Voice reminding method and system based on nuclear power high-risk operation prompt

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to a voice prompting method and system based on nuclear power high-risk operation prompting. Comprising the following steps of 1, preprocessing a text; step 2, feature extraction; 3, acoustic modeling is carried out; and 4, training and optimizing. The system has the beneficial effects that the working efficiency and accuracy are improved: through the voice intelligent reminding system, a worker can directly hear important step reminding of high-risk operation without depending on text notification or manual voice broadcast, so that the time and errors of information transmission are reduced, the possibility of text understanding or reading errors is avoided, and the working efficiency and accuracy are improved. And errors caused when the worker executes high-risk operation are reduced, and the working efficiency is improved.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and in particular relates to a voice reminder method and system based on high-risk nuclear power operation prompts. Background Art

[0002] In the field of nuclear power, the execution of high-risk nuclear power operations is a complex and important process. In order to ensure the accurate execution and safety of work order operations, it is necessary to provide an effective reminder method. At present, most power plant reminder systems mainly rely on the system's push text or sound method. Users need to use electronic devices to view relevant information about high-risk operations, but this method has problems such as untimely information transmission and limited effect. Summary of the invention

[0003] The purpose of the present invention is to provide a voice reminder method and system based on nuclear power high-risk operation prompts, which realizes voice broadcast reminders of important steps of work orders through end-to-end speech synthesis. It has the characteristics of high efficiency, accuracy and safety, and can be widely used in the nuclear power industry and other high-risk operation fields to improve work efficiency and safety.

[0004] The technical solution of the present invention is as follows: A voice reminder method based on nuclear power high-risk operation prompts, comprising the following steps:

[0005] Step 1: Text preprocessing;

[0006] Step 2: Feature extraction;

[0007] Step 3: Acoustic modeling;

[0008] Step 4: Training and optimization.

[0009] In the step 1, the input work order text is preprocessed, including word segmentation and punctuation processing, and the text is converted into a discrete word sequence.

[0010] In the step 2, the text sequence is encoded using a Transformer network structure, which includes multiple encoder layers, each encoder layer includes a self-attention mechanism and a feedforward neural network, the self-attention mechanism models the relationship between words and captures global context information, and the feedforward neural network is used to perform nonlinear transformation on the features of each position, and after processing by multiple encoder layers, the feature representation of each position is obtained.

[0011] In the step 3, the feature representation of the encoder output is taken as input, and acoustic modeling is performed through the decoder to generate a speech waveform. The decoder consists of multiple layers, each of which contains a self-attention mechanism, an encoder-decoder attention mechanism and a feedforward neural network. The self-attention mechanism is used to model the context information inside the decoder, the encoder-decoder attention mechanism is used to associate the output of the encoder with the current position of the decoder, and the feedforward neural network is used to perform nonlinear transformation on the features of each position. Finally, the output of the decoder is mapped to a speech waveform through a linear layer.

[0012] In step 4, a training data set with a target speech waveform is used to train the speech synthesis model. During the training process, a cross entropy loss function is used to calculate the difference between the model output and the target speech waveform, and the parameters of the model are updated through a back propagation algorithm.

[0013] A voice reminder system based on nuclear power high-risk operation prompts includes a model training module, a high-risk prompt setting module, a voice synthesis module, a voice broadcast module, a reminder strategy module and a user interaction module.

[0014] The model training module trains the text data and speech data set of nuclear power work orders. The data are from work orders in the nuclear power field, containing nuclear power specific data and proper nouns. The training data set with the target speech waveform is used to train the speech synthesis model based on the Transformer network structure. The model training module trains the text data and speech data set of nuclear power work orders. The training data set with the target speech waveform is used to train the speech synthesis model based on the Transformer network structure, including the following:

[0015] Data preparation: Collect text data of nuclear power work orders, including historical work orders and standard operating procedure texts;

[0016] Voice data set: record or obtain voice data corresponding to the work order text;

[0017] Data preprocessing: perform word segmentation and punctuation processing on text data, and extract features from speech data;

[0018] Model training:

[0019] Model selection: Use the Transformer network structure, including encoder and decoder;

[0020] Training process: Use a training dataset with a target speech waveform to train the Transformer model, use the cross entropy loss function to calculate the difference between the model output and the target speech waveform, and update the model parameters through the back propagation algorithm.

[0021] The high-risk warning setting module analyzes the work order data sent by the user through the terminal, receives and parses the work order to be executed, which can be input in text form or phonetic character form, including:

[0022] Data reception and analysis:

[0023] User input: Provide a user interface or API interface for users to enter work order data;

[0024] Data parsing: Parse the input work order data to extract key steps and important information.

[0025] The speech synthesis module adopts the Transformer network structure neural network model to convert the important step text or phonetic characters in the work order into speech. The specific neural network model includes an encoder-decoder architecture, in which the encoder converts the text sequence into an intermediate representation, and the decoder generates the corresponding speech waveform according to the intermediate representation. Between the encoder and the decoder, an attention mechanism is used to enhance the attention to the input information and the accuracy of the generated speech. The neural network model architecture adopts the Transformer network structure and is trained in combination with the cross entropy loss function and the optimization algorithm.

[0026] The voice broadcast module plays the generated voice through an audio playback device or a voice presentation device, and the voice broadcast module supports different audio output modes; the reminder strategy module controls the timing and frequency of voice broadcast according to the importance, urgency and sequence of execution steps of the work order to ensure the accuracy and timeliness of the reminder information.

[0027] The beneficial effects of the present invention are: improving work efficiency and accuracy: through the voice intelligent reminder system, workers do not need to rely on text notifications or manual voice broadcasts, and can directly hear reminders of important steps in high-risk operations, reducing the time and errors in information transmission, avoiding the possibility of errors in text comprehension or interpretation, reducing errors made by workers when performing high-risk operations, and improving work efficiency.

[0028] Increase safety of nuclear power work: Through the voice intelligent reminder system, reminders of important steps can be conveyed to staff in a timely manner, reducing potential safety hazards in high-risk operations and improving the safety of the work site.

[0029] Application of end-to-end speech synthesis technology: The end-to-end speech synthesis technology is combined with a specific neural network model to achieve direct conversion from text to speech, eliminating the intermediate links in traditional speech synthesis and improving the naturalness and accuracy of synthesized speech.

[0030] System modular design: The system of the present invention adopts modular design, which not only makes the system structure clear and the role of each module clear, but also facilitates the expansion and maintenance of the system, and improves the flexibility and scalability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A flow chart of a voice reminder method based on nuclear power high-risk operation prompts provided by the present invention;

[0032] Figure 2 This is a schematic diagram of a voice reminder system based on nuclear power high-risk operation prompts provided by the present invention. DETAILED DESCRIPTION

[0033] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] The present invention provides a voice reminder method and system based on nuclear power high-risk operation prompts. When an AWT work order is started, the high-risk operation precautions are targeted to achieve accurate, timely and efficient voice prompt execution of high-risk operations, thereby ensuring the stability of the execution of nuclear power high-risk operations and improving safety production efficiency.

[0035] The present invention relates to the field of natural language processing, and provides a voice reminder method and system based on nuclear power high-risk operation prompts. The scheme is based on a nuclear power workflow management information system, and the present invention uses a Transformer network structure. Through the application and practice of the method and system, the implementation efficiency of nuclear power high-risk operations is improved, and the problem that the traditional method has many user viewing prompts and subsequent processing and cannot meet actual needs is solved.

[0036] Compared with traditional text or voice push methods, this method achieves accuracy and timeliness of information transmission through voice reminders. Workers do not need to rely on electronic devices to view relevant information, and can obtain important step reminders for high-risk operations through voice broadcasts, thereby improving work efficiency.

[0037] Using the Transformer network structure and natural language processing technology, this method can better model the relationship between texts and capture global context information, thereby improving the quality and accuracy of speech synthesis.

[0038] Through the reminder strategy module, the method can intelligently control the timing and frequency of voice broadcasts according to the characteristics of the work order and the sequence of execution steps, ensure the accuracy and timeliness of the reminder information, and reduce unnecessary interference.

[0039] The user interaction module provides personalized setting options, allowing users to customize according to their needs and preferences, improving the user experience.

[0040] The reason why this method chooses to use the Transformer network structure is that Transformer has achieved remarkable results in natural language processing tasks and has the following advantages:

[0041] 1. Capturing long-distance dependencies: Traditional recurrent neural networks (RNNs) are prone to gradient vanishing or gradient exploding problems when processing long sequences, making it difficult to capture long-distance dependencies. Transformer, on the other hand, can directly model the relationship between words through the self-attention mechanism, without relying on a fixed-length window, thereby better capturing long-distance dependencies.

[0042] 2. Strong parallel computing capability: Due to the time-series dependency of RNN, they need to perform calculations in sequence, which makes it difficult to perform parallel calculations, limiting their efficiency on large-scale data. However, the self-attention mechanism in Transformer can be calculated in parallel, making the model training faster and able to process larger-scale data.

[0043] 3. Global context information: Traditional RNN models process words one by one, and each word can only see the previous context information. The self-attention mechanism in Transformer can model all words at once, capture global context information, and help generate more accurate and fluent speech.

[0044] 4. Parameter sharing and efficiency: Traditional RNN models need to store and calculate a large number of intermediate states when processing long sequences, resulting in a large number of model parameters and low computational efficiency. Transformer achieves global information interaction through the self-attention mechanism, avoiding RNN state transfer, better parameter sharing, and higher computational efficiency.

[0045] like Figure 1 As shown, a voice reminder method based on nuclear power high-risk operation prompts includes the following steps:

[0046] Step 1: Text Preprocessing

[0047] The input work order text is preprocessed, including word segmentation and punctuation processing, to convert the text into a discrete word sequence.

[0048] Word segmentation: Perform word segmentation on the input work order text, that is, segment the text string into word sequences. For example, the input text "Please check reactor pressure" is segmented into "please", "check", "reactor", and "pressure".

[0049] Punctuation processing: Process the punctuation in the text, you can choose to remove or retain it, and adjust it according to specific needs. For example, you can remove symbols such as periods and commas to make subsequent processing more concise.

[0050] Word vector representation: Convert the word sequence after word segmentation into a word vector. You can use a pre-trained word embedding model (such as Word2Vec, GloVe, or BERT) to represent each word as a vector of fixed dimension. For example, the word "check" may be represented as a 300-dimensional vector.

[0051] Step 2: Feature Extraction

[0052] The text sequence is encoded using a Transformer network structure, which includes multiple encoder layers, each of which contains a self-attention mechanism and a feedforward neural network. The self-attention mechanism can model the relationship between words and capture global context information. The feedforward neural network is used to perform nonlinear transformation on the features of each position. After processing by multiple encoder layers, the feature representation of each position is obtained.

[0053] In the feature extraction step, the text sequence is encoded using the Transformer network structure. The Transformer network consists of multiple encoder layers, each of which contains a self-attention mechanism and a feedforward neural network.

[0054] The self-attention mechanism is used to calculate the importance of each word in the sequence and capture global context information. The formula is as follows:

[0055]

[0056] Among them: Q (query matrix), K (key matrix), V (value matrix) are obtained by linear transformation of the input word vector. k is a scaling factor, typically equal to the dimension of the key vector.

[0057] The feedforward neural network is used to perform nonlinear transformation on the features at each position. The formula is as follows:

[0058] FFN(x)=max(0,xW1+b1)W2+b2

[0059] Where: W1 and W2 are weight matrices. b1 and b2 are bias vectors. max(0,x) is the ReLU activation function. The calculation steps are as follows:

[0060] The first layer of linear transformation and ReLU activation: After the input feature vector x undergoes linear transformation xW1+b1, the ReLU activation function is applied.

[0061] The second layer of linear transformation: The activated output undergoes another linear transformation (max(0,xW1+b1))W2+b2.

[0062] After being processed by multiple encoder layers, feature representations are obtained for each position. These feature representations contain the relationship and context information between words.

[0063] Step 3: Acoustic Modeling

[0064] The feature representation of the encoder output is used as input, and acoustic modeling is performed through the decoder to generate a speech waveform. The decoder also consists of multiple layers, each of which contains a self-attention mechanism, an encoder-decoder attention mechanism, and a feedforward neural network. The self-attention mechanism is used to model the contextual information inside the decoder. The encoder-decoder attention mechanism is used to associate the output of the encoder with the current position of the decoder. The feedforward neural network is used to perform nonlinear transformations on the features of each position. Finally, the output of the decoder is mapped to a speech waveform through a linear layer.

[0065] The acoustic modeling part uses a decoder to convert the feature representation of the encoder output into a speech waveform. The decoder also consists of multiple layers, each of which contains a self-attention mechanism, an encoder-decoder attention mechanism, and a feedforward neural network.

[0066] The encoder-decoder attention mechanism is used to associate the output of the encoder with the current position of the decoder. The formula is as follows:

[0067]

[0068] Here, Q is the query matrix at the decoder input, and K and V are the key matrix and value matrix at the encoder output.

[0069] The feedforward neural network in the decoder is similar to the encoder and is used to perform nonlinear transformation on the features at each position. Finally, the output of the decoder is mapped to a speech waveform through a linear layer:

[0070] Output=Linear(x)

[0071] Here, x is the output feature of the decoder, and the linear layer converts it into a speech signal.

[0072] The details are as follows:

[0073] Linear transformation: The feature vector x output by the decoder undergoes a linear transformation y=xW+b.

[0074] Among them, W is the weight matrix and b is the bias vector.

[0075] Generate speech waveform: Use the output of the linear layer as speech waveform data.

[0076] The speech synthesis method based on the Transformer network uses an encoder-decoder structure to effectively convert text input into speech output. Through the combination of self-attention mechanism and feedforward neural network, it can capture the global context information of the text and generate high-quality speech prompts.

[0077] Step 4: Training and Optimization

[0078] The speech synthesis model is trained using a training dataset with a target speech waveform. During the training process, the cross entropy loss function is used to calculate the difference between the model output and the target speech waveform, and the model parameters are updated through the back propagation algorithm.

[0079] Among them, the self-attention mechanism formula is expressed as:

[0080]

[0081] Where Q is the query vector, K is the key vector, V is the value vector, and d k is the vector dimension.

[0082] The encoder-decoder attention mechanism is expressed as:

[0083] MultiHead(Q,K,V)=Concat(h1,...,hn)W O

[0084] hi=Attention(QWQi,KWKi,VWVi)

[0085] Where h is the number of heads, W Qi , W Ki , W Vi , W O is the parameter matrix.

[0086] Among them, the objective function uses the cross entropy loss function to measure the difference between the model output and the target speech waveform:

[0087] LOSS=-i∑tilog(pi)

[0088] Among them, ti is the i-th value of the target speech waveform, and pi is the i-th value of the model output.

[0089] like Figure 2 As shown, a voice reminder system based on nuclear power high-risk operation prompts is based on the field of artificial intelligence and natural language processing technology, including: a model training module, a high-risk prompt setting module, a speech synthesis module, a speech broadcast module, a reminder strategy module and a user interaction module.

[0090] Model training module: The data trained on nuclear power work order text data and voice data sets come from work orders in the nuclear power field, contain nuclear power-specific data and proper nouns, and have nuclear power characteristics. The training data set with the target speech waveform is used to train the speech synthesis model based on the Transformer network structure.

[0091] By training the nuclear power work order text data and speech data set, the speech synthesis model based on the Transformer network structure is trained using the training data set with the target speech waveform. The main technical means are as follows:

[0092] 1. Data preparation:

[0093] Nuclear power work order text data: Collect a large amount of nuclear power work order text data, including historical work orders and standard operating procedure texts.

[0094] Voice data set: Record or obtain voice data corresponding to the work order text to ensure that the voice data can accurately reflect the text content.

[0095] Data preprocessing: perform word segmentation and punctuation processing on text data, and perform feature extraction on speech data, such as MFCC (Mel-frequency cepstral coefficient) feature extraction.

[0096] 2. Model training:

[0097] Model selection: The Transformer network structure is adopted, which includes two main parts: encoder and decoder.

[0098] Training process: Use the training dataset with the target speech waveform to train the Transformer model. Use the cross entropy loss function to calculate the difference between the model output and the target speech waveform, and update the model parameters through the back propagation algorithm.

[0099] Optimization algorithm: Commonly used optimization algorithms such as Adam are used to accelerate model convergence.

[0100] High-risk warning setting module: By analyzing the work order data sent by the user through the terminal, receiving and parsing the work orders to be executed, the work orders can be entered in text form or phonetic character form.

[0101] Analyze the work order data sent by the user through the terminal, receive and parse the work order to be executed, which can be input in text form or phonetic character form. The main technical means include:

[0102] Data reception and analysis:

[0103] User input: Provide a user interface (GUI) or API interface for users to enter work order data.

[0104] Data parsing: Parse the input work order data, extract key steps and important information, and use natural language processing techniques such as named entity recognition (NER) and dependency parsing.

[0105] Speech synthesis module: Using end-to-end speech synthesis technology and a Transformer network structure neural network model, the important step text or phonetic characters in the work order are converted into natural and fluent speech. The specific neural network model includes an encoder-decoder architecture, in which the encoder converts the text sequence into an intermediate representation, and the decoder generates the corresponding speech waveform based on the intermediate representation. Between the encoder and the decoder, an attention mechanism can be used to enhance the attention to the input information and the accuracy of the generated speech. The neural network model architecture adopts the Transformer network structure and is trained in combination with the cross entropy loss function and optimization algorithm.

[0106] The main technical means include:

[0107] 1. Text encoding:

[0108] Encoder: The preprocessed text sequence is input into the encoder to generate an intermediate representation. The encoder contains multiple layers of self-attention mechanism and feedforward neural network, and each layer extracts features from the input sequence.

[0109] 2. Voice Generation:

[0110] Decoder: The decoder receives the intermediate representation of the encoder output and combines the self-attention mechanism and the encoder-decoder attention mechanism to generate the corresponding speech waveform.

[0111] Linear layer: The output of the decoder is mapped into a speech waveform through a linear layer.

[0112] 3. Model training:

[0113] Loss function: The cross entropy loss function is used to measure the difference between the generated speech and the target speech.

[0114] Optimization process: Optimize model parameters through gradient descent method and back propagation algorithm.

[0115] Voice broadcast module: The generated voice is played through an audio playback device or a voice presentation device so that the staff can hear the important step reminders of high-risk operations. The voice broadcast module can support different audio output methods, such as speakers, headphones, etc. The main technical means include:

[0116] Audio device: supports multiple audio output methods, such as speakers, headphones, etc.

[0117] Playback control: Implement audio playback control functions, such as play, pause, stop, etc., through audio playback libraries (such as PyAudio, ALSA, etc.).

[0118] Reminder strategy module: According to the importance, urgency and sequence of execution steps of the work order, the reminder strategy module is designed to control the timing and frequency of voice broadcasts to ensure the accuracy and timeliness of reminder information. The main technical means include:

[0119] 1. Strategy design:

[0120] Priority setting: Define the importance and urgency of different types of tickets and set the priority of reminders.

[0121] Time control: Control the timing and frequency of voice broadcast according to the sequence of execution steps and actual operation conditions.

[0122] 2. Scheduling algorithm: Use the scheduling algorithm priority queue to manage reminder tasks to ensure that high-priority tasks are reported in a timely manner.

[0123] User interaction module: According to user preferences, the reminder method can be customized including volume, whether to enable, male or female voice version and other parameters. The main technical means include:

[0124] 1. User settings:

[0125] Settings interface: Provides a user interface for users to personalize settings, including volume, voice gender selection, enabling or disabling reminders, etc.

[0126] 2. Configuration storage:

[0127] Storage method: User settings can be saved in a local configuration file or a cloud database for the system to read and apply at runtime.

[0128] Data interaction between system modules: The preprocessing module converts the work order text into a word sequence and passes it to the feature extraction module. The intermediate representation generated by the feature extraction module is passed to the acoustic modeling module for generating a speech waveform. The speech waveform generated by the acoustic modeling module is passed to the speech broadcast module for actual playback. The model training module provides a trained model for use by the speech synthesis module. The work order data parsed by the high-risk warning setting module is passed to the speech synthesis module. The reminder strategy module controls the timing and frequency of the speech broadcast module. The setting parameters of the user interaction module affect the behavior of the reminder strategy module and the speech broadcast module.

[0129] Example:

[0130] A voice reminder method based on nuclear power high-risk operation prompts includes the following steps:

[0131] Step 001: Model training. This step includes: Text preprocessing: preprocess the input work order text, including word segmentation, punctuation processing, etc., to convert the text into a discrete word sequence. Feature extraction, acoustic modeling, training and optimization. The specific neural network model adopts an encoder-decoder architecture, in which the encoder converts the text sequence into an intermediate representation, and the decoder generates the corresponding speech waveform based on the intermediate representation. In order to enhance the focus on input information and generation accuracy, the model combines the attention mechanism. The system uses a cross-entropy loss function and an optimization algorithm to train the neural network model to improve the quality and accuracy of speech synthesis.

[0132] Step 002: Work order data analysis. In this step, the system receives and analyzes the work order data sent by the user terminal. The system can parse the work order to be executed. The work order can be input in text form or phonetic character form. By analyzing the work order data, the system can extract the important step information required to perform high-risk operations.

[0133] Step 003: Reminder strategy, including two parts: 1. Analyze whether the data format of the work order meets the requirements. If not, remind the user that the data format of the work order does not meet the prompt requirements. 2. According to the importance, urgency and sequence of execution steps of the work order, the system designs a reminder strategy module to control the timing and frequency of voice broadcast to ensure the accuracy and timeliness of reminder information. Through flexible reminder strategies, the system can provide effective voice prompts for high-risk operations according to actual conditions, thereby improving the efficiency and safety of operation execution.

[0134] Step 004: User interaction. The system provides a user interaction module, which can customize the reminder method according to the user's preferences and settings. The user can adjust the volume, enable or disable the voice reminder function, and select the male and female version of the voice to meet personalized needs and preferences.

[0135] Step 005: Speech synthesis, in the speech synthesis module, the system adopts end-to-end speech synthesis technology and uses a neural network model with a Transformer network structure. This model converts the important step text or phonetic characters in the work order into natural and fluent speech.

[0136] Step 006: Voice broadcast, the generated voice is played through an audio playback device or a voice presentation device so that the staff can hear the important step reminders of high-risk operations. The voice broadcast module supports different audio output methods, such as speakers, headphones, etc., to adapt to different usage scenarios and needs.

Claims

1. A voice reminder method based on nuclear power high-risk operation prompts, characterized in that: The steps include: Step 1: Text preprocessing; Step 2: Feature extraction; Step 3: Acoustic modeling; Step 4: Training and optimization.

2. A voice reminder method based on nuclear power high-risk operation prompts as claimed in claim 1, characterized in that: In the step 1, the input work order text is preprocessed, including word segmentation and punctuation processing, and the text is converted into a discrete word sequence.

3. A voice reminder method based on nuclear power high-risk operation prompts as claimed in claim 1, characterized in that: In the step 2, the text sequence is encoded using a Transformer network structure, which includes multiple encoder layers, each encoder layer includes a self-attention mechanism and a feedforward neural network, the self-attention mechanism models the relationship between words and captures global context information, and the feedforward neural network is used to perform nonlinear transformation on the features of each position, and after processing by multiple encoder layers, the feature representation of each position is obtained.

4. A voice reminder method based on nuclear power high-risk operation prompts as claimed in claim 1, characterized in that: In the step 3, the feature representation of the encoder output is taken as input, and acoustic modeling is performed through the decoder to generate a speech waveform. The decoder consists of multiple layers, each of which contains a self-attention mechanism, an encoder-decoder attention mechanism and a feedforward neural network. The self-attention mechanism is used to model the context information inside the decoder, the encoder-decoder attention mechanism is used to associate the output of the encoder with the current position of the decoder, and the feedforward neural network is used to perform nonlinear transformation on the features of each position. Finally, the output of the decoder is mapped to a speech waveform through a linear layer.

5. A voice reminder method based on nuclear power high-risk operation prompts as claimed in claim 1, characterized in that: In step 4, a training data set with a target speech waveform is used to train the speech synthesis model. During the training process, a cross entropy loss function is used to calculate the difference between the model output and the target speech waveform, and the parameters of the model are updated through a back propagation algorithm.

6. A voice reminder system based on high-risk nuclear power operations, characterized in that: It includes model training module, high-risk warning setting module, speech synthesis module, speech broadcast module, reminder strategy module and user interaction module.

7. A voice reminder system based on nuclear power high-risk operation prompts as claimed in claim 6, characterized in that: The model training module trains the text data and speech data set of nuclear power work orders. The data are from work orders in the nuclear power field, containing nuclear power specific data and proper nouns. The training data set with the target speech waveform is used to train the speech synthesis model based on the Transformer network structure. The model training module trains the text data and speech data set of nuclear power work orders. The training data set with the target speech waveform is used to train the speech synthesis model based on the Transformer network structure, including the following: Data preparation: Collect text data of nuclear power work orders, including historical work orders and standard operating procedure texts; Voice data set: record or obtain voice data corresponding to the work order text; Data preprocessing: perform word segmentation and punctuation processing on text data, and extract features from speech data; Model training: Model selection: Use the Transformer network structure, including encoder and decoder; Training process: Use a training dataset with a target speech waveform to train the Transformer model, use the cross entropy loss function to calculate the difference between the model output and the target speech waveform, and update the model parameters through the back propagation algorithm.

8. A voice reminder system based on nuclear power high-risk operation prompts as claimed in claim 6, characterized in that: The high-risk warning setting module analyzes the work order data sent by the user through the terminal, receives and parses the work order to be executed, which can be input in text form or phonetic character form, including: Data reception and analysis: User input: Provide a user interface or API interface for users to enter work order data; Data parsing: Parse the input work order data to extract key steps and important information.

9. A voice reminder system based on nuclear power high-risk operation prompts as claimed in claim 6, characterized in that: The speech synthesis module adopts the Transformer network structure neural network model to convert the important step text or phonetic characters in the work order into speech. The specific neural network model includes an encoder-decoder architecture, in which the encoder converts the text sequence into an intermediate representation, and the decoder generates the corresponding speech waveform according to the intermediate representation. Between the encoder and the decoder, an attention mechanism is used to enhance the attention to the input information and the accuracy of the generated speech. The neural network model architecture adopts the Transformer network structure and is trained in combination with the cross entropy loss function and the optimization algorithm.

10. A voice reminder system based on nuclear power high-risk operation prompts as claimed in claim 6, characterized in that: The voice broadcast module plays the generated voice through an audio playback device or a voice presentation device, and the voice broadcast module supports different audio output modes; the reminder strategy module controls the timing and frequency of voice broadcast according to the importance, urgency and sequence of execution steps of the work order to ensure the accuracy and timeliness of the reminder information.

Images