module test platform is shown in Figure 5 The robot arm test platform includes an encoder, a torque and speed sensor, a reducer, a torque motor, and the like.

[0087] The test platform of the present embodiment can accurately test the performance parameter changes of each component of the robot arm joint module before and after irradiation damage.

[0088] Finally, it should be noted that: the above only describes the preferred embodiments of the present application, and is not used to limit the present application, although the present application has been described in detail with reference to the foregoing embodiments, and for those skilled in the art, the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for diagnosing robotic arm faults under nuclear radiation environment based on deep learning, characterized in that: The method comprises the following steps: S1: obtaining a fault data set of the mechanical arm, wherein the fault data set comprises data collected by each sensor when the mechanical arm is normally operated and data when each part of the mechanical arm appears a fault after being irradiated; S2: using the above data as fault data samples and the operation conditions of each component of the corresponding mechanical arm as sample labels to construct a training sample set; wherein the operation conditions comprise normal operation and fault types; S3: training a data generation model using the training sample set; wherein the data generation model is a multi-discriminator auxiliary classification generative adversarial network model, the data generation model takes real data samples as input and approximates real data samples as output; S4: using real fault data samples and pseudo data samples generated by the data generation model as mixed final training samples to train a preset fault diagnosis model; wherein the fault diagnosis model is a deep convolutional neural network model, the fault diagnosis model takes mechanical arm operation data as input and operation conditions as output; S5: collecting real-time operation data of a mechanical arm to be fault diagnosed and inputting the collected real-time operation data into the deployed fault diagnosis model to realize the fault diagnosis model; wherein the deployed fault model is a trained model deployed to a web page by flask.

2. The method according to claim 1, wherein the method is characterized in that: The fault types comprise driver irradiation damage faults, joint motor irradiation damage faults, encoder damage faults, and combinations of any one or more thereof.

3. The method of claim 1, wherein the method comprises: The data generation model in S3 comprises three discriminator modules, one generator module, and one classifier module.

4. The method according to claim 3, wherein the method is characterized in that: The discriminator module determines whether the data source is real data or pseudo data generated by the generator, the generator module extracts data features to fit the distribution and generates pseudo data samples approximating real samples, and the classifier module classifies the data.

5. The method of claim 1, wherein the method is based on deep learning. The fault diagnosis model in S4 comprises multiple convolution modules, one global average pooling layer, and one fully connected layer.

6. The method of claim 5, wherein the method further comprises: The convolution module comprises multiple convolution kernels, one batch normalization layer, and one activation layer.

7. The method according to claim 6, wherein the method is characterized in that: The convolution kernel size is 3x3, the number is 64 or 128 or 256 or 512, and the activation function of the activation layer is a ReLu function.

8. The method according to claim 7, wherein the method is characterized in that: In the fault diagnosis model, the number of convolution kernels is doubled after passing through two convolution modules each time.

9. The method of claim 1-8, wherein the method is based on deep learning. The nuclear radiation environment-based mechanical arm fault diagnosis method based on deep learning provides a mechanical arm fault diagnosis system, which comprises a mechanical arm fault data acquisition module, a fault data sample set construction module, a data generation model training module, a fault diagnosis model training module, and a fault diagnosis module.

10. The method of claim 1-8, wherein the method is based on deep learning. The nuclear radiation environment-based mechanical arm fault diagnosis method based on deep learning provides a mechanical arm joint module test platform, which is used to constitute the mechanical arm fault data acquisition module. The mechanical arm joint module test platform comprises an encoder, a torque and speed sensor, a reducer, and a torque motor.

Images