Intelligent prediction method and system for TOI quality control list of nuclear power plant
By extracting TOI text features using the AM-LSTM model, the problem of low efficiency in manual review of TOI quality control forms in nuclear power plants was solved, enabling efficient and accurate generation of TOI quality control forms and reducing safety risks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NAT NUCLEAR INFORMATION TECH CO LTD
- Filing Date
- 2026-01-07
- Publication Date
- 2026-05-19
AI Technical Summary
The review of TOI quality control sheets at nuclear power plants relies on manual judgment, resulting in low work efficiency and potential safety hazards.
The AM-LSTM model is used for single-intelligence prediction of TOI quality control. Contextual features are extracted through BiLSTM layer, and semantic abstraction is performed by combining attention mechanism and fully connected layer to generate multi-dimensional probability vector and output the pre-judgment result.
This improved the efficiency and accuracy of TOI quality control sheet review and reduced security risks caused by human error.
Smart Images
Figure CN122066014A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent control technology for nuclear power plants, and in particular relates to an intelligent prediction method and system for TOI quality control sheets in nuclear power plants. Background Technology
[0002] A Temporary Operation Instruction (TOI) for a nuclear power plant is a temporary operating procedure. It specifies a special operating procedure for a limited period of time when the operating mode of a system or equipment is not included in the existing normal operating procedures, or when the system or equipment has been modified or altered, rendering some existing procedures inapplicable. The TOI quality control sheet is the process by which the checker, reviewer, and approver review the quality of the TOI when it is applied for, ensuring that reliable documentation is provided as the basis for operational monitoring and control activities even when the operating procedures are inapplicable.
[0003] Therefore, reviewing Temporary Operating Instructions (TOIs) through TOI quality control sheets is a crucial aspect of nuclear power plant safety. However, currently, the review of TOIs at nuclear power plants requires manual assessment of each TOI's purpose and content, and necessitates review of the TOI's quality by the checker, approver, and authorizer of the TOI application. These issues significantly impact the efficiency of nuclear power plant operations and may also lead to safety problems due to human error. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and to propose an intelligent prediction method and system for TOI quality control in nuclear power plants.
[0005] In view of this, the present invention proposes an intelligent prediction method for TOI quality control in nuclear power plants, comprising: Step S1: Obtain the metadata of the Temporary Operation Instruction (TOI) of the nuclear power plant, wherein the metadata includes at least the purpose of the TOI, the content of the TOI, and text information of the execution prerequisites; Step S2: Perform data preprocessing to obtain multidimensional tensors; Step S3: Input the pre-trained AM-LSTM model, in which the AM-LSTM model sequentially extracts context features through BiLSTM layers, calculates the weights of key information in the context features through attention mechanism layers, and performs element-level semantic abstraction and association mapping through fully connected layers; outputs multiple sets of three-dimensional probability vectors; Step S4: Based on the output of the AM-LSTM model, generate pre-judgment results for multiple inspection elements in the TOI quality control form, where each pre-judgment result is "yes", "no" or "not applicable". Step S5: Generate a pre-filled TOI quality control form based on the pre-judgment results.
[0006] As an improvement to the above method, step S2 includes: Structured fields are extracted using regular expressions and domain rules, and noisy data is cleaned. Perform domain-enhanced Chinese word segmentation; Irrelevant words are filtered out by a dynamic stop word list, while key quantifiers and negative adverbs are retained; We introduce entity chaining and spatiotemporal context related to equipment, system, and unit, and use a hybrid embedding method of pre-trained power word vectors and TF-IDF weighting to achieve semantic enhancement; The output is a multidimensional tensor containing lexical features, document-level security identifiers, and contextual alert status.
[0007] As an improvement to the above method, the AM-LSTM model includes an input layer, an embedding layer, a BiLSTM layer, an attention mechanism layer, a fully connected layer, and an output layer connected in sequence. The output layer comprises 13 independent 3-neuron subnetworks, each corresponding to a specific inspection element. The activation function transforms the original scores into a probability distribution, outputting a three-dimensional probability vector representing the probabilities of the checked element being "yes", "no", and "not applicable".
[0008] As an improvement to the above method, the fully connected layer performs feature-level semantic abstraction and association mapping, including: When the keyword-level features extracted by the attention layer are test-exempt devices or identification protection devices, they are upgraded to element-level semantic representations. When the keyword-level features extracted by the attention layer are layout steps, operability-related neurons are activated; A mapping relationship between text features and inspection elements is established using a weight matrix; When the keyword-level features extracted by the attention layer are kernel-safe, the activation value of the check element "TOI meets and will not deviate from TS requirements" is forcibly increased.
[0009] As an improvement to the above method, the inspection elements generated in step S4 include: Is TOI necessary? TOI is technically feasible in terms of nuclear safety and power plant availability; TOI complies with and does not deviate from TS requirements; Whether the qualifications of the drafters and reviewers meet the authorization requirements; The instructions should be accurate and easy to understand, avoiding ambiguity. The instructions are operable; Does TOI provide clear intervention criteria? For monitoring and recording, a record form must be provided and the monitoring frequency must be clearly stated; If the TOI involves operations, operation documents or a list must be provided; Are backup plans in place for the lost equipment? Whether the capabilities or functions of the backup measures are equivalent to the initial design; Have the feedback from the simulation execution been addressed? Whether temporary modifications and settings can be restored is already reflected in the TOI.
[0010] As an improvement to the above method, after step S5, the method further includes: Step S6: Receive the user's manual review and confirmation instruction for the pre-judgment result; Step S7: The confirmed TOI data and its corresponding pre-judgment results are fed back to the model training process as new training samples for continuous optimization of the AM-LSTM model.
[0011] As an improvement to the above method, the method further includes a training step of the AM-LSTM model using an established dataset; each temporary run instruction in the dataset includes: basic TOI information, unit, TOI purpose, TOI content, execution prerequisites, TOI cancellation conditions, estimated start time, estimated end time, whether it is related to nuclear safety, whether it is related to personnel safety, whether it is related to major equipment, and involves status setting fields; it also includes 13 inspection elements of the quality control sheet.
[0012] On the other hand, the present invention provides an intelligent prediction system for TOI quality control in nuclear power plants, comprising: The data acquisition module is used to acquire metadata of the Temporary Operation Instruction (TOI) of a nuclear power plant. The metadata includes at least textual information about the purpose of the TOI, the content of the TOI, and the execution prerequisites. The data preprocessing module is used to perform data preprocessing to obtain multidimensional tensors; The intelligent prediction module includes a pre-trained AM-LSTM model, which processes the input data of the model to generate pre-judgment results for multiple inspection elements in the TOI quality control form. Each pre-judgment result is "yes", "no" or "not applicable". The AM-LSTM model sequentially extracts contextual features through a BiLSTM layer, calculates the weights of key information in the contextual features through an attention mechanism layer, and performs feature-level semantic abstraction and association mapping through a fully connected layer; it outputs multiple sets of three-dimensional probability vectors; and The results generation module is used to generate a pre-filled TOI quality control form based on the pre-judgment results.
[0013] As an improvement to the above system, the data acquisition module is built on the Spring Cloud server-side framework and is used to obtain basic information and attachments of the TOI application form from the web form; the result generation module sends the pre-filled TOI quality control form back to the user page through the RESTful API.
[0014] Compared with the prior art, the advantages of the present invention are: 1. Construct a BiLSTM model to capture information such as the purpose and content of the TOI in the temporary operation instruction application form, understand the complex semantics, contextual relationships and implicit information in the instruction text, and construct a 3-neuron output layer to determine whether the temporary operation instruction meets the inspection elements in the TOI quality control form, thereby improving the review efficiency.
[0015] 2. An attention mechanism (AM) is added to the BiLSTM model to construct an AM-BiLSTM model. Adding the attention mechanism allows the attention layer to label the importance of different features, thereby increasing the accuracy and efficiency of TOI quality control. Attached Figure Description
[0016] Figure 1 This is a structural diagram of the AM-LSTM model; Figure 2 This is a flowchart of the intelligent prediction method for quality control sheets of the present invention; Figure 3 This is a diagram illustrating the basic information of the TOI; Figure 4 This is a schematic diagram of the TOI quality control sheet. Detailed Implementation
[0017] This invention proposes an intelligent prediction method for TOI quality control in nuclear power plants, comprising: Step S1: Obtain the metadata of the Temporary Operation Instruction (TOI) of the nuclear power plant, wherein the metadata includes at least the purpose of the TOI, the content of the TOI, and text information of the execution prerequisites; Step S2: Perform data preprocessing to obtain multidimensional tensors; Step S3: Input the pre-trained AM-LSTM model, in which the AM-LSTM model sequentially extracts context features through BiLSTM layers, calculates the weights of key information in the context features through attention mechanism layers, and performs element-level semantic abstraction and association mapping through fully connected layers; outputs multiple sets of three-dimensional probability vectors; Step S4: Based on the output of the AM-LSTM model, generate pre-judgment results for multiple inspection elements in the TOI quality control form, where each pre-judgment result is "yes", "no" or "not applicable". Step S5: Generate a pre-filled TOI quality control form based on the pre-judgment results.
[0018] Step S6: Receive the user's manual review and confirmation instruction for the pre-judgment result; Step S7: The confirmed TOI data and its corresponding pre-judgment results are fed back to the model training process as new training samples for continuous optimization of the AM-LSTM model.
[0019] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.
[0020] Example 1 Embodiment 1 of this invention proposes an intelligent prediction method for TOI quality control sheets in nuclear power plants. In this method, to improve the efficiency and accuracy of intelligent judgment in TOI quality control sheets, an AM-LSTM model is constructed, such as... Figure 1 As shown, the attention mechanism is incorporated into the LSTM model. The model mainly consists of an input layer, an embedding layer, a BiLSTM layer, an attention layer, and a fully connected layer. The specific model construction process is as follows: (1) Dataset preparation The dataset used for training the AM-LSTM intelligent prediction method model for TOI quality control orders was obtained through business requirement interactions with various nuclear power plants. Each temporary operation instruction in the dataset contains basic TOI information, including unit, TOI purpose, TOI content, execution prerequisites, TOI cancellation conditions, estimated start time, estimated end time, whether it is related to nuclear safety, whether it is related to personnel safety, whether it is related to critical equipment, and status setting fields. It also includes 13 inspection elements of the quality control order, as shown in Table 1.
[0021] Table 1
[0022] (2) Data preprocessing For the preprocessing of TOI text, a multi-stage specialized processing strategy is adopted: First, structured fields (such as TO purpose and TO content) are extracted using regular expressions and domain rules, and noisy data is cleaned; then, domain-enhanced Chinese word segmentation is performed; subsequently, irrelevant words are filtered out using a dynamic stop word list, while retaining key quantifiers and negation adverbs. In the semantic enhancement stage, entity chain references (equipment-system-unit association) and spatiotemporal context are introduced, and a hybrid embedding method combining pre-trained power word vectors and TF-IDF weighted embedding is used. The final output is a multi-dimensional tensor containing lexical features, document-level security identifiers, and contextual alarm status, supporting subsequent models to accurately predict inspection elements.
[0023] (3) BiLSTM model Compared to traditional Recurrent Neural Networks (RNNs), LSTM (Long Short-Term Memory) models use cleverly designed gate structures to determine whether information passes through, thus mitigating the vanishing or exploding gradient problems of RNNs to some extent. BiLSTM models, building upon LSTM, incorporate two LSTM layers (forward and backward) to extract not only preceding but also following information, establishing a connection between the two contexts. When h is the hidden layer output of a BiLSTM unit, c is the value of a BiLSTM memory cell, and x is the input data, the BiLSTM computation process is as follows: Calculate the value of the forget gate The calculation formula is shown in (1). Calculate the input node at the current time. , , The weights are the input data and the output of the BiLSTM unit at the previous time step, respectively, and the calculation formula is shown in (2).
[0024]
[0025] Calculate the value of the input gate The input gate is used to control the influence of the current input data on the state value of the memory cell, and the calculation formula is shown in (3). Calculate the state value of the memory cell at the current moment. The calculation formula is shown in (4). Calculate the output gate. The output gate is used to control the output of the state value of the memory unit, and the calculation formula is shown in (5).
[0026]
[0027] Finally, the output of the BiLSTM unit is calculated, and the calculation formulas are shown in (6) and (7).
[0028]
[0029] (4) Attention mechanism Attention mechanisms (AM) are an important research direction in deep learning, especially in fields such as natural language processing, and have broad application prospects. Attention mechanisms employ parallel computation to calculate the similarity between elements within the same sentence to obtain overall information, rather than information from a single context. Therefore, attention mechanisms can provide a more complete understanding of the entire sequence and better grasp the complex relationships between various elements. Assuming the input sequence... Output sequence The computational principle of the attention mechanism is described below.
[0030] when For query vector, For key vectors, For value vectors, , , If the matrix is a linear transformation matrix, then (Query) (Key) The formula for calculating (Value) is as follows:
[0031] The weights are calculated as shown in formula (11). Here is the attention weight matrix, where The length of the input sequence. To query the dimensions of a vector or key vector, Let be the transpose matrix of the key vectors. The function converts the dot product result into a probability distribution. The final output sequence is shown in equation (12), where... For the output sequence, Let be the transpose of the value vectors, and n and j be the positions of the input and output vectors, respectively.
[0032]
[0033] (5) Fully connected layer The fully connected layer acts as a high-level feature processor for the model, receiving context vectors (typically 128-256 dimensions) from the attention layer. This layer is implemented through non-linear transformations. Feature abstraction: Keyword-level features extracted by the attention layer (such as "exempt equipment" and "identification protection device") are upscaled to element-level semantic representations. For example, when the input contains "deployment steps," the fully connected layer activates neurons related to "operability." Element association modeling: A mapping relationship between text features and inspection elements is established through a weight matrix. For example, the strong connection weight between the feature "Unit 2" and the element "unit" (≈0.85), and the negative weight between the feature "none" and the element "record table" (≈-1.2). Domain knowledge injection: Bias constraints are set in the nuclear power scenario, such as forcibly increasing the activation value of element 3 (TS compliance) when the text contains "nuclear safety."
[0034] (6) Output results The output layer consists of 13 independent 3-neuron subnetworks, each corresponding to a specific inspection element. Each sub-output layer applies... The activation function transforms the raw scores into a probability distribution. By adjusting the weights of the loss function or employing stratified sampling, sufficient training samples are ensured for all three states of each element. The model ultimately outputs 13 sets of three-dimensional probability vectors, which are then used to label the most likely judgment.
[0035] Example 2 Embodiment 2 of the present invention proposes an intelligent prediction method for TOI quality control in nuclear power plants using an algorithmic model, such as... Figure 2 As shown. This method mainly consists of three layers: the data source layer, the algorithm model layer, and the result output layer. Using this method, the TOI quality control form can be filled out accurately and efficiently. The detailed implementation process is as follows: (1) First, use the Spring Cloud server-side framework to obtain the basic information of the TOI application form actively submitted by the web client, related attachments, and the affected procedure number, attachment / section data. For example Figure 3 As shown. Further explanation is needed. Figure 3 This is a diagram illustrating the basic information of TOI, and aims to explain the 13 inspection elements included in Table 1.
[0036] (2) Secondly, based on the Restful API, the obtained metadata is input into the AM-LSTM algorithm model built in Python; then the metadata is preprocessed to meet the standards of model training and prediction; then the trained and optimized AM-LSTM model is used to make predictions and judgments based on the data input from the input layer; finally, the judgment results of each TOI quality control check element are obtained through the multi-output classification layer.
[0037] (3) Then, the judgment results obtained from the model are parsed, and the multi-class results output by the model (such as “qualified”, “required for review”, “unqualified”, etc.) are mapped to each specific inspection element corresponding to the TOI quality control sheet, and structured judgment result data is generated.
[0038] (4) Finally, the parsed structured judgment result data is sent back to the user page via the Restful API to generate a preliminary TOI quality control form, which is then manually reviewed and finally confirmed. Figure 4 As shown. Further explanation is needed. Figure 4 This is a schematic diagram of a TOI quality control sheet. It is intended to illustrate the TOI number, TOI purpose, TOI content, execution prerequisites, and TOI cancellation conditions. The fields to be filled in are for illustrative purposes only.
[0039] (5) Knowledge base update: The final result confirmed by the user and the corresponding input data (under the premise of desensitization and compliance) are used as new training samples and are periodically fed back to the model training process for continuous iterative optimization of the model.
[0040] Example 3 Embodiment 3 of the present invention provides an intelligent prediction system for TOI quality control sheets in nuclear power plants, implementing the above-mentioned method example, including: The data acquisition module is used to acquire metadata of the Temporary Operation Instruction (TOI) of a nuclear power plant. The metadata includes at least textual information about the purpose of the TOI, the content of the TOI, and the execution prerequisites. The data preprocessing module is used to perform data preprocessing to obtain multidimensional tensors; The intelligent prediction module includes a pre-trained AM-LSTM model, which processes the input data of the model to generate pre-judgment results for multiple inspection elements in the TOI quality control form. Each pre-judgment result is "yes", "no" or "not applicable". The AM-LSTM model sequentially extracts contextual features through a BiLSTM layer, calculates the weights of key information in the contextual features through an attention mechanism layer, and performs feature-level semantic abstraction and association mapping through a fully connected layer; it outputs multiple sets of three-dimensional probability vectors; and The results generation module is used to generate a pre-filled TOI quality control form based on the pre-judgment results.
[0041] The data acquisition module is built on the Spring Cloud server-side framework and is used to obtain basic information and attachments of the TOI application form from the web form; the result generation module sends the pre-filled TOI quality control form back to the user page through the RESTful API.
[0042] It is worth noting that in the embodiments of the above system, the modules included are divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional module are only for easy differentiation and are not used to limit the scope of protection of the present invention.
[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for intelligent prediction of TOI quality control sheets in nuclear power plants, comprising: Step S1: Obtain the metadata of the Temporary Operation Instruction (TOI) of the nuclear power plant, wherein the metadata includes at least the purpose of the TOI, the content of the TOI, and text information of the execution prerequisites; Step S2: Perform data preprocessing to obtain multidimensional tensors; Step S3: Input into the pre-trained AM-LSTM model, the AM-LSTM model sequentially extracts context features through the BiLSTM layer, calculates the weights of key information in the context features through the attention mechanism layer, and performs element-level semantic abstraction and association mapping through the fully connected layer; Output multiple sets of three-dimensional probability vectors; Step S4: Based on the output of the AM-LSTM model, generate pre-judgment results for multiple inspection elements in the TOI quality control form, where each pre-judgment result is "yes", "no" or "not applicable". Step S5: Generate a pre-filled TOI quality control form based on the pre-judgment results.
2. The intelligent prediction method for TOI quality control sheets in nuclear power plants according to claim 1, characterized in that, Step S2 includes: Structured fields are extracted using regular expressions and domain rules, and noisy data is cleaned. Perform domain-enhanced Chinese word segmentation; Irrelevant words are filtered out by a dynamic stop word list, while key quantifiers and negative adverbs are retained; We introduce entity chaining and spatiotemporal context related to equipment, system, and unit, and use a hybrid embedding method of pre-trained power word vectors and TF-IDF weighting to achieve semantic enhancement; The output is a multidimensional tensor containing lexical features, document-level security identifiers, and contextual alert status.
3. The intelligent prediction method for TOI quality control sheets in nuclear power plants according to claim 1, characterized in that, The AM-LSTM model comprises an input layer, an embedding layer, a BiLSTM layer, an attention mechanism layer, a fully connected layer, and an output layer, connected sequentially. The output layer consists of 13 independent 3-neuron subnetworks, each corresponding to a specific inspection element. The activation function transforms the original scores into a probability distribution, outputting a three-dimensional probability vector representing the probabilities of the checked element being "yes", "no", and "not applicable".
4. The intelligent prediction method for TOI quality control sheets in nuclear power plants according to claim 1, characterized in that, The fully connected layer performs feature-level semantic abstraction and association mapping, including: When the keyword-level features extracted by the attention layer are test-exempt devices or identification protection devices, they are upgraded to element-level semantic representations. When the keyword-level features extracted by the attention layer are layout steps, operability-related neurons are activated; A mapping relationship between text features and inspection elements is established using a weight matrix; When the keyword-level features extracted by the attention layer are kernel-safe, the activation value of the check element "TOI meets and will not deviate from TS requirements" is forcibly increased.
5. The intelligent prediction method for TOI quality control sheets in nuclear power plants according to claim 4, characterized in that, The inspection elements generated in step S4 include: Is TOI necessary? TOI is technically feasible in terms of nuclear safety and power plant availability; TOI complies with and does not deviate from TS requirements; Whether the qualifications of the drafters and reviewers meet the authorization requirements; The instructions should be accurate and easy to understand, avoiding ambiguity. The instructions are operable; Does TOI provide clear intervention criteria? For monitoring and recording, a record form must be provided and the monitoring frequency must be clearly stated; If the TOI involves operations, operation documents or a list must be provided; Are backup plans in place for the lost equipment? Whether the capabilities or functions of the backup measures are equivalent to the initial design; Have the feedback from the simulation execution been addressed? Whether temporary modifications and settings can be restored is already reflected in the TOI.
6. The intelligent prediction method for TOI quality control sheets in nuclear power plants according to claim 1, characterized in that, Following step S5, the method further includes: Step S6: Receive the user's manual review and confirmation instruction for the pre-judgment result; Step S7: The confirmed TOI data and its corresponding pre-judgment results are fed back to the model training process as new training samples for continuous optimization of the AM-LSTM model.
7. The intelligent prediction method for TOI quality control sheets in nuclear power plants according to claim 1, characterized in that, The method also includes a training step of the AM-LSTM model using the established dataset; each temporary operation instruction in the dataset includes: basic TOI information, unit, TOI purpose, TOI content, execution prerequisites, TOI cancellation conditions, estimated start time, estimated end time, whether it is related to nuclear safety, whether it is related to personnel safety, whether it is related to major equipment, and involves status setting fields; it also includes 13 inspection elements of the quality control sheet.
8. A smart prediction system for TOI quality control sheets in nuclear power plants, characterized in that, include: The data acquisition module is used to acquire metadata of the Temporary Operation Instruction (TOI) of a nuclear power plant. The metadata includes at least textual information about the purpose of the TOI, the content of the TOI, and the execution prerequisites. The data preprocessing module is used to perform data preprocessing to obtain multidimensional tensors; The intelligent prediction module includes a pre-trained AM-LSTM model, which processes the input data of the model to generate pre-judgment results for multiple inspection elements in the TOI quality control form. Each pre-judgment result is "yes", "no" or "not applicable". The AM-LSTM model extracts contextual features sequentially through a BiLSTM layer, calculates the weights of key information in the contextual features through an attention mechanism layer, and performs element-level semantic abstraction and association mapping through a fully connected layer. Output multiple sets of three-dimensional probability vectors; and The results generation module is used to generate a pre-filled TOI quality control form based on the pre-judgment results.
9. The intelligent prediction system for nuclear power plant TOI quality control sheets according to claim 8, characterized in that, The data acquisition module is built on the Spring Cloud server-side framework and is used to obtain basic information and attachments of the TOI application form from the web form; the result generation module sends the pre-filled TOI quality control form back to the user page through the RESTful API.