Intelligent inspection and fault diagnosis system for nuclear power robot

Abstract

The invention relates to the technical field of nuclear power robot maintenance, and discloses an intelligent inspection and fault diagnosis system for a nuclear power robot, which integrates a high-precision sensor network, captures operation environment data of nuclear power facilities in real time, and constructs a rich training data set in combination with operation and maintenance records and historical fault data. Through a machine learning algorithm, a nuclear power facility inspection model and a fault diagnosis model are systematically trained, and automatic identification of abnormal conditions, preliminary classification of fault types and deep analysis of specific reasons and positions of faults are realized. Meanwhile, the system innovatively constructs an error correction model, and further optimizes the accuracy of inspection and fault diagnosis. And the data analysis unit automatically generates a detailed inspection report and a fault diagnosis report by using the trained model, thereby providing powerful technical support for safe and stable operation of the nuclear power station.

Description

Technical Field

[0001] The present invention relates to the technical field of nuclear power robot maintenance, and in particular to a nuclear power robot intelligent inspection and fault diagnosis system. Background Art

[0002] In the nuclear power industry, ensuring the safe and stable operation of nuclear power plants is of vital importance. As a highly complex and dangerous industrial environment, the operation and maintenance management of nuclear power plants requires extremely high professionalism and precision. Traditional nuclear power facility inspections and fault diagnosis mainly rely on manual work, which is not only inefficient but also poses extremely high safety risks. Manual inspections are difficult to achieve comprehensive coverage, and it is often difficult to conduct effective inspections in some hidden or inaccessible areas. At the same time, manual fault diagnosis relies on experience and judgment, which is easily affected by personal ability and cognitive limitations, resulting in inaccurate or missed diagnostic results.

[0003] With the advancement of science and technology, especially the rapid development of sensor technology and artificial intelligence technology, new possibilities have been provided for the intelligent operation and maintenance of the nuclear power industry. Sensor networks can capture the operating environment data of nuclear power facilities in real time, including radiation intensity, temperature, humidity, vibration, sound, and visual information, providing a data basis for comprehensive monitoring of the status of nuclear power facilities. However, how to efficiently integrate and utilize these massive data to achieve accurate inspections and fault diagnosis has become an urgent problem to be solved.

[0004] Although existing nuclear power facility monitoring systems can collect data in real time and conduct a certain degree of monitoring, they often lack intelligent analysis and diagnosis capabilities. They can usually only provide simple alarms or abnormal indications, but cannot accurately locate faults and analyze their causes. In addition, existing systems often ignore the correction of model prediction errors, resulting in limited accuracy of diagnostic results. Summary of the invention

[0005] The purpose of the present invention is to provide a nuclear power robot intelligent inspection and fault diagnosis system to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: a nuclear power robot intelligent inspection and fault diagnosis system, the system comprising:

[0007] Data acquisition unit: used to capture operating environment data in real time through a sensor network, including radiation intensity sensors, temperature sensors, humidity sensors, vibration sensors, sound sensors, and visual sensors; at the same time, it collects operation and maintenance records of nuclear power facilities, historical fault records and their solutions to form annotated fault data;

[0008] A data processing unit, used to integrate and pre-process the real-time collected data and the annotated fault data to construct a training data set;

[0009] Model development unit, used to establish nuclear power facility inspection model and fault diagnosis model. The inspection model is used to identify abnormal conditions of nuclear power facilities and preliminarily classify fault types, while the fault diagnosis model deeply analyzes current and historical data to determine the specific cause and location of the fault. The inspection model and fault diagnosis model are trained based on the training data set using machine learning algorithms, and the prediction errors of the two models are further combined to build an error correction model.

[0010] The data analysis unit uses the trained model and combines the real-time operation data of nuclear power facilities to automatically generate inspection reports and fault diagnosis reports.

[0011] Preferably, a convolutional neural network (CNN) algorithm is used to construct an inspection model, and deep features of environmental data are extracted through multi-layer convolution and pooling operations to achieve recognition and classification of abnormal conditions of nuclear power facilities.

[0012] Preferably, the training process of the inspection model includes:

[0013] Step 1: Design the CNN architecture: determine the number of convolutional layers, convolution kernel size, pooling method, and fully connected layer structure;

[0014] Step 2: Model training: Input the preprocessed training data set and optimize the model parameters through the back propagation algorithm;

[0015] Step 3: Model verification and optimization: Use the cross-validation strategy to evaluate the recognition accuracy of the model, and adjust the network structure and learning rate based on the F1 score and accuracy index.

[0016] Preferably, a long short-term memory network (LSTM) algorithm is used to construct a fault diagnosis model, and the memory capacity of LSTM is used to process time series data to predict the development trend and specific location of the fault.

[0017] Preferably, the training process of the fault diagnosis model includes:

[0018] Step 1: Feature engineering: extract time series features from the training data set, including trend features and periodic features;

[0019] Step 2: Build the LSTM model: set the number of LSTM units, the number of layers, and the dropout ratio;

[0020] Step 3: Model training: Input the time series features into the LSTM model and train the model through sequence-to-sequence learning mode;

[0021] Step 4: Model evaluation and tuning: Evaluate the prediction accuracy and stability of the model on the validation set, and adjust the hyperparameters of the LSTM model based on the results, including increasing the number of layers or adjusting the dropout ratio.

[0022] Preferably, the steps of constructing the error correction model include:

[0023] Step 1: Calculate the difference between the predicted value and the true value of the inspection model and the fault diagnosis model, which are defined as the inspection error and the diagnosis error respectively;

[0024] Step 2: Combine the inspection error, diagnosis error and original features to form an error feature vector;

[0025] Step 3: Use the XGBoost algorithm to train the error feature vector and build an error correction model;

[0026] Step 4: The error correction model consists of an input layer, an integration layer consisting of multiple decision trees, and an output layer. The input layer receives the error feature vector, and the output layer outputs the corrected inspection and fault diagnosis results, and continuously reduces the prediction error through the iterative improvement mechanism of XGBoost.

[0027] Preferably, during the fault diagnosis model training process, the mean square error (MSE) algorithm is used to evaluate the model performance, and the specific implementation steps include:

[0028] Step 1: After each training iteration, use the validation set data to evaluate the performance of the LSTM model.

[0029] Step 2: Calculate the mean square error (MSE) between the predicted output of the LSTM model and the actual output of the validation set. The calculation formula for MSE is:

[0030]

[0031] Among them, N is the number of samples in the validation set, y i is the actual output value of the i-th sample in the validation set, y i is the predicted output value of the LSTM model for the i-th sample;

[0032] Step 3: Evaluate the performance of the LSTM model based on the calculated MSE value. The smaller the MSE value, the higher the prediction accuracy of the model and the better the performance.

[0033] Step 4: Based on the changing trend of the MSE value and the preset performance threshold, decide whether to continue training the model, adjust the model parameters, or stop training early.

[0034] Preferably, the data processing unit pre-processes the collected data by means of data cleaning and standardization.

[0035] Preferably, the method of cleaning the data includes:

[0036] For missing values, linear interpolation algorithm is used to fill them;

[0037] For outliers, the threshold judgment method is used to regard data points that exceed the preset threshold range as outliers and replace them with the mean of their neighboring points.

[0038] Preferably, the method for standardizing the data includes: using the Z-score standardization method to standardize the continuous features, specifically: subtracting the mean of each numerical data and dividing it by its standard deviation, so that the processed data conforms to the standard normal distribution, and the standardization formula is: Z = (X-μ) / σ, wherein X is the original data, μ is the mean of the original data, σ is the standard deviation of the original data, and Z is the standardized data.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] Improve inspection efficiency and comprehensiveness: Through the sensor network deployed by the data acquisition unit, the system can capture the operating environment data of nuclear power facilities in real time and comprehensively, including radiation intensity, temperature, humidity, vibration, sound and visual information, etc., and realize all-round and no-dead-angle monitoring of nuclear power facilities. This greatly improves the efficiency of inspections and ensures the comprehensiveness of inspections, especially for the detection of areas that are difficult to access or hidden by humans.

[0041] Enhance the accuracy and precision of fault diagnosis: The system integrates a data processing unit and a model development unit, which can efficiently integrate real-time collected data and annotated fault data to build a rich training data set. Using machine learning algorithms, the system trains nuclear power facility inspection models and fault diagnosis models, which can automatically identify abnormal conditions, preliminarily classify fault types, and conduct in-depth analysis to determine the specific cause and location of the fault. This intelligent analysis and diagnosis capability significantly improves the accuracy and precision of fault diagnosis and reduces errors and omissions caused by manual experience judgment.

[0042] Introducing an error correction mechanism to improve model reliability: In the model development unit, the present invention further combines the prediction errors of the inspection model and the fault diagnosis model to construct an error correction model. This mechanism can dynamically adjust and optimize the model prediction results, effectively reducing the uncertainty of the model prediction and improving the reliability and stability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a working principle diagram of the nuclear power robot intelligent inspection and fault diagnosis system of the present invention;

[0044] Figure 2 A flowchart for building an inspection model;

[0045] Figure 3 Flowchart for the construction of the error correction model. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] See also Figure 1-3 The present invention provides a technical solution: a nuclear power robot intelligent inspection and fault diagnosis system, the system comprising:

[0048] Data acquisition unit: The data acquisition unit consists of a series of high-precision sensors that are deployed in key parts of nuclear power facilities, including but not limited to the reactor area, cooling system, electrical control system, etc. Radiation intensity sensors are used to monitor radiation levels, temperature sensors and humidity sensors are used to monitor ambient temperature and humidity, vibration sensors are used to detect the vibration of equipment, sound sensors are used to capture abnormal sounds, and visual sensors are used to obtain intuitive image information of the facility. At the same time, the unit is also responsible for collecting operation and maintenance records, historical fault records and their solutions from the operation and maintenance management system of nuclear power facilities. These data will be used to annotate fault data and provide supervision information for subsequent model training.

[0049] Data processing unit: The data processing unit is responsible for integrating and preprocessing the real-time data collected by the data acquisition unit with the annotated fault data. First, the raw data is cleaned to remove invalid or erroneous data points. Then, the data is normalized or standardized to ensure the comparability between different sensor data. Next, the processed data is aligned with the annotated fault data to form a training data set containing feature vectors and corresponding fault labels. In addition, this unit is also responsible for the division of the data set into training set, validation set and test set for subsequent model training and evaluation.

[0050] Model development unit: The model development unit is the core part of this system. It is responsible for establishing the inspection model and fault diagnosis model of nuclear power facilities. The inspection model uses machine learning algorithms to train the training data set to identify abnormal conditions of nuclear power facilities and preliminarily classify the fault type. The fault diagnosis model uses a more complex deep learning algorithm to conduct in-depth analysis of time series data to determine the specific cause and location of the fault. During the model training process, the cross-validation strategy is used to evaluate the model performance, and the model parameters are adjusted based on the evaluation results. In addition, the unit also combines the prediction errors of the inspection model and the fault diagnosis model to construct an error correction model to further improve the accuracy of diagnosis. The error correction model can be trained using algorithms such as gradient boosting tree (GBDT).

[0051] Data analysis unit: The data analysis unit is responsible for automatically generating inspection reports and fault diagnosis reports using the trained model and real-time operation data of nuclear power facilities. The unit first obtains real-time data through the data acquisition unit and inputs it into the inspection model and fault diagnosis model. Then, based on the output of the model, a report containing abnormal condition description, preliminary classification of fault type, fault cause and location analysis is generated. Finally, the report is presented to the operation and maintenance personnel in a visual or textual form so that they can take timely measures to deal with it.

[0052] The present invention will be further described below in conjunction with Examples 1 to 3:

[0053] Embodiment 1:

[0054] In the nuclear power robot intelligent inspection system of the present invention, the construction of the inspection model is based on the convolutional neural network (CNN) algorithm, which aims to effectively extract the deep features of environmental data through multi-layer convolution and pooling operations, thereby realizing accurate identification and classification of abnormal conditions of nuclear power facilities. The following is a specific implementation method of the inspection model training process:

[0055] Step 1: Design CNN architecture

[0056] First, the CNN network architecture is designed based on the characteristics of nuclear power facility environmental data and the requirements of inspection tasks. Specifically, it includes but is not limited to the determination of the following key parameters:

[0057] Number of convolutional layers and kernel size: Initially set a number of convolutional layers, each layer uses a kernel of different sizes to capture features of different scales. For example, the first layer kernel may be larger to capture global features, while subsequent layers use smaller kernels to refine local features.

[0058] Pooling method: Choose an appropriate pooling operation, such as max pooling or average pooling, to reduce the size of the feature map while retaining important information. This helps reduce the amount of computation and prevent overfitting.

[0059] Fully connected layer structure: After the convolution and pooling layers, a fully connected layer is designed to map the extracted features to the classification space. The number of fully connected layers and the number of neurons are adjusted according to the specific task to ensure that the model has sufficient expressive power.

[0060] Step 2: Model training

[0061] Data preparation: The raw data obtained by the data acquisition unit is preprocessed, including normalization, standardization, data enhancement and other steps, to generate a training data set suitable for CNN input.

[0062] Input data: Input the preprocessed training dataset into the designed CNN model.

[0063] Back propagation and parameter optimization: The back propagation algorithm is used to calculate the gradient according to the loss function (such as cross entropy loss), and the model parameters are updated through the optimizer (such as Adam, SGD, etc.), so that the model gradually learns the ability to distinguish different abnormal states.

[0064] Step 3: Model validation and optimization

[0065] Cross-validation strategy: A K-fold cross-validation strategy is used to divide the training data set into K subsets, and one of the subsets is used as the validation set and the rest as the training set in turn to evaluate the generalization ability of the model.

[0066] Evaluation indicators: F1 score and precision are selected as the main evaluation indicators. F1 score is the harmonic mean of precision and recall, which can comprehensively reflect the classification performance of the model; precision measures the proportion of samples predicted by the model as positive that are actually positive.

[0067] Adjustment of network structure and learning rate: Analyze the performance bottleneck of the model based on the results of cross-validation. If the model is overfitting, consider reducing the number of convolutional layers, increasing the regularization term, or reducing the learning rate; if the model is underfitting, consider increasing the number of convolutional layers, adjusting the convolution kernel size, or increasing the learning rate. Continuously iterate and adjust until satisfactory recognition accuracy and generalization ability are achieved.

[0068] Suppose that during the inspection of nuclear power facilities, it is necessary to identify abnormal temperature rise in the reactor area. First, when designing the CNN architecture, three convolutional layers may be selected, with convolution kernel sizes of 7x7, 5x5, and 3x3, respectively, and each layer is followed by a maximum pooling operation. The fully connected layer is set to two layers, each containing 128 neurons. In the model training phase, the preprocessed temperature sensor data is used as input, and the model is optimized by the back propagation algorithm. In the model verification and optimization phase, a 5-fold cross-validation is used. According to the F1 score and accuracy index, the network structure (such as increasing the number of convolutional layers, adjusting the convolution kernel size) and the learning rate (such as gradually reducing from 0.001 to 0.0001) are continuously adjusted, and finally an inspection model that can accurately identify abnormal temperature rise in the reactor area is obtained.

[0069] Embodiment 2:

[0070] In the fault diagnosis of nuclear power facilities, the long short-term memory network (LSTM) algorithm is used to build a fault diagnosis model to fully utilize the memory capacity of LSTM in processing time series data, predict the development trend of the fault and determine the specific location. The following is the specific implementation of the fault diagnosis model training process:

[0071] Step 1: Feature Engineering

[0072] Time series feature extraction: Extract key time series features from the training data set, including but not limited to trend features and periodic features. Trend features reflect the changing trend of data over time, such as the gradual increase in temperature or the gradual increase in vibration amplitude, etc.; periodic features capture possible periodic changes in the data, such as the periodic fluctuations in the operation of certain equipment.

[0073] Feature selection and transformation: According to the fault type and target, the most representative features are selected, and the expressiveness of the features is enhanced through appropriate transformations (such as difference, logarithmic transformation, etc.).

[0074] In the cooling system of a nuclear power facility, a fault may manifest as an abnormal increase in the cooling water temperature. From the data collected by the temperature sensor, the temperature change trend over time can be extracted as a feature, while considering periodic features such as daily cycles (such as temperature changes during the day and night) and weekly cycles (such as load changes between weekdays and weekends).

[0075] Step 2: Build LSTM model

[0076] Set the number of LSTM units: Determine the number of LSTM units based on the complexity of the task and the size of the data. More units may improve the expressiveness of the model, but also increase the computational cost and the risk of overfitting.

[0077] Determine the number of layers: Multi-layer LSTM can capture more complex temporal dependencies, but it may also make training difficult. Usually start with a single layer and gradually increase the number of layers as needed.

[0078] Set dropout ratio: Introduce dropout between LSTM layers to prevent overfitting. The dropout ratio determines the proportion of neurons that are randomly dropped during training.

[0079] Step 3: Model training

[0080] Data preparation: Organize the time series features obtained by feature engineering into input sequences in chronological order and prepare corresponding labels (such as fault type or location).

[0081] Sequence-to-sequence learning: Using a sequence-to-sequence learning model, the input sequence is fed into the LSTM model, and the model parameters are optimized through the back propagation algorithm and the time back propagation algorithm (BPTT), so that the model can predict the development trend and specific location of the fault.

[0082] Step 4: Model evaluation and tuning

[0083] Validation set evaluation: Evaluate the prediction accuracy and stability of the model on an independent validation set. Accuracy measures the correctness of the model in predicting the fault type, while stability reflects the consistency of the model's performance under different data conditions. Use the mean square error (MSE) algorithm to evaluate model performance. The specific implementation steps include:

[0084] Step 1: After each training iteration, use the validation set data to evaluate the performance of the LSTM model.

[0085] Step 2: Calculate the mean square error (MSE) between the predicted output of the LSTM model and the actual output of the validation set. The calculation formula for MSE is:

[0086]

[0087] Among them, N is the number of samples in the validation set, y i is the actual output value of the i-th sample in the validation set, is the predicted output value of the LSTM model for the i-th sample.

[0088] Step 3: Evaluate the performance of the LSTM model based on the calculated MSE value. The smaller the MSE value, the higher the prediction accuracy of the model and the better the performance.

[0089] Step 4: Based on the changing trend of the MSE value and the preset performance threshold, decide whether to continue training the model, adjust the model parameters, or stop training early.

[0090] Hyperparameter adjustment: According to the evaluation results on the validation set, the hyperparameters of the LSTM model are adjusted, including increasing the number of layers, adjusting the number of LSTM units, changing the dropout ratio, etc., to further improve the performance of the model.

[0091] In the fault diagnosis of the cooling system, the trend characteristics and periodic characteristics of the temperature time series are first extracted through feature engineering. Then, a model with two layers of LSTM is constructed, each layer has 128 LSTM units, and a dropout ratio of 0.5 is set between the layers. In the model training stage, the sequence-to-sequence learning mode is used to input the feature sequence into the model, and the model parameters are optimized to minimize the prediction error. In the model evaluation and tuning stage, by evaluating the accuracy and stability of the model on the validation set, it is found that adding a layer of LSTM and adjusting the dropout ratio to 0.3 can further improve the performance of the model. Finally, an LSTM fault diagnosis model that can accurately predict the development trend and specific location of cooling system faults is obtained.

[0092] Embodiment 3:

[0093] In the inspection and fault diagnosis of nuclear power facilities, in order to further improve the accuracy of prediction, an error correction model is constructed. The model calculates the prediction errors of the inspection model and the fault diagnosis model, combines the original features, and uses the XGBoost algorithm for training to achieve the correction of the prediction error. The following is the specific implementation of the steps for building the error correction model:

[0094] Step 1: Calculate the prediction error

[0095] Inspection error calculation: First, use the trained inspection model to predict the inspection of nuclear power facilities and obtain the predicted value. Then, compare the predicted value with the true value (data obtained through actual inspection) and calculate the difference between them, which is defined as the inspection error.

[0096] Calculation of diagnostic error: Similarly, the fault diagnosis model is used to predict the fault diagnosis of nuclear power facilities and obtain the prediction results. The prediction results are compared with the actual fault conditions and the diagnostic error is calculated.

[0097] Assume that in a certain area of ​​a nuclear power facility, the inspection model predicts a temperature of 50°C, while the actual measured temperature is 52°C, and the inspection error is -2°C. Similarly, the fault diagnosis model predicts that the probability of a cooling system failure in the area is 80%, while the actual failure probability is 90%, and the diagnosis error is -10%.

[0098] Step 2: Form the error feature vector

[0099] Combining original features with errors: The calculated inspection errors and diagnostic errors are combined with the original features (such as temperature, pressure, vibration, etc.) to form a new feature vector, called the error feature vector. This vector contains both the information of the original data and the information of the model prediction error.

[0100] In the above example, the original features may include temperature, pressure, vibration, etc. These features are combined with the inspection error (-2°C) and the diagnosis error (-10%) to form an error feature vector containing multiple dimensions.

[0101] Step 3: Build the Error Correction Model

[0102] Select the XGBoost algorithm: Since the XGBoost algorithm performs well in processing complex features and large-scale data and has strong generalization ability, XGBoost is selected as the algorithm for building the error correction model.

[0103] Training model: Use the error feature vector as input and the true value (inspection results and fault diagnosis results) as output to train the XGBoost model. Minimize the prediction error by continuously adjusting model parameters such as learning rate, maximum depth, sub-sample ratio, etc.

[0104] Step 4: Model structure and iterative improvement

[0105] Model structure: The error correction model consists of an input layer, an integration layer consisting of multiple decision trees, and an output layer. The input layer is responsible for receiving the error feature vector; the integration layer consists of multiple decision trees, each of which classifies or regresses the input data; the output layer outputs the corrected inspection and fault diagnosis results.

[0106] Iterative Boosting: XGBoost uses an iterative boosting mechanism, which means that each round of training will be adjusted based on the prediction results of the previous round to gradually reduce the prediction error. Through multiple rounds of iterations, the model can continuously learn and optimize, and finally obtain more accurate prediction results.

[0107] During the training process, the error feature vector is input into the XGBoost model. Through multiple rounds of iterative training, the model gradually learns how to correct the prediction results based on the error feature vector. Finally, the performance of the model is verified on the test set, and it is found that the accuracy of the inspection and fault diagnosis results after error correction has been significantly improved.

[0108] Data preprocessing mainly includes two steps: data cleaning and standardization. The following is a detailed description of the specific implementation methods of these two steps:

[0109] Missing value processing: Traverse each field in the data set and mark the missing data points. For the data points marked as missing, use the linear interpolation algorithm to fill them. Specifically, if a data point is missing, the estimated value of the missing point is calculated by the linear interpolation formula based on the values ​​of the two non-missing data points before and after it, and used to fill it.

[0110] Assume that in the temperature time series data, the temperature value at a certain moment is missing, and the temperature before and after it is 20℃ and 24℃ respectively. If the time interval between these two moments is equal, the temperature value of the missing point can be calculated by linear interpolation as 22℃.

[0111] Outlier processing: According to the normal operating parameter range of nuclear power facilities, a reasonable threshold range is set for each feature. The data set is traversed, and data points that exceed the preset threshold range are considered outliers. Then, the outlier is replaced with the mean of several non-outlier data points adjacent to the outlier (such as one before and one after).

[0112] In the pressure sensor data, if the pressure value of a data point far exceeds the normal working pressure range (such as set to 10-100MPa), such as reaching 150MPa, it is considered an abnormal value. If the pressure of the previous and next data points is 90MPa and 95MPa respectively, the average of these two points (92.5MPa) can be used to replace the abnormal value.

[0113] Data standardization: Z-score standardization is used. For each continuous feature, its mean and standard deviation in the entire data set are first calculated.

[0114] Apply the standardization formula: Then, subtract the mean from the original value of each data point and divide it by its standard deviation to get the standardized data. The standardization formula is: Z = (X-) / σ, where X

[0115] μ

[0116] is the original data, μ is the mean of the original data, σ is the standard deviation of the original data, and Z is the standardized data.

[0117] Assume that the mean value of a feature (such as temperature) in a data set is 25°C and the standard deviation is 5°C. For a data point in the feature, its original value is 30°C. After Z-score standardization, the value of the data point is: Z = (30-25) / 5 = 1.

[0118] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0119] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. Nuclear power robot intelligent inspection and fault diagnosis system, characterized by: The system comprises: Data acquisition unit: used to capture operating environment data in real time through a sensor network, including radiation intensity sensors, temperature sensors, humidity sensors, vibration sensors, sound sensors, and visual sensors; at the same time, it collects operation and maintenance records of nuclear power facilities, historical fault records and their solutions to form annotated fault data; A data processing unit, used to integrate and pre-process the real-time collected data and the annotated fault data to construct a training data set; Model development unit, used to establish nuclear power facility inspection model and fault diagnosis model. The inspection model is used to identify abnormal conditions of nuclear power facilities and preliminarily classify fault types, while the fault diagnosis model deeply analyzes current and historical data to determine the specific cause and location of the fault. The inspection model and fault diagnosis model are trained based on the training data set using machine learning algorithms, and the prediction errors of the two models are further combined to build an error correction model. The data analysis unit uses the trained model and combines the real-time operation data of nuclear power facilities to automatically generate inspection reports and fault diagnosis reports.

2. The nuclear power robot intelligent inspection and fault diagnosis system according to claim 1 is characterized by: The convolutional neural network (CNN) algorithm is used to build an inspection model. Through multi-layer convolution and pooling operations, the deep features of environmental data are extracted to realize the recognition and classification of abnormal conditions of nuclear power facilities.

3. The nuclear power robot intelligent inspection and fault diagnosis system according to claim 2 is characterized in that: The training process of the inspection model includes: Step 1: Design the CNN architecture: determine the number of convolutional layers, convolution kernel size, pooling method, and fully connected layer structure; Step 2: Model training: Input the preprocessed training data set and optimize the model parameters through the back propagation algorithm; Step 3: Model verification and optimization: Use the cross-validation strategy to evaluate the recognition accuracy of the model, and adjust the network structure and learning rate based on the F1 score and accuracy index.

4. The nuclear power robot intelligent inspection and fault diagnosis system according to claim 1 is characterized in that: The long short-term memory network (LSTM) algorithm is used to build a fault diagnosis model, and the memory capacity of LSTM is used to process time series data to predict the development trend and specific location of the fault.

5. The nuclear power robot intelligent inspection and fault diagnosis system according to claim 4 is characterized in that: The training process of the fault diagnosis model includes: Step 1: Feature engineering: extract time series features from the training data set, including trend features and periodic features; Step 2: Build the LSTM model: set the number of LSTM units, the number of layers, and the dropout ratio; Step 3: Model training: Input the time series features into the LSTM model and train the model through sequence-to-sequence learning mode; Step 4: Model evaluation and tuning: Evaluate the prediction accuracy and stability of the model on the validation set, and adjust the hyperparameters of the LSTM model based on the results, including increasing the number of layers or adjusting the dropout ratio.

6. The nuclear power robot intelligent inspection and fault diagnosis system according to claim 5 is characterized in that: The steps to construct the error correction model include: Step 1: Calculate the difference between the predicted value and the true value of the inspection model and the fault diagnosis model, which are defined as the inspection error and the diagnosis error respectively; Step 2: Combine the inspection error, diagnosis error and original features to form an error feature vector; Step 3: Use the XGBoost algorithm to train the error feature vector and build an error correction model; Step 4: The error correction model consists of an input layer, an integration layer consisting of multiple decision trees, and an output layer. The input layer receives the error feature vector, and the output layer outputs the corrected inspection and fault diagnosis results, and continuously reduces the prediction error through the iterative improvement mechanism of XGBoost.

7. The nuclear power robot intelligent inspection and fault diagnosis system according to claim 6 is characterized by: During the fault diagnosis model training process, the mean square error (MSE) algorithm is used to evaluate the model performance. The specific implementation steps include: Step 1: After each training iteration, use the validation set data to evaluate the performance of the LSTM model. Step 2: Calculate the mean square error (MSE) between the predicted output of the LSTM model and the actual output of the validation set. The calculation formula for MSE is: Among them, N is the number of samples in the validation set, y i is the actual output value of the i-th sample in the validation set, is the predicted output value of the LSTM model for the i-th sample; Step 3: Evaluate the performance of the LSTM model based on the calculated MSE value. The smaller the MSE value, the higher the prediction accuracy of the model and the better the performance. Step 4: Based on the changing trend of the MSE value and the preset performance threshold, decide whether to continue training the model, adjust the model parameters, or stop training early.

8. The nuclear power robot intelligent inspection and fault diagnosis system according to claim 1 is characterized in that: The data processing unit pre-processes the collected data in a manner including data cleaning and standardization.

9. The nuclear power robot intelligent inspection and fault diagnosis system according to claim 8 is characterized in that: Methods for cleaning data include: For missing values, linear interpolation algorithm is used to fill them; For outliers, the threshold judgment method is used to regard data points that exceed the preset threshold range as outliers and replace them with the mean of their neighboring points.

10. The nuclear power robot intelligent inspection and fault diagnosis system according to claim 8, characterized in that: The method of standardizing the data includes: using the Z-score standardization method to standardize the continuous features, specifically: subtracting the mean of each numerical data and dividing it by its standard deviation, so that the processed data conforms to the standard normal distribution. The standardization formula is: Z = (X-μ) / σ, where X is the original data, μ is the mean of the original data, σ is the standard deviation of the original data, and Z is the standardized data.

Images