A method for detecting the position state of a valve core of an accident pressure distribution valve of a hydro-turbine governor

Abstract

The application discloses a kind of water turbine set governor accident pressure distribution valve valve core position state detection methods, it includes the following steps: step 1: pressure distribution valve valve core position state data sampling;Step 2: data preprocessing;Step 3: feature coding;Step 4: reconstruct sample feature acquisition;Step 5: model training;Step 6: classification detection;Step 7: system development;The application is by to the pressure distribution valve sample feature coding and the construction of reconstruct model, respectively set coding, decoding and discriminator and several loss functions such as judge each other, reach optimal effect;According to the coding of input sample feature, take deconvolution operation to realize the transformation of sample data, increase the nonlinearity of state sample data, increase the robustness of model, effectively reduce the influence of noise label on model training.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of the accident pressure distribution valve of the speed regulator of the hydraulic turbine unit, and particularly relates to a valve core position state detection method of the accident pressure distribution valve of the speed regulator of the hydraulic turbine unit. BACKGROUND

[0002] The accident pressure distribution valve of the speed regulator of the hydraulic turbine unit is an overspeed protection device for the hydraulic turbine generator unit of the hydropower station. At present, the action condition of the accident pressure distribution valve of the hydropower station is collected by using the feedback contact point of the travel switch, which is installed on the accident pressure distribution valve. In the traditional working environment of the hydropower station, the identification of the state of the accident pressure distribution valve mainly depends on manual observation. However, due to the insufficient light condition and the subtle state change of the accident pressure distribution valve in the actual environment, it is difficult to effectively meet the actual production demand and the precision requirement by only relying on the manual observation method. At present, in order to improve the accuracy of the identification of the action of the equipment and record the full stroke record and analysis of the equipment in the slow change process, the artificial intelligence technology based on visual operation has attracted more attention in the detection method of the micro-motion track of the equipment of the hydropower station. Different from the traditional observation of the state change of the accident pressure distribution valve by the naked eye, the state detection technology of the accident pressure distribution valve based on computer vision relies more on artificial intelligence and deep learning. Through the training and optimization of the model, the automatic state detection and judgment are realized. The existing state recognition method of the accident pressure distribution valve based on vision mainly includes the following steps: (1) feature extraction of the state sample of the accident pressure distribution valve; (2) model update iteration; (3) state classification detection of the accident pressure distribution valve. Through the above steps, the labor and time cost required in the state detection process of the accident pressure distribution valve is reduced to a certain extent. The disadvantage is that in the feature extraction and model optimization process, due to the small number of state change types of the accident pressure distribution valve and the poor on-site working environment, noise is easily generated, which affects the accuracy of the overall state judgment. One of the reasons for the generation of noise is that in the environment of the accident pressure distribution valve, the differences in the view point, brightness and scale are caused due to the change of the visual angle and the light, which reduces the state discrimination and recognition performance of the accident pressure distribution valve. At present, the state detection method of the accident pressure distribution valve based on various visual features has attracted widespread attention. However, due to the particularity of the change of the accident pressure distribution valve, there is still room for improvement in the existing method.

[0003] At present, the existing action feedback for the similar accident pressure distribution valve mainly relies on the sensor equipment and the related equipment operation technology to detect the accident pressure distribution valve (micro-motion switch). The position travel switch is installed at the lower end of the valve core indicating rod of the accident pressure distribution valve. When the valve core indicating rod falls down, the cam of the position travel switch is impacted, and the contact point of the position travel switch is actuated. The specific existing related research technical scheme is as follows.

[0004] CN220418755U proposes a micro switch airtightness detection equipment, including shell, air cylinder, etc. A plurality of physical devices ensure the stability of the detection of micro switch, and ensure the stability of the work of micro switch after detection.

[0005] CN220340353U proposes a detection device for micro switch, which includes a base, a support, a micro switch and a trigger part. The support is connected with the base. The micro switch is installed on the support. The trigger part is installed on the support and used for triggering the micro switch.

[0006] The existing technology is a contact type reaction working principle of position travel switch. The travel switch must rely on the mutual contact of the contact and the detected object to work. After a long time of use, the target object and the travel switch will cause certain damage to the travel switch itself during the impact process. And the installation position of the travel switch is narrow, and the later maintenance and verification are difficult. The specific manifestations are as follows:

[0007] (1) The new device mainly relies on sensor equipment, which has high installation cost and is prone to line damage, resulting in identification accuracy error and poor reliability. It is difficult to maintain subsequently.

[0008] (2) In addition, based on the traditional machine vision method, although it has great improvement compared with the manual observation mode, due to the single data of the accident pressure regulating valve state change, it is difficult to collect diversified and sufficient data, and it is difficult to effectively train the model, resulting in insufficient prediction accuracy. In addition, due to the difference of view point, light change, view point, light and scale in the water and electricity station dam, it is difficult to construct the machine learning model, and the traditional machine learning method is difficult to effectively complete the accurate identification of the pressure regulating valve state. SUMMARY

[0009] The purpose of the present application is to overcome the above-mentioned shortcomings, and to provide a water turbine unit governor accident pressure regulating valve core position state detection method to solve the problems in the background art.

[0010] To solve the above technical problems, the technical solution adopted by the present application is: a water turbine unit governor accident pressure regulating valve core position state detection method, which comprises the following steps:

[0011] Step 1: sampling of pressure regulating valve core position state data;

[0012] Step 2: data preprocessing;

[0013] Step 3: feature encoding;

[0014] Step 4: Reconstructing sample feature acquisition;

[0015] Step 5: Model training;

[0016] Step 6: Classification detection;

[0017] Step 7: System development.

[0018] Preferably, the step 1 is specifically: through the acquisition device, the position state data of the pressure regulating valve core is photographed and collected, and different types of position state data of the pressure regulating valve core are obtained; the pressure regulating valve core position state sample data set is divided into multiple batches (batch), wherein each batch contains a plurality of state sample images; in subsequent model training, a plurality of sample data (images) in different states are selected from the multiple batches in turn, and then the data is trained; the state data set collected by this operation is represented as , wherein D is represented as a state data set collection, is represented as the image sample, is the total number of data sets.

[0019] Preferably, the step 2 is specifically: the collected pressure regulating valve core position state sample image is preprocessed, including random cropping, scaling, enhancement, and twisting operation, so as to facilitate subsequent feature extraction and matching, and enhance the robustness of the data and the generalization ability of the model.

[0020] Preferably, the step 3 is specifically: using the ResNet50 network model , wherein is the model parameter, is the feature expression of the pressure regulating valve sample data; the pressure regulating valve core position state sample is encoded, and the feature reconstruction of the pressure regulating valve core position state sample image is completed based on the self-encoding model.

[0021] Preferably, the step 3 includes the following sub-steps:

[0022] 3.1, Feature model construction: input image, the image is passed through the convolutional neural network (ResNet50), the input pressure regulating valve core position state sample image is labeled, and all defect samples contain corresponding label information;

[0023] 3.2 Feature extraction (decoding): based on the ResNet50 network, the above preprocessed pressure regulating valve core position state data is input into the network model in batches, and the feature extraction of the input image is completed based on the convolution and pooling operation of the image; the specific operation is shown in formulas (1) and (2), wherein is a parameter set, a weight set, a bias term, an extracted feature, an original input sample, an intermediate layer output of the neural network, a transpose operation;

[0024] (1)

[0025] (2)

[0026] 3.3 Sample reconstruction (encoding): after completing the feature extraction of the position state sample of the proportional valve spool, this step uses a deconvolution-based operation to complete the reconstruction of the sample , and the specific operation is as shown in formula (3);

[0027] +1 (3)

[0028] wherein is the size of the input sample, which can be the same size as , is the size of the convolution kernel, is the number of layers filled with 0 in the deconvolution process, , and the step size of the convolution is represented by

[0029] Preferably, the step 4 comprises the following sub-steps:

[0030] 4.1, reconstruction sample feature acquisition: after the deconvolution operation, the reconstructed sample is again subjected to a ResNet50-based backbone network to obtain the corresponding reconstruction feature feature maps, and the specific operation is shown in the above formula (1) and (2) operations, and then a fully connected layer is used to obtain the final reconstruction feature;

[0031] 4.2 original feature extraction: considering the error information generated by the reconstruction feature, the original image data needs to be supplemented and improved, and the original preprocessed state image data is also subjected to forward propagation to complete feature extraction; combined with the reconstruction feature, the final double combined feature is formed to complete the nonlinearly processed proportional valve spool position state sample hierarchical feature; the original feature and the reconstructed feature are double combined for learning, and in order to make up for the information loss in the model reconstruction process, the specific operation is shown in formula (4);

[0032]

[0033] wherein The joint learning of the reconstructed feature and the original extracted feature is a double feature, which is a final matching valve core state sample feature, The original state sample feature is a feature, The reconstructed feature is a feature, The joint learning parameter is used to adjust the joint learning proportion of the two features to prevent redundant information after the fusion of the two types of features.

[0034] Preferably, the step 5 comprises the following sub-steps:

[0035] 5.1, defect sample classification training: after obtaining the matching valve core position state sample image and its corresponding label information, the ResNet50 model is used to obtain the double feature, and then the classification loss function is trained, which is specifically represented by the following formulas (5) and (6), to realize the classification training module of the defect sample;

[0036] (5)

[0037] (6)

[0038] Wherein represents the predicted label, represents the true label, that is, the real class label, is a commonly used multi-part function, and log is a logarithmic function, is a binary classification loss function;

[0039] 5.2, overall training of the matching valve frame: since the number of state samples of the matching valve working scene has the characteristic of few state features, if only binary classification is used, it is not enough to completely detect and classify all states; therefore, the sample feature reconstruction is used to complete the diversity of the input sample, and its reconstruction loss function is represented by the following formula (7);

[0040] (7)

[0041] 5.3, by combining the above classification loss and reconstruction loss, the overall loss function of the matching valve core position state detection combined with feature reconstruction and double discrimination is shown in formula (8);

[0042] (8)

[0043] Wherein is the overall loss function, is a hyperparameter for controlling classification and reconstruction.

[0044] Preferably, the step 6 comprises the following sub-steps:

[0045] 6.1 Through the optimal model training, the feature vectors of all valve core state sample pictures are extracted using the trained model, and the vectors are saved for subsequent state classification detection;

[0046] 6.2 When performing classification test, the state of the pressure regulating valve core is detected and classified through function, the state of the pressure regulating valve core is detected and classified through inputting the corresponding feature map and performing classification test through the classifier, the probability of the class to which each input sample feature belongs is calculated, and the state recognition is completed.

[0047] Preferably, the step 7 comprises the following sub-steps:

[0048] 7.1 System development: the development of the pressure regulating valve core position state online detection software system adopts a micro-service architecture, and combines software development processes, and according to requirement analysis, functional system design, implementation development, test application and other links, the online defect detection system is specifically developed;

[0049] 7.2 Model application: the pressure regulating valve core position state detection model trained by the method is transplanted to the corresponding developed good software platform, interface calling is realized, and the pressure regulating valve core state online classification detection in the actual production scene is completed.

[0050] Advantages of the present application:

[0051] (1) The present application achieves the optimal effect by constructing the encoding and reconstruction model of the pressure regulating valve sample features, setting several loss functions such as encoding, decoding and discriminator to mutually restrict each other; according to the encoding of the input sample features, the inverse convolution operation is adopted to realize the transformation of the sample data, the nonlinearity of the state sample data is increased, the robustness of the model is increased, and the influence of noise labels on the model training is effectively reduced.

[0052] (2) In addition, the present application ensures the robustness of the model through double discrimination of the original sample and the encoded features, increases the diversity of the training samples from the feature level, and the final recognition effect is better than that of the traditional method of increasing the diversity of the target features at the image level; in general, the detection precision and efficiency are enhanced, and the method has good migratability; the software system developed based on the algorithm can realize effective state detection effect. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 It is an accident pressure regulating valve state monitoring system architecture flow chart. DETAILED DESCRIPTION

[0054] The present application will be further described in detail below in combination with the drawings and specific embodiments.

[0055] As Figure 1As shown, the state detection problem can essentially be viewed as a binary classification problem, grouping open states into one category and closed states into another. Different shooting angles, lighting conditions, and scale can all affect the accuracy of state detection. To address these issues, this invention designs a method for detecting the position state of the valve core in a turbine generator set during an accident. This method mainly involves three parts: feature acquisition of the valve core position state, construction of a feature generation and discrimination framework model, and design of a state classification detection model. The first part is primarily used for feature extraction of the valve core position state samples. The second part, through the design of a corresponding deep learning model, transforms, maps, and enhances the features of the valve core position state samples, utilizing a production adversarial model to achieve diversified fusion of the valve core position state features. The third part, through the design of a classification loss function, achieves classification detection of the valve core position state. The specific steps include:

[0056] Step 1: Sampling of Pressure Control Valve Spool Position Status Data; Using various acquisition devices such as sensors, cameras, and mobile phones, the position status data of the pressure control valve spool is captured and obtained, acquiring different types of pressure control valve spool position status data. The pressure control valve spool position status sample dataset is divided into multiple batches, with each batch containing several state sample images. During subsequent model training, several sample data (images) of different states are randomly selected from multiple batches, and then used for training. The state dataset collected in this operation can be represented as... ,in D Represented as a set of state datasets, Represented as the first Zhang image samples, This represents the total number of data points in the dataset. The specific steps are as follows:

[0057] 1.1 Acquire sample data (images or video sequences) of the valve core position status of the pressure regulating valve to be tested. Acquisition is completed using data acquisition devices such as cameras, sensors, etc. The quality of image acquisition directly affects the accuracy of subsequent analysis and testing; therefore, it is necessary to ensure that the images are clear, sufficient, and capable of capturing the details of the surface of the device to be analyzed.

[0058] 1.2 The collected data is filtered and cleaned to remove low-quality images or video sequences and retain high-quality data samples. Simultaneously, duplicate data with excessively high similarity is removed to avoid class bias caused by too many similar defective samples, while increasing the diversity of data categories and improving the robustness of the model.

[0059] 1.3 Constructing a data set, from which a sequence of data samples is randomly selected as a batch, and for a batch of image samples, as many samples as possible are selected with different perspectives, different scales, etc. The types of samples are as diverse as possible, and the categories they present are as diverse as possible.

[0060] Step 2: Data preprocessing; the collected pressure regulating valve spool position state sample images are preprocessed, including random cropping, scaling, enhancement, and twisting operations, to facilitate subsequent feature extraction and matching, and to enhance the robustness of the data and the generalization ability of the model. The specific steps are as follows:

[0061] 2.1 Data inspection. Check whether the number of state sample data (record number) meets the minimum requirement of deep model feature extraction, whether the pressure regulating valve spool position state is comprehensive and meets the standard and requirement of this method. Specifically, the obtained state image is normalized and the mean value is subtracted, so that it has the same size and proportion. Adjust the image to a fixed size.

[0062] 2.2 Data cleaning. This step uses appropriate methods to clean the "dirty" data to make it "cleaner", which is beneficial to subsequent statistical analysis to obtain reliable feature expression.

[0063] 2.3 Data conversion. Data analysis emphasizes the comparability of the analysis object. Comprehensive analysis of some statistical indicators requires data transformation before analysis, such as dimensionless processing, linear transformation, summarization and aggregation, moderate generalization, normalization, and attribute construction.

[0064] 2.4 Data verification. The purpose of this step is to preliminarily evaluate and judge whether the data meets the needs of statistical analysis, and to decide whether to increase or decrease the amount of data. In order to increase the diversity and robustness of the data, the data is enhanced by rotation, cropping, scaling, etc. to expand the size of the data set and improve the generalization ability of the model. And use simple linear models, scatter plots, histograms, and other image transformations for exploratory analysis, use correlation analysis, consistency checks, and other methods to verify the accuracy of the data, to ensure that no errors and biases are brought into the actual data analysis.

[0065] Step 3: Feature encoding; use ResNet50 network model , where is the model parameter, The feature expression of the pressure regulating valve sample data is expressed, the feature coding of the pressure regulating valve spool position state sample is performed, and the feature reconstruction of the pressure regulating valve spool position state sample image is completed based on a self-coding model; different from the existing method, the patent based on the ResNet50 model realizes the coding of the input sample features in a self-coding mode, and realizes the decoding of the compressed features in a transposed convolution mode, so as to complete the reconstruction of the pressure regulating state input data. The step 3 comprises the following sub-steps:

[0066] 3.1, feature model construction: the input image is input into a convolutional neural network (ResNet50), the input pressure regulating valve spool position state sample image is labeled, and all defect samples contain corresponding label information;

[0067] 3.2 feature extraction (decoding): based on the ResNet50 network, the preprocessed pressure regulating valve spool position state data is input into the network model in batches, and the feature extraction of the input image is completed based on the convolution and pooling operation of the image; the specific operation is shown in formulas (1) and (2), wherein is a parameter set, is a weight set, is a bias term, is an extracted feature, is an original input sample, is an intermediate layer output of the neural network, is a transposition operation;

[0068] (1)

[0069] (2)

[0070] 3.3 sample reconstruction (coding): after the feature extraction of the pressure regulating valve spool position state sample is completed, the reconstruction sample is completed by using the operation based on the transposed convolution in this step , and the specific operation is shown in formula (3);

[0071] +1 (3)

[0072] wherein is the size of the input sample, which can be the same size as , is the size of the convolution kernel, is the number of layers filled with 0 in the transposed convolution process, , and the step size of the convolution is represented.

[0073] Step 4: Reconstructed sample feature acquisition; through the reconstruction of the above pressure regulating valve state sample data, and in combination with the original sample data, the sample state is distinguished. As above, ResNet50 network is used as the backbone network in this step, and feature extraction is performed on the reconstructed sample and the original sample to complete the feature acquisition of the real sample and the reconstructed sample. Step 4 includes the following sub-steps:

[0074] 4.1, Reconstructed sample feature acquisition: after the deconvolution operation, the reconstructed sample Again, the corresponding reconstructed feature feature maps are obtained by taking ResNet50 as the skeleton network, and the specific operation is shown in the above formulas (1) and (2). Then, through the full connection layer, the final reconstructed feature is obtained;

[0075] 4.2 Original feature extraction: considering the error information generated by the reconstructed feature, the original image data needs to be supplemented and improved. The pre-processed state image data is also extracted through the forward propagation method. Combined with the reconstructed feature, the final double combined feature is formed, and the non-linear processed pressure regulating valve spool position state sample hierarchical feature is completed. The original feature and the reconstructed feature are double combined learning, in order to make up for the information loss in the model reconstruction process, and the specific operation is shown in formula (4);

[0076]

[0077] Wherein is the double feature after the joint learning of the reconstructed feature and the original extracted feature, that is, the final pressure regulating valve spool state sample feature, is the original state sample feature, is the reconstructed feature, is the joint learning parameter of the two, used to adjust the joint learning proportion of the two features, so as to prevent the generation of redundant information after the fusion of the two features.

[0078] Step 5: model training; since the method of the present application is a supervised state detection method, each spool position state sample of each pressure regulating valve has corresponding label information. For each randomly extracted batch sample, according to its corresponding information label, we set the corresponding binary classification loss function . In addition, since the original sample data has completed the image reconstruction of the deconvolution operation, we set the corresponding reconstruction loss function accordingly, and combine the two to complete the entire pressure regulating valve spool position state detection framework and perform corresponding model training. Step 5 includes the following sub-steps:

[0079] 5.1, Defect sample classification training: for the obtained pressure regulating valve spool position state sample image and its corresponding label information, after obtaining the double features through the ResNet50 model, the classification loss function is trained, which is specifically represented as formulas (5) and (6), to realize the classification training module of the defect sample.

[0080] (5)

[0081] (6)

[0082] wherein represents the predicted label, represents the true label, that is, the true class label, is a commonly used multi-function, and log is a logarithmic function, is a binary classification loss function.

[0083] 5.2, Overall training of the pressure regulating valve frame: since the number of state samples of the pressure regulating valve working scene has the characteristic of few states, if only binary classification is used, it is not enough to completely detect and classify all states; therefore, the diversity of the input sample is completed by using sample feature reconstruction, and the reconstruction loss function is represented as formula (7).

[0084] (7)

[0085] 5.3, by combining the above classification loss and reconstruction loss, then the overall loss function of the pressure regulating valve spool position state detection combined with feature reconstruction and double discrimination is shown in formula (8).

[0086] (8)

[0087] wherein is the overall loss function, is a hyperparameter for controlling classification and reconstruction.

[0088] Step 6: classification detection; after completing the network model optimization training, the trained model is used to test the pressure regulating valve spool position state. For each state sample data, its feature vector can be stored in a database, and when state detection and recognition are needed, the feature vector of the query image is compared with all sample features in the database to find similar image models as the classification result, realizing the detection and recognition of the state. The step 6 includes the following steps:

[0089] 6.1, using the optimal model obtained by training, the feature vectors of all valve core state sample pictures are extracted, and the vectors are saved for subsequent state classification detection;

[0090] 6.2 When conducting classification tests, through The function detects and classifies the state of the pressure regulating valve core. It inputs a corresponding feature map and performs classification tests through a classifier. For each input sample feature, it calculates the probability of its class and completes the identification of its state.

[0091] Step 7: System Development. After completing the development of the corresponding pressure regulating valve spool position status detection algorithm, the corresponding model is encapsulated and integrated into the developed software platform to realize real-time spool position status detection of the pressure regulating valve. Step 7 includes the following sub-steps:

[0092] 7.1 System Development: The development of the online inspection software system for the valve core position status of the pressure regulating valve adopts a microservice architecture and combines the software development process. Based on the requirements analysis, functional system design, implementation development, testing and application, a defect online detection system is specifically developed.

[0093] 7.2 Model Application: The valve core position status detection model trained by this method is ported to a well-developed software platform to implement interface calls and complete the online classification and detection of valve core status in actual production scenarios.

[0094] This invention is used for the detection and identification of the valve core position status of a generator unit in emergency situations. This invention primarily addresses the difficulty in identifying the status of emergency pressure regulating valves due to their limited range of status types, subtle changes, varying working viewpoints, and differences in observation scales. Based on a deep learning framework, it employs next-generation artificial intelligence technology to perform valve core position status detection and identification from an image recognition perspective, thereby improving the working efficiency of hydropower station units and protecting their safe operation. The key technical aspects of this invention are mainly reflected in the following aspects:

[0095] (1) Acquisition and diversified transformation preprocessing of valve core position status data of pressure regulating valve in order to obtain diversified state feature sample data and realize model training optimization;

[0096] (2) The reconstruction encoding of the valve core position status sample features of the pressure distribution valve is mainly based on ResNet50 to complete feature extraction, and then the input features are decoded and reconstructed through deconvolution operation.

[0097] (3) Combine the original input samples and the encoded reconstructed samples to achieve joint learning of the two modal features, and use a binary classifier to classify and identify the valve core position status samples of the pressure regulating valve.

[0098] In addition, since the position state of the pressure regulating valve spool is relatively single (only two states of opening and closing), it can be considered as a binary classification problem, and the design requirements of the model structure are simple and convenient compared with other multi-classification problems. Therefore, ResNet50, as a relatively lightweight model, is more suitable in this scenario, which can ensure the detection accuracy without the complexity of the transformer model structure, and has better recognition accuracy than LeNet-5 (the most initialized convolutional neural network). Under the premise of ensuring the accuracy and model complexity, the use of ResNet50 network structure has better cost performance. In addition, by jointly training and discriminating the reconstructed samples and the original samples, the nonlinearity of the sample data is increased, which is helpful to extract the discriminative features of the sample features. Therefore, considering the above factors, the technical solution involved in the present application is the technical solution with the highest cost performance at present.

[0099] Overall, based on the technology adopted in the present application, the algorithm verification and system development and application are carried out, which effectively improves the management efficiency of the existing hydroelectric generating set governor accident pressure regulating valve in the actual hydroelectric generating set plant application scene. Especially more than 50% of human resources are saved (4 people are needed to form a shift to pay attention to the state of the valve core of the pressure regulating valve, and the intelligent video monitoring technology based on deep learning only needs two people to form a shift to complete the corresponding task).

[0100] The above embodiments are only preferred technical solutions of the present application, and should not be regarded as a limitation of the present application. The protection scope of the present application should be based on the technical solutions recited in the claims, including the equivalent replacement solutions of the technical features recited in the claims. That is, the equivalent replacement improvement within this range is also within the protection scope of the present application.

Claims

1. A method for detecting the position state of the valve core of the accident pressure distribution valve of the speed regulator of a hydraulic turbine unit, characterized in that: It comprises the following steps: Step 1: data sampling of the position state of the valve core of the pressure regulating valve; Step 2: data preprocessing; Step 3: feature encoding; Step 4: obtaining reconstructed sample features; Step 5: model training; Step 6: classification detection; Step 7: system development; The step 3 is specifically using a ResNet50 network model wherein is a model parameter, is a feature expression of the pressure regulating valve sample data; the valve core position state sample of the pressure regulating valve is feature coded, and feature reconstruction of the valve core position state sample image of the pressure regulating valve is completed based on a self-encoding model; 3.1, feature model construction: the input image is processed through the convolutional neural network ResNet50, and the input pressure regulating valve core position state sample image is labeled, and all defect samples contain corresponding label information; 3.2, feature extraction: based on the ResNet50 network, the preprocessed pressure regulating valve core position state data is input into the network model in batches, and the feature extraction of the input image is completed based on the convolution and pooling operation of the image; The specific operation is shown in formulas (1) and (2), wherein is a parameter set, is a weight set, is a bias term, is an extracted feature, is an original input sample, is an intermediate layer output of a neural network, is a transpose operation; （1）； （2）； 3.3, Sample reconstruction: After the feature extraction of the sample of the position state of the valve core of the pressure regulating valve is completed, this step uses the operation based on deconvolution to complete the reconstruction of the sample The specific operation is as follows formula (3); +1（3）； wherein is the input sample size, which can be the same as , is the kernel size, is the number of layers filled with 0 in the deconvolution process, denotes the stride size of the convolution; The step 4 comprises the following sub-steps: 4.1, reconstruction sample feature acquisition: after the deconvolution operation, the acquired reconstruction sample Again, take ResNet50 as the backbone network to obtain the corresponding reconstruction feature feature maps, see the above formulas (1) and (2) for specific operations, and then through the full connection layer to obtain the final reconstruction feature; 4.2, original feature extraction: considering the error information generated by the reconstructed feature, the original image data needs to be supplemented and improved, and the original preprocessed state image data is also extracted through forward propagation; combined with the reconstructed feature, the final double combined feature is formed, and the pressure regulating valve core position state sample hierarchical feature after nonlinear processing is completed; the original feature and the reconstructed feature are double combined learning, in order to make up for the information loss in the model reconstruction process, and the specific operation is shown in formula (4); （4）； wherein is the double feature after the joint learning of the reconstructed feature and the original extracted feature, i.e., the final matching valve spool state sample feature, is the original state sample feature, is the reconstructed feature, is the joint learning parameter of the two, used to adjust the joint learning proportion of the two features, to prevent redundant information after the fusion of the two types of features; The step 7 comprises the following sub-steps: 7.1, system development: the development of the online detection software system of the position state of the valve core of the pressure regulating valve, adopting micro service architecture, combining software development process, according to the links of demand analysis, function system design, implementation development, test application, the defect online detection system is developed in detail; 7.2, model application: the trained pressure regulating valve core position state detection model is transplanted to the corresponding developed software platform, the interface calling is realized, and the online classification detection of the pressure regulating valve core state in the actual production scene is completed.

2. The method for detecting the position state of the valve core of the emergency pressure distribution valve of the speed regulator of a hydraulic turbine unit according to claim 1, characterized in that: The step 1 is specifically: through the acquisition device, the position state data of the proportional valve core is shot and collected, different types of proportional valve core position state data are acquired; the proportional valve core position state sample data set is divided into multiple batches batch, wherein each batch contains several state sample images; in subsequent model training, sample data in several different states are selected from multiple batches in turn, and then the data is trained; the state data set collected by the operation is represented as Wherein D is represented as a state data set collection, is represented as the first image sample, is the total number of data sets.

3. The method for detecting the position state of the valve core of the emergency pressure distribution valve of the speed regulator of a hydraulic turbine unit according to claim 1, characterized in that: The step 2 is specifically: the collected pressure regulating valve core position state sample image is preprocessed, including random cropping, scaling, enhancement and twisting operation, so as to facilitate the subsequent feature extraction and matching, and enhance the robustness of data and the generalization ability of model.

4. The method for detecting the position state of the valve core of the emergency pressure distribution valve of the speed regulator of a hydraulic turbine unit according to claim 1, characterized in that: The step 5 comprises the following sub-steps: 5.1, classification training of defect samples: after obtaining the pressure regulating valve core position state sample image and its corresponding label information, the double feature is obtained through the ResNet50 model, and the classification loss function is trained, which is specifically expressed as formula (5) and (6), realizing the classification training module of defect samples; （5）； （6）； wherein denotes the predicted label, denotes the true label, i.e. the true class label, is a commonly used multivariate function, log is the logarithm function, is a binary classification loss function; 5.2, The overall training of the pressure regulating valve frame: Because the state sample number of the working scene of the pressure regulating valve has the characteristics of few state characteristics, if only two classification recognition is used, it is not enough to completely detect and classify all states; Therefore, the sample feature reconstruction method is used to complete the diversity of the input sample, and its reconstruction loss function is represented as formula (7) as follows: （7）； 5.3, by combining the above classification loss and reconstruction loss, the overall loss function of the pressure regulating valve core position state detection of the feature reconstruction and double discrimination cooperation is shown as formula (8); （8）； wherein is the overall loss function, are hyperparameters that control classification and reconstruction.

5. The method for detecting the position state of the valve core of the emergency pressure distribution valve of the speed regulator of a hydraulic turbine unit according to claim 1, characterized in that: The step 6 comprises the following sub-steps: 6.1, the optimal model is obtained by training, and the trained model is used to extract the feature vector of all valve core state sample pictures, and the vector is saved for subsequent state classification detection; 6.2、When the classification test is performed, the state of the valve core of the pressure regulating valve is detected and classified by The function detects and classifies the state of the valve core of the pressure regulating valve by inputting the corresponding feature map and performing classification test through the classifier. For each input sample feature, the probability of the class to which it belongs is calculated to complete the recognition of its state.

Images