Method and device for identifying mistaken-walking-prevention interval human factor event of nuclear power station

Abstract

The invention discloses a nuclear power station mistaken-walking-prevention interval human factor event identification method and device, and relates to the technical field of nuclear power station human factor accident prevention. The method comprises the following steps: S1, before an operator receiving a task instruction executes a task, carrying out voice acquisition and recognition on voice record of the operator, converting record voice information of the operator into text information, and carrying out comparative analysis on the text information and reference text information corresponding to the task instruction, according to the analysis result, prompt information of successful identification or early warning information of failed identification is given; and S2, after ticket reading voice recognition succeeds, image acquisition and recognition are performed on an operation button pointed by an operator, an acquired image is converted into character information, the character information and reference character information corresponding to the task instruction are subjected to comparative analysis, and prompt information about recognition success or early warning information about recognition failure is given according to an analysis result. The method can effectively prevent the nuclear power station from misplacement interval human factors.

Description

Technical Field

[0001] The invention relates to the technical field of prevention of human-caused accidents in nuclear power plants, and in particular to a method and device for identifying human-caused events in a nuclear power plant to prevent wrong interval walking. Background Art

[0002] With the development of nuclear power, the reliability of equipment is getting higher and higher, but the operation and maintenance process is basically completed by humans. There are many uncertainties in humans, and human errors may occur during the execution of tasks. According to statistics, 60% to 70% of the events (accidents) in nuclear power plants are directly or indirectly related to human factors, which shows that human errors are an important cause of events.

[0003] As one of the types of human error, the wrong interval has a high proportion in human error events. It will more or less bring some adverse consequences, and in some cases the consequences are quite serious. Therefore, it is very important to prevent the wrong interval. At present, nuclear power plants basically adopt the traditional way of preventing the wrong interval, that is, according to the requirements of the human error prevention tool card, the operation is carried out according to the process; on the other hand, the training of personnel is strengthened. However, due to the dynamic and complexity of people, and the diversity and complexity of the tasks themselves, human error accidents caused by wrong intervals occur from time to time. Summary of the invention

[0004] One of the purposes of the present invention is to provide a method for identifying human-caused events that can effectively prevent wrong interval walking.

[0005] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme: a method for identifying human-caused events in a nuclear power plant to prevent wrong intervals, comprising:

[0006] S1. Before the operator who receives the task instruction performs the task, the operator's voice vote counting is collected and recognized, the operator's voice vote counting information is converted into text information, and the text information is compared and analyzed with the reference text information corresponding to the task instruction, and prompt information of successful recognition or warning information of failed recognition is given according to the analysis result;

[0007] S2. After successful voice recognition of the vote counting, the operation button currently pointed to by the operator (such as with a finger) is imaged and recognized, the image is converted into character information, and the character information is compared and analyzed with the reference character information corresponding to the task instruction. A prompt message of successful recognition or a warning message of failed recognition is given according to the analysis result.

[0008] Furthermore, the specific steps of collecting and recognizing the voice of the operator in S1 include:

[0009] The operator performs voice counting, collects the operator's voice counting information (such as voice signals) and performs feature extraction to obtain a spectrogram, which is then input into a neural network model for processing to obtain an output result in the form of One-hot encoding, performs text conversion on the One-hot encoding to obtain the text information of the current voice counting, and then compares the text information with the reference text information corresponding to the task instruction, and determines whether the voice counting recognition is successful based on the comparison result.

[0010] Furthermore, the neural network model includes an input layer, a hidden layer and an output layer, the number of nodes in the input layer is i_nodes, the number of nodes in the hidden layer is H_nodes, and the number of nodes in the output layer is 0_nodes;

[0011] The output H_output of the hidden layer is:

[0012]

[0013] The output Output_v of the output layer is:

[0014]

[0015] Among them, W ih is the connection weight between the input layer and the hidden layer neurons, W ho H_v is the connection weight between the hidden layer and the output layer neurons, h is the hidden layer threshold, O_v0 is the output layer threshold, x h is the input vector. h Output_v is the output value of the hidden layer. o is the output value of the output layer.

[0016] Furthermore, through data training, the weight W ih , W ho Adjust the weight W ih , W ho Substitute the output value calculation formula of the hidden layer and the output layer to correct the output value of the output layer to obtain the corrected output value. The specific steps include:

[0017] a. Calculate the output error E of the output layer as follows:

[0018]

[0019] Among them, D_v and D_v0 are the expected output values ​​of the output layer.

[0020] Furthermore, the output error E of the output layer is calculated as follows:

[0021]

[0022] Furthermore, the output error E of the output layer is calculated as follows:

[0023]

[0024] Among them, x i_nodes is the input vector.

[0025] Furthermore, the above steps also include:

[0026] b. If the output error E of the output layer is greater than the minimum output error, the weight W is calculated as follows: ih , W ho Adjust to reduce output error:

[0027] The weight adjustment formula of the output layer is:

[0028]

[0029] in, represents the output layer weight adjustment value; β is a constant parameter; d_v0 represents the oth expected value; represents the pth output value of the hth hidden layer;

[0030] The weight adjustment formula of the hidden layer is:

[0031]

[0032] in, Represents the hidden layer weight adjustment value; β is a constant, and the initial value is given before training; is the value of the oth neuron node in the kth hidden layer; is the connection weight between the oth neuron in the hth hidden layer and a neuron in the h+1th hidden layer; Represents the output value of the pth neuron in the hth hidden layer; is the output value of the i-th input node and the h-1-th neuron in the first hidden layer;

[0033]

[0034] in, is the output of the hth neuron node in the pth hidden layer; is the value of the oth neuron node in the kth hidden layer;

[0035] c. Adjust the weight W after step b ih , W hoSubstitute into the calculation formula of the output error to calculate the output error E. When the output error E reaches the minimum output error or the number of training times reaches the maximum number of training times, terminate the training and output the corrected output value as the final output value. Otherwise, repeat steps a and b until the output error E reaches the minimum output error or the number of training times reaches the maximum number of training times.

[0036] Furthermore, the specific steps of performing image acquisition and recognition on the operation button currently pointed to by the operator in S2 include:

[0037] When the operator points to an operation button, an image of the mark on the operation button pointed to by the operator is captured, and the captured image is smoothed and character segmented, and then character recognition is performed to obtain the character sequence of the operation button. The character sequence is then compared with the reference character information corresponding to the task instruction, and based on the comparison result, it is determined whether the operation button character recognition is successful (that is, whether the button currently pointed to by the operator is consistent with the task instruction button that needs to be pointed to).

[0038] Furthermore, the image smoothing process adopts a rotating mask averaging method, the character segmentation process adopts a vertical projection segmentation method, and the character recognition is performed using a convolutional neural network.

[0039] Another object of the present invention is to provide a device for identifying human factors events in a nuclear power plant to prevent the wrong interval from being walked into, the device comprising a server, a workstation, and a collection device, the collection device being used to collect the voice of the operator's voice counting and to collect images of the operation buttons pointed to by the operator, and to transmit the collected data to the server, the server being used to receive the data from the collection device and to execute the steps in the method for identifying human factors events in a nuclear power plant to prevent the wrong interval from being walked into, outputting the recognition results and corresponding prompts or warnings and transmitting them to the workstation, the workstation being used to issue prompts or warning information.

[0040] The present invention can issue early warnings before and when operators perform tasks through the dual operations of ticket-calling voice recognition and operation button character recognition, which can greatly reduce the probability of going into the wrong interval and effectively prevent human factors such as going into the wrong interval. Moreover, the above operations are implemented through devices and equipment, and the equipment itself is deterministic and timely. Compared with the methods of preventing human factors such as tool cards or strengthening personnel training in the prior art, preventing accidents of going into the wrong interval through equipment can play a fundamental role. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 The architecture and functional diagram of the human-caused event recognition device for preventing wrong interval walking in a nuclear power plant;

[0042] Figure 2 Provide a voice recognition flow chart for operators to count votes;

[0043] Figure 3 Identify the image recognition flow chart for the operation position;

[0044] Figure 4 This is a diagram of the character recognition process based on CNN. DETAILED DESCRIPTION

[0045] In order to facilitate those skilled in the art to better understand the improvements of the present invention relative to the prior art, the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0046] In a narrow sense, the wrong interval refers to the operation and maintenance personnel walking into the wrong equipment room, such as walking into the wrong room of other units. In a broad sense, the wrong interval refers to the wrong working object. The types of wrong interval human factors events are divided into three categories: walking into different physical intervals, operating the wrong object in the same physical interval, and operating the wrong object in the control system software. Walking into different physical intervals means: the two erroneous equipment objects are in different spaces, such as different units, different rooms or different cabinets; operating the wrong object in the same physical interval means: the operator operates adjacent or similar wrong objects in the same physical interval; operating the wrong object in the control system software means: in the distributed control system (DCS) of a nuclear power plant, temporary control changes are required due to regular testing or maintenance work, and wrong interval events are prone to occur when the control logic or control signal is manually changed. Based on the above-mentioned wrong interval types, the present invention designs a prevention device (such as a nuclear power plant anti-wrong interval human factors event identification device) to avoid accidents.

[0047] The present invention selects the high-risk industry of nuclear power plants, the background is on-site operation tasks, and the means used is to prevent wrong intervals through equipment and devices, which is of great help to the on-site operation efficiency of nuclear power plants and the prevention of human-caused accidents, and has certain practical significance and application prospects. Among them, the present invention takes the on-site operation tasks of nuclear power plants as the background, and designs a human-caused event identification device to prevent wrong intervals. Because the equipment itself is deterministic and timely, compared with the anti-human-caused tool card or by strengthening personnel training, preventing wrong interval accidents through equipment will inevitably play a fundamental role.

[0048] Based on the above concept, this embodiment provides a device (system) for identifying human error events in a field operation task of a nuclear power plant. The device mainly includes: a central server, a fixed workstation, a mobile workstation, a remote data acquisition device and a software system. The architecture and functional relationship of the device are as follows: Figure 1 shown.

[0049] (1) Central server: As the core of the entire system, it is mainly responsible for data reception, analysis and processing, information push, storage, etc. Specifically, it can recognize the images of the operation position or logo obtained by the monitoring equipment, recognize the operator's voice of the vote, retrieve the command information, etc. It mainly includes: task operation command database, task corresponding button character database and voice database; main function: through AI artificial intelligence technology, it can realize the functions of picture feature recognition, voice comparison and operation deviation warning.

[0050] (2) Fixed workstation: Generally speaking, a fixed workstation can be established if conditions permit at a certain task site. Its purpose is to confirm the operator's task list, capture images of the location of the operation task, collect the voice of the vote counting, and provide warnings for erroneous operations.

[0051] (3) Mobile workstation: If conditions do not allow or cannot establish a fixed workstation at a certain task site, a mobile workstation can be used. Its purpose is the same as that of a fixed workstation. The difference is that the data acquisition equipment in a fixed workstation can be fixed, while the mobile workstation is a remote interaction system that integrates human-computer interaction and data acquisition equipment.

[0052] (4) Remote data acquisition equipment: The acquisition equipment is mainly used to collect voice and image data of the vote-calling voice and operation position points during task execution, and transmit the collected data to the server for processing. The acquisition equipment is usually set in a fixed workstation or a mobile workstation, and usually includes a microphone and a monitoring device.

[0053] The functions of the human-caused event recognition device for preventing wrong intervals during on-site operation tasks in nuclear power plants and the identification and early warning methods are described below.

[0054] 1. Voice recognition function for operators to count votes

[0055] Generally speaking, after receiving a task instruction, the operator should call out the votes according to the instruction requirements before executing the task. The purpose of voice recognition of the call out is to determine whether the operator's call out is consistent with the task instruction. If not, the server will generate an alarm after judgment, and transmit the alarm information to the on-site mobile workstation or fixed workstation to remind the operator. The process of building a speech recognition system mainly includes: speech signal acquisition, speech signal preprocessing, feature parameter extraction of speech signals, quantization, and adaptive feedback module error correction. The language acquisition device uses a microphone. If it is a mobile workstation, the microphone can be installed at a certain location on the mobile terminal. Based on the above description, the operator's voice recognition process for calling out the votes is shown in Figure 1. Figure 2 .

[0056] from Figure 2It can be seen that the main process of voice recognition for operators counting tickets to prevent wrong intervals in nuclear power plant on-site tasks consists of three parts:

[0057] (1) Neural network model for voice command recognition

[0058] 1) Data preprocessing

[0059] The speech data is retrieved from the task instruction speech database, and features are extracted to form a spectrogram. The spectrogram is then feature enhanced as data sample input for the neural network training process. On the other hand, the neural network uses a supervised learning process, so the expected results for each input sample must be pre-set and input. In order to keep the expected results consistent with the sample input, when retrieving the task instruction speech, the speech data must be text encoded synchronously to obtain the One-hot encoding of the text data as the output of the neural network training process. Simply put, the spectrogram of the task instruction is used as the input of the neural network model, and the One-hot encoding is used as the model output. After training, a neural network classification model can be obtained to construct a mapping relationship between the spectrogram and the One-hot encoding.

[0060] 2) Neural Network Model

[0061] In the supervised neural network model method, the multilayer perceptron of the BP algorithm is a widely used neural network. The voice command recognition neural network of this embodiment adopts the BP neural network model.

[0062] The BP neural network model mainly consists of three layers: input layer, hidden layer and output layer. The number of neurons in the input layer and output layer is determined according to the number of input vectors and the number of target outputs respectively. The number of hidden layers and the number of neurons in each hidden layer are continuously adjusted and determined after iteration during the network training process. The initial number of neurons in the hidden layer is the square root of the sum of the number of input nodes and the number of output nodes, plus a constant between 1 and 10.

[0063] Based on the background of solving the problem in this embodiment, the voice command recognition BP neural network training mathematical model is defined as:

[0064] ① Initialization

[0065] According to the task of voting speech category, determine the number of input layer nodes (i_nodes), hidden layer nodes (H_nodes), and output layer nodes (0_nodes). The weights of the input layer and hidden layer neurons are represented by W. ih Indicates that the hidden layer and the output layer neuron connection weights are represented by W ho The initial threshold of the hidden layer is represented by H_v, the initial threshold of the output layer is represented by O_v, and the input vector is represented by (x1, x2, .....xi_nodes ) T Indicates that the output vector is (y1,y2,......y O_nodes )express.

[0066] ②Hidden layer output

[0067] According to the input variable (the spectrogram after the task speech voting preprocessing), the weight W of the connection between the input layer and the hidden layer neurons ih , hidden layer threshold H_v, calculate the hidden layer output H_output:

[0068]

[0069] f is the hidden layer excitation function. The excitation function has many forms. In this embodiment, the unipolar Sigmoid function is selected as the excitation function:

[0070]

[0071] ③ Output layer calculation results

[0072] Based on the hidden layer output H_output, the hidden layer and output layer neuron connection weight W ho , output layer threshold O_v, calculate the output layer result (Output_v):

[0073]

[0074] ④ Error calculation

[0075] When the resulting output is inconsistent with the expected value (D_v), the error is defined as:

[0076]

[0077] If the error definition is extended to the hidden layer, we have:

[0078]

[0079] If the error definition is extended to the input layer, we have:

[0080]

[0081] ⑤Weight adjustment

[0082] The purpose of weight adjustment is to reduce the output error and make the output closer to the expected value. Weight adjustment adopts the chain derivation rule of BP neural network. Based on the derivation of the chain derivation rule of BP neural network, the weight adjustment formulas of the output layer and the hth hidden layer are:

[0083] The output layer weight adjustment formula is:

[0084]

[0085] The weight adjustment formula of the hth hidden layer is:

[0086]

[0087] in:

[0088]

[0089] (2) Voice recognition for counting votes

[0090] The implementation process is similar to data preprocessing. Feature extraction is performed on the acquired task voice counting to obtain its spectrum, which is used as the input of the voice command recognition neural network model. One-hot encoding is obtained through neural network recognition, and then the One-hot encoding is converted into text to obtain the text corresponding to the task command counting voice.

[0091] (3) Early warning analysis

[0092] After the task instruction vote-calling voice is recognized by the neural network model to obtain the corresponding text, the text information of the corresponding task in the database (i.e., the reference text information) is retrieved and compared. If the text information is consistent with the database instruction text information, the correct prompt information is sent to the fixed workstation or mobile workstation. Otherwise, an early warning message is sent to remind the operator that the vote-calling is incorrect and to be careful not to go to the wrong interval. At this time, the operator needs to count the votes again.

[0093] 2. Operation position character image recognition function

[0094] When the voice counting is successfully completed, according to the requirements of the nuclear power plant operation process tool card, when the operator points his finger to the button to be operated, it will pause to prevent pressing the wrong button. However, due to the uncertainty of the personnel themselves, it is difficult to ensure that the wrong position is not pressed or the wrong interval is not reached. In addition, sometimes the operator does not necessarily follow the operation process completely, which makes it easier to reach the wrong interval. Therefore, in order to ensure that the task is executed without going to the wrong interval, when the operator points his finger to the button corresponding to the task, the icon character logo presented by the button is collected and then recognized. After recognition, it is compared and analyzed with the logo corresponding to the task specified library. If the two are consistent, it is considered that the pointed operation button or position is correct. Otherwise, an alarm is issued and the position is incorrect. Based on the operation position identification image recognition process, function and image recognition technology, its implementation process is as follows: Figure 3 .

[0095] 1. Preprocessing

[0096] (1) Image smoothing

[0097] Image smoothing can effectively suppress noise, but it may also cause edges to become blurred. The rotating mask averaging method is a nonlinear method that can avoid edge blurring. Therefore, this embodiment uses the classic smoothing method of rotating mask averaging to maintain edge clarity while removing noise. The average brightness is calculated based on a certain area, and a brightness spread β is used as a consistency metric for the image area. The spread β is defined as:

[0098]

[0099] The general execution process of the rotating mask averaging method is:

[0100] 1) Evaluate each pixel (i,j) of the image.

[0101] 2) According to the definition of spread β, calculate the spread of all possible rotation masks for pixel (i, j).

[0102] 3) Select the mask with the smallest scatter.

[0103] 4) Calculate the average brightness of the selected mask and then assign the average brightness to the pixel (i, j) in the output image.

[0104] (2) Character line segmentation

[0105] The main purpose of the character threshold segmentation on the button pointed by the operator is to separate the target and the background in the image, so as to facilitate the analysis of the target. The main idea of ​​threshold segmentation of the grayscale image is as follows: First, for each image, the minimum gray value (G_L min ), maximum gray value (G_L max ), and then take a grayscale threshold Threshold between the minimum and maximum grayscale values; then compare the grayscale value of each pixel in the image with the threshold, those greater than the threshold are represented by white, and those less than the threshold are represented by black.

[0106] (3) Single character segmentation

[0107] Character segmentation uses vertical projection segmentation method. Projection operation is the first step in character segmentation. First, the image characters on the task command button after binarization are scanned in sequence to obtain a vertical projection image. In the histogram, the peak is the projection of the character, and the trough value is the interval point between two characters. In the second step, the center position of all the peaks is selected, and the corresponding characters are obtained according to the center position. In this way, the character segmentation is completed, and then it is saved as the input for character recognition.

[0108] 2. Character Recognition

[0109] Character recognition is to identify the letters or numbers of the button image corresponding to the current command. Convolutional neural networks are widely used in image processing. They are feedforward neural networks that include convolution transformation and have deep results. They are one of the most representative methods of deep learning. This embodiment uses convolutional neural networks to complete character recognition.

[0110] (1) CNN processing flow

[0111] According to the structural characteristics of CNN and the research background of this embodiment, the character recognition process on the button image corresponding to the current instruction is as follows: Figure 4 .

[0112] (2) Character Recognition Convolutional Network Structure (CNN)

[0113] There are multiple categories of convolutional neural networks. This embodiment selects the simpler LeNet network as the character recognition of the command button image. In CNN, the LeNet network mainly includes: input layer, convolution layer CV1, pooling layer SL2, convolution layer CV3, pooling layer SL4, convolution layer CL5, full connection layer, output layer. The specific description of each layer is shown in Table 1.

[0114] Table 1 Convolutional structure based on character recognition

[0115]

[0116]

[0117] (3) Character comparison

[0118] The purpose of character comparison is to ensure that the button currently pointed to by the operator is consistent with the task instruction button currently required to be pointed to. The specific comparison process is as follows: After the characters on the task instruction button are recognized by the convolutional neural network, the recognized characters are saved in a one-dimensional array in the order of character segmentation. Then, the corresponding character string (i.e., reference character information) is called out from the task instruction button character database, and the characters on the corresponding task button in the database are compared with the recognized characters in turn. If they are completely equal, it is considered that the task instruction button pointed to by the operator is correct and there is no wrong interval. Otherwise, an alarm is issued, and the operator's pointing does not match the current task instruction button and the interval is wrong.

[0119] In addition, the character comparison algorithm pseudo code is as follows:

[0120]

[0121]

[0122] The above embodiments are preferred implementation schemes of the present invention. In addition, the present invention may also be implemented in other ways. Any obvious replacement without departing from the concept of the present technical solution is within the protection scope of the present invention.

[0123] Finally, it should be emphasized that in order to allow ordinary technicians in the field to more easily understand the improvements of the present invention over the prior art, some descriptions of the present invention have been simplified, and for the sake of clarity, some other elements are omitted in this application document. Those skilled in the art should realize that these omitted elements may also constitute the content of the present invention.

Claims

1. A method for identifying human-caused events in a nuclear power plant to prevent wrong intervals, characterized in that: include: S1. Before the operator who receives the task instruction performs the task, the operator's voice vote counting is collected and recognized, the operator's voice vote counting information is converted into text information, and the text information is compared and analyzed with the reference text information corresponding to the task instruction, and prompt information of successful recognition or warning information of failed recognition is given according to the analysis result; S2. After successful voice recognition of the vote counting, the operation button pointed to by the operator is imaged and recognized, the image is converted into character information, and the character information is compared and analyzed with the reference character information corresponding to the task instruction. A prompt message of successful recognition or a warning message of failed recognition is given according to the analysis result.

2. The method for identifying human-caused events in preventing wrong intervals in a nuclear power plant according to claim 1 is characterized in that: The specific steps of voice collection and recognition of the operator's voice counting in S1 include: The operator performs voice counting, collects the operator's voice counting information and performs feature extraction to obtain a spectrogram, which is then input into the neural network model for processing to obtain an output result in the form of One-hot encoding, and performs text conversion on the One-hot encoding to obtain the text information of the current voice counting, which is then compared with the reference text information corresponding to the task instruction, and whether the voice counting recognition is successful is determined based on the comparison result.

3. The method for identifying human-caused events in preventing wrong intervals in a nuclear power plant according to claim 2 is characterized in that: The neural network model includes an input layer, a hidden layer and an output layer, the number of nodes in the input layer is i_nodes, the number of nodes in the hidden layer is H_nodes, and the number of nodes in the output layer is 0_nodes; The output H_output of the hidden layer is: The output Output_v of the output layer is: Among them, W ih is the connection weight between the input layer and the hidden layer neurons, W ho H_v is the connection weight between the hidden layer and the output layer neurons, h is the hidden layer threshold, O_v0 is the output layer threshold, x h is the input vector.

4. The method for identifying human-caused events in preventing wrong intervals in a nuclear power plant according to claim 3 is characterized in that: Through data training, the weight W ih , W ho Adjust the weight W ih , W ho Substitute the output value calculation formula of the hidden layer and the output layer to correct the output value of the output layer to obtain the corrected output value. The specific steps include: a. Calculate the output error E of the output layer as follows: Among them, D_v is the expected output value of the output layer.

5. The method for identifying human-caused events in preventing wrong intervals in a nuclear power plant according to claim 4 is characterized in that: The output error E of the output layer is calculated as follows:

6. The method for identifying human-caused events in preventing wrong intervals in a nuclear power plant according to claim 5 is characterized in that: The output error E of the output layer is calculated as follows: Among them, x i_nodes is the input vector.

7. The method for identifying human-caused events in preventing wrong intervals in a nuclear power plant according to claim 6 is characterized in that: Also includes: b. If the output error E of the output layer is greater than the minimum output error, the weight W is calculated as follows: ih , W ho Make adjustments: The weight adjustment formula of the output layer is: in, represents the output layer weight adjustment value; β is a constant parameter; d_v0 represents the oth expected value; represents the pth output value of the hth hidden layer; The weight adjustment formula of the hidden layer is: in, represents the hidden layer weight adjustment value; β is a constant; is the value of the oth neuron node in the kth hidden layer; is the connection weight between the oth neuron in the hth hidden layer and the neurons in the h+1th hidden layer; Represents the output value of the pth neuron in the hth hidden layer; is the output value of the i-th input node and the h-1-th neuron in the first hidden layer; in, is the output of the hth neuron node in the pth hidden layer; is the value of the oth neuron node in the kth hidden layer; c. Adjust the weight W after step b ih , W ho Substitute into the calculation formula of the output error to calculate the output error E. When the output error E reaches the minimum output error or the number of training times reaches the maximum number of training times, terminate the training and output the corrected output value as the final output value. Otherwise, repeat steps a and b until the output error E reaches the minimum output error or the number of training times reaches the maximum number of training times.

8. The method for identifying human-caused events in preventing wrong intervals in a nuclear power plant according to claim 1 is characterized in that: The specific steps of performing image acquisition and recognition on the operation button pointed to by the operator in S2 include: When the operator points to an operation button, an image of the mark on the operation button is captured, and the captured image is smoothed and character segmented. Character recognition is then performed to obtain the character sequence of the operation button. The character sequence is then compared with the reference character information corresponding to the task instruction. Whether the operation button character recognition is successful is determined based on the comparison result.

9. The method for identifying human-caused events in preventing wrong intervals in a nuclear power plant according to claim 8, characterized in that: The image smoothing process adopts the rotating mask averaging method, the character segmentation process adopts the vertical projection segmentation method, and the character recognition is completed by using a convolutional neural network.

10. A device for identifying human-caused events in a nuclear power plant to prevent wrong intervals, characterized in that: It includes a server, a workstation, and a collection device. The collection device is used to collect the voice of the operator's voice counting and to collect images of the operation buttons pointed to by the operator, and transmit the collected data to the server. The server is used to receive the data from the collection device and execute the steps of the method for identifying human factors events in preventing wrong intervals in a nuclear power plant as described in any one of claims 1 to 9, output the identification results and corresponding prompts or warnings and transmit them to the workstation. The workstation is used to issue prompts or warning information.

Images