Artificial intelligence broadcasting system based on Beidou satellite positioning and TTS framework

Abstract

The present invention discloses an artificial intelligence broadcasting system based on Beidou satellite positioning and TTS framework, the system includes a positioning acquisition and calibration module, a regional management and redline processing module, a positioning calibration machine learning algorithm application module, a TTS speech synthesis module, and a user data security analysis module; obtains geographic environment and positioning data, integrates and calibrates the geographic environment and positioning data, analyzes and trains the geographic environment and positioning data and outputs visualization results; converts the geographic environment and positioning data into voice signals and broadcasts them in real time; and encrypts the user positioning data. The system can improve the voice emotion intensity of the broadcast and the triggerability of the broadcast content, promote the automation of the broadcast by artificial intelligence, and introduce Beidou satellite positioning to accurately locate the user's area and range, focusing on the combination of user actual needs and travel services.

Description

Technical Field

[0001] The present invention relates to the field of navigation broadcasting, and in particular to an artificial intelligence broadcasting system based on Beidou satellite positioning and a TTS framework. Background Art

[0002] Smart broadcasting refers to the use of artificial intelligence technology to integrate text, audio, video and other media forms, and to achieve automated and personalized information broadcasting services through speech synthesis, image recognition, and natural language processing technologies. It can automatically select the appropriate broadcasting method, content, and time according to user needs and preferences, and provide users with efficient and convenient information acquisition services. At the same time, the smart broadcasting system can also continuously optimize and improve its own broadcasting effects and service quality based on user feedback and evaluation, thereby improving user satisfaction and loyalty. In addition, there have been certain attempts in the research of intelligent broadcasting in China, but there are still some shortcomings, which are mainly manifested in the following aspects: First, the lack of emotional expression: Although intelligent broadcasting can convert text information into speech, it lacks emotional expression and cannot convey emotions like human anchors, which makes intelligent broadcasting sound a bit stiff and mechanical; second, the voice intonation is unnatural: the voice intonation of intelligent broadcasting is often not natural enough, lacking the ups and downs and sense of language of human anchors, which makes intelligent broadcasting sound not smooth and comfortable enough; third, information security and privacy protection issues: intelligent broadcasting needs to collect users' voice data and other personal information, which may cause information security and privacy protection issues, and data security and privacy protection measures need to be strengthened; fourth, there is a lack of integration with the surrounding environment or existing business. Many intelligent broadcasts are still in the system manually triggered broadcast type, and lack integration with the business.

[0003] Therefore, it is of great practical value to develop an artificial intelligence broadcasting system based on Beidou satellite positioning and TTS framework to provide triggered intelligent broadcasting for users' travel. Summary of the invention

[0004] In view of the above-mentioned prior art, the purpose of the present invention is to provide an artificial intelligence broadcasting system based on Beidou satellite positioning and TTS framework, which can strengthen the active broadcasting mechanism with actual business, improve the voice emotion intensity and triggerability of the broadcasting content, promote the automation of artificial intelligence for broadcasting, and introduce Beidou satellite positioning to accurately locate the user's area and range, focusing on the combination of user's actual needs and travel services.

[0005] In order to achieve the above object, the technical solution provided by the present invention is:

[0006] An artificial intelligence broadcasting system based on Beidou satellite positioning and TTS framework, including a positioning acquisition and calibration module, an area management and redline processing module, a positioning calibration machine learning algorithm application module, a TTS speech synthesis module, and a user data security analysis module;

[0007] The positioning acquisition and calibration module is used to obtain geographic environment and positioning data, and integrate and calibrate the geographic environment and positioning data, wherein the geographic environment and positioning data include user positioning data, POI data in the area, and vehicle operation data in the area;

[0008] The regional management and redline processing module is used to process the POI data in the region and perform regional redline processing;

[0009] The positioning calibration machine learning algorithm application module is used to analyze the training geographic environment and positioning data and output visualization results;

[0010] The TTS speech synthesis module is used to convert the geographical environment and positioning data into speech signals and broadcast them in real time;

[0011] The user data security analysis module is used to encrypt user location data.

[0012] Preferably, the positioning acquisition and calibration module includes a geographic environment and positioning data acquisition submodule, and the geographic environment and positioning data acquisition submodule includes a user positioning data acquisition unit, a POI data acquisition unit in the area, and a vehicle data acquisition unit in the area;

[0013] The user positioning data acquisition unit is used to acquire user positioning data and other basic information data. The user positioning data is acquired through the GPS positioning device at the user end, and the basic information data is acquired through the Beidou satellite data interface;

[0014] The POI data acquisition unit in the area is used to collect POI data in the relevant area, and the operating vehicle data acquisition unit in the area is used to acquire operating vehicle data in the area and analyze traffic operation conditions in the area.

[0015] Preferably, the positioning acquisition and calibration module further includes a data integration management unit and a data calibration unit;

[0016] The data integration management unit is used to integrate the collected geographic environment and positioning data, establish a data platform, and finally store the integrated geographic environment and positioning data in the data platform;

[0017] The data calibration unit is used to calibrate the user positioning data through machine learning and intersatellite link technology, and adopts the existing traffic road network model to offset the coordinates of the road network where the user positioning data is located, and offset it to the correct road network symmetric information coordinate range.

[0018] Preferably, the area management and redline processing module includes a redline preprocessing unit, an area boundary extraction unit, and a result visualization unit;

[0019] The redline preprocessing unit is used to preprocess the map vector obtained from the third-party data interface according to the OGC standard, and preprocess the map vector according to the GB / T17791-2007 Geographic Information Highway Network Data Exchange Format Standard Specification, and the preprocessing includes removing noise around the map vector elements and geographic information, geometric correction and radiation correction;

[0020] The region boundary extraction unit uses an image segmentation algorithm and a target extraction algorithm to extract the boundaries of buildings or scenic spots in the POI data in the region, distinguishes the specific boundaries of each region of the POI data in the region by identifying the boundary collision areas between different regions, and generates a POI region;

[0021] The result visualization unit is used to visualize the results of the POI area and the red line processing unit, generate a vector boundary map of the building or scenic area and refine it into a GIS service.

[0022] Preferably, the positioning calibration machine learning algorithm application module includes a data preprocessing unit, a feature extraction unit, and a model training unit;

[0023] The data preprocessing unit is used to clean and process the geographical environment and positioning data entered into the data platform;

[0024] The feature extraction unit screens features related to the POI area based on computer graphics knowledge to reduce feature area coverage or chaos;

[0025] The model training unit uses K-Means clustering training and an autoencoder model to extract features of user positioning data and POI areas, divides the original data into a training set and a validation set, and uses the training set to train the autoencoder model. The K-Means clustering training is used to determine the number K of clusters to be clustered, clusters the original data using the K-Means algorithm, assigns the original data points to the nearest cluster center, and repeats the iteration until the cluster center no longer changes significantly or a predetermined number of iterations is reached;

[0026] The autoencoder model includes an encoder and a decoder part, and the network structure of the autoencoder is designed, specifically: the number of nodes is estimated according to the maximum number of the input original data set, the activation function ReLU is used to act on the hidden layer, the sigmoid function is used to act on the output layer, and then the encoded dimension is determined; the training set is used to train the autoencoder so that the autoencoder can learn the effective representation of the data in the training set, the network weights are adjusted by minimizing the reconstruction error, and the autoencoder model parameters are adjusted by optimization algorithm and cross-validation method to achieve the best autoencoder model performance.

[0027] Preferably, the positioning calibration machine learning algorithm application module further includes a model verification and adjustment unit;

[0028] The model verification and adjustment unit includes a model evaluation subunit and a model optimization subunit;

[0029] The model evaluation subunit uses a cross-validation technique to first divide the data set into K subsets of equal size; each subset is used as a test set in turn, and the remaining K-1 subsets are used as evaluation training sets; then in each iteration, the evaluation training set is used to train the autoencoder model, and the performance of the autoencoder model is evaluated on the test set; the accuracy, precision, recall and F1 score of each iteration are calculated, and the F1 score is an indicator used to measure the accuracy of a binary classification model; finally, the results of all iterations are averaged to obtain the average accuracy, precision, recall and F1 score of the autoencoder model; by calculating the accuracy, precision, recall and F1 score indicators of the autoencoder model, the performance and generalization ability of the autoencoder model are evaluated;

[0030] The model optimization subunit adjusts the hyperparameters of the autoencoder model, improves feature engineering or adjusts the sample weight method according to the evaluation results, uses cross-validation, confusion matrix, ROC curve, precision, and recall rate indicators to evaluate the current performance of the autoencoder model, and then finds the optimal hyperparameter combination through random search; sets a noise threshold, continues to remove irrelevant features or features with noise greater than the noise threshold, creates new features, and then combines existing features or uses the positioning domain knowledge in the existing positioning model; uses feature selection technology to reduce dimensions and retain the most informative vector features. During this period, if the data set is unbalanced, uses oversampling technology, namely Oversampling, to balance the categories; uses sample weights to give more attention to minority classes, which is achieved by setting weight vector values, and then applies L1 regularization combined with Dropout technology to reduce overfitting, where L1 is the minimum absolute value deviation, and finally uses Stacking ensemble learning method to improve the robustness of the autoencoder model.

[0031] Preferably, the positioning calibration machine learning algorithm application module further includes a model analysis result and visualization unit, a model update and maintenance unit, and an encryption and decryption processing unit;

[0032] The model analysis result and visualization unit uses the auto-encoding model trained by the model training unit to perform regional red line analysis on the original data of the new date, and outputs the irregular regional red line preprocessing result;

[0033] The model updating and maintenance unit is used for appending original data and retraining the autoencoder model;

[0034] The encryption and decryption processing unit encrypts the collected geographic environment and positioning data according to unified industry standards and establishes a unified encryption and decryption process.

[0035] Preferably, the TTS speech synthesis module includes a speech synthesis unit and a transfer learning training unit;

[0036] The speech synthesis unit is used to learn the correspondence between speech and text, optimizes using the Deep Voice model, converts text into frequency spectrum, and then synthesizes speech through a neural network;

[0037] The transfer learning training unit is used to improve the performance of the Deep Voice model. It uses the knowledge of the already trained speech synthesis models, namely the Tacotron model and the WaveNet model, to improve the performance of the current task. The knowledge of the Tacotron model and the WaveNet model is transferred to the Deep Voice model to accelerate the training of the Deep Voice model and improve its performance.

[0038] Preferably, the speech synthesis unit specifically comprises the following steps:

[0039] S1. Increase the amount of training data based on the existing language data. The training data includes speech samples with different accents and recordings in different environments.

[0040] S2. Use recurrent neural network (RNN) to improve the performance and accuracy of the Deep Voice model.

[0041] S3. Introducing the attention mechanism into the Deep Voice model to improve the performance of the neural network;

[0042] S4. Use the Adam optimizer to update the Deep Voice model parameters, including:

[0043] S41. In each iteration, the mini-batch gradient descent method is used to calculate the gradient of the loss function with respect to the model parameters; according to the gradient algorithm:

[0044]

[0045] in represents the gradient change value, ω and b are the optimal solution parameters; Representation function The directional derivative along the x-axis at point x is Any point x of the set of functions on the x-axis, Represents the objective function The partial derivative of parameter b reflects the function The rate of change in the b direction; according to the gradient results, the parameters are updated:

[0046]

[0047]

[0048] Among them, α in the formula k is a learning rate that changes with the number of iterations k. Both ω and b can be determined by α k Calculated by the above formula; finally, when the gradient changes The absolute value approaches the threshold δ grad After that, the final result parameters after training are returned: ω and b, δ grad The calculation formula is as follows:

[0049]

[0050] The symbol (||…||) represents the norm of the vector, which is used to measure the size of the gradient; the colon is a mathematical symbol that represents a ratio or a proportion, which means the relative size between two numbers or two quantities; the lowercase i in the formula is an index, which represents the parameter of the i-th neuron or the i-th layer. It is a variable. If i=1, it means the gradient of the first neuron or the first layer. If i=2, it means the gradient of the second neuron or the second layer, and so on.

[0051] At this point, the values ​​of the optimal parameters ω and b can be obtained;

[0052] S42, the learning rate, beta1 and beta2 values ​​are adjusted according to the center amount of each returned result; specifically: first, for learning rate decay, the learning rate is gradually reduced as the training progresses, using exponential decay; at the same time, the adaptive learning rate of the Adam optimizer is increased, and the learning rate is automatically adjusted according to the historical information of the gradient; in the Adam optimizer, beta1 and beta2 are the decay rates of the unbiased variance that controls the first-order moment estimate, i.e. momentum, and the second-order moment estimate, i.e. gradient; beta1 is used to calculate the exponential moving average of the gradient; beta2 is used to calculate the exponential moving average of the square of the gradient; then, the learning rate scheduler is used to automatically adjust the learning rate according to the verification performance of the model during training; the ReduceLROnPlateau callback function is used to reduce the learning rate when the verification performance of the model stops improving; finally, Bayesian optimization is used to find the best hyperparameter combination;

[0053] S43. Repeat steps S41 and S42 until the preset number of iterations is reached or other stopping conditions are met; other stopping conditions include that the performance on the validation set no longer improves. If the performance of the model on the validation set does not improve significantly in several consecutive iterations, the model can be considered to have converged and training can be stopped. In addition, when the value of the loss function is lower than a preset threshold, the model can also be considered to have achieved sufficient accuracy.

[0054] Preferably, the user data security analysis module includes a data protection unit, a data backup and recovery unit, a data access control unit, a data audit and monitoring unit, a security incident response unit, a security policy management unit, a risk assessment and reporting unit, a data classification and marking unit, an anomaly detection and alarm unit, and an integration and linkage unit;

[0055] The data protection unit is used to encrypt and securely store the user location data to ensure the security and confidentiality of the user location data;

[0056] The data backup and recovery unit is used to regularly back up user location data and restore data when necessary to ensure the integrity and availability of user location data;

[0057] The data access control unit is used to perform access control on user location data, and only authorized users can access sensitive data;

[0058] The data audit and monitoring unit is used to monitor and audit user location data in real time, and promptly detect and handle abnormal operations and potential security threats;

[0059] The security incident response unit is used to respond immediately when a security incident is detected and take corresponding measures to prevent the further spread of the security incident;

[0060] The security policy management unit is used to manage and configure security policies to ensure that only users who comply with the security policies can access user location data;

[0061] The risk assessment and reporting unit is used to perform risk assessment on user location data, identify potential security risks, and generate corresponding risk reports;

[0062] The data classification and marking unit classifies and marks the user location data according to the importance and sensitivity of the user data, so as to better manage and protect the user location data;

[0063] The anomaly detection and alarm unit is used to monitor user location data in real time, and once abnormal behavior or potential threats are found, an alarm is immediately issued and corresponding measures are taken;

[0064] The integration and linkage unit is used to integrate and link with other units of the user data security analysis module, supervise the working conditions of other units of the user data security analysis module, and form a comprehensive security protection system.

[0065] The beneficial effects of the present invention are as follows: the positioning acquisition and calibration module is used to obtain geographical environment and positioning data, and integrate and calibrate the geographical environment and positioning data, obtain various types of data information, and improve the accuracy of the voice broadcast system; the regional management and red line processing module is used for POI data in the region and regional red line processing; the positioning calibration machine learning algorithm application module is used to analyze and train geographical environment and positioning data and output visualization results, which is convenient for accurately distinguishing different regions and roads, and further improving the accuracy of the broadcast; the TTS speech synthesis module is used to convert geographical environment and positioning data into voice signals and broadcast in real time, using the knowledge of Tacotron model and WaveNet model as the basis and introducing a weight mechanism, and introducing a neural network and an attention mechanism in the speech synthesis unit, which can better synthesize natural and fluent speech; the user data security analysis module is used to encrypt the user positioning data, which can ensure the security of the collected information and improve the privacy protection performance of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only preferred embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0067] Figure 1A schematic diagram of the functional modules of an artificial intelligence broadcasting system based on Beidou satellite positioning and TTS framework provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0068] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The enumerated embodiments are only used to explain the present invention and are not used to limit the scope of the present invention.

[0069] Reference Figure 1 This embodiment provides an artificial intelligence broadcasting system based on Beidou satellite positioning and TTS framework. The artificial intelligence broadcasting system based on Beidou satellite positioning and TTS framework includes a positioning acquisition and calibration module, an area management and red line processing module, a positioning calibration machine learning algorithm application module, a TTS speech synthesis module, and a user data security analysis module.

[0070] The positioning acquisition and calibration module is used to obtain geographic environment and positioning data, and integrate and calibrate the geographic environment and positioning data, wherein the geographic environment and positioning data include user positioning data, POI data in the area, and vehicle operation data in the area;

[0071] The regional management and redline processing module is used to process the POI data in the region and perform regional redline processing;

[0072] The positioning calibration machine learning algorithm application module is used to analyze the training geographic environment and positioning data and output visualization results;

[0073] The TTS speech synthesis module is used to convert the geographical environment and positioning data into speech signals and broadcast them in real time;

[0074] The user data security analysis module is used to encrypt user location data.

[0075] The positioning acquisition and calibration module includes a geographic environment and positioning data acquisition submodule, and the geographic environment and positioning data acquisition submodule includes a user positioning data acquisition unit, a POI data acquisition unit in the area, and a vehicle data acquisition unit in the area;

[0076] The user positioning data acquisition unit is used to acquire user positioning data and other basic information data. The user positioning data includes latitude and longitude information, timestamp, device information, and location information. The basic information data includes name, gender, and age. The user positioning data is acquired through the GPS positioning device at the user end, and the basic information data is acquired through the Beidou satellite data interface;

[0077] The POI data acquisition unit in the area is used to collect POI data in the relevant area. The POI data in the area includes name, category, coordinates, address, telephone number, business hours, and photos. The POI data in the area is acquired through a web crawler or an enterprise almanac data interface; it provides a basis for travel planning and early warning and forecast of dangerous goods vehicles. The operating vehicle data acquisition unit in the area is used to acquire operating vehicle data in the area and analyze traffic operation conditions in the area.

[0078] These data will provide the necessary information basis for the screening of travel broadcast range and trigger point information, comprehensively understand the traffic conditions in the user's location area, mark location information, and improve the depth and relevance of broadcast content to specific business entities.

[0079] The positioning acquisition and calibration module also includes a data integration management unit and a data calibration unit;

[0080] The data integration management unit integrates the collected geographic environment and positioning data according to the unified industry standard "GB / T 39616-2020", establishes a data platform, and finally stores the integrated geographic environment and positioning data in the data platform. The data platform includes a user information library, an operating vehicle library, and a regional location information library;

[0081] The data calibration unit is used to calibrate the user positioning data through machine learning and inter-satellite link technology, and adopts the existing traffic road network model to offset the coordinates of the road network at the location of the user positioning data to the correct road network symmetry information coordinate range; specifically: first determine the nearest target road network based on the user positioning data, then frame the range of the traffic road network model, calculate the offset and direction between the user positioning data and the range of the traffic road network model, and reference the calculated offset to the positioning coordinates to obtain new coordinates in the range of the traffic road network model, ensure that the new coordinates are aligned with the target road network, and then perform three-star positioning calibration on the coordinates according to the inter-satellite link, and confirm the precise coordinates of the coordinate center position through the three-star positioning deviation value, thereby providing support for subsequent triggering and broadcasting information.

[0082] The area management and redline processing module includes a redline preprocessing unit, an area boundary extraction unit, and a result visualization unit;

[0083] The red line preprocessing unit is used to preprocess the map vector obtained from the third-party data interface according to the OGC standard, and the elements and geographic information contained in the map vector are preprocessed according to the GB / T17791-2007 Geographic Information Highway Network Data Exchange Format Standard Specification. The preprocessing includes removing noise around the elements and geographic information in the map vector, geometric correction and radiation correction; these preprocessing operations can improve the quality of regional red line extraction, eliminate uncertain factors, and make subsequent analysis more accurate and reliable.

[0084] The region boundary extraction unit uses an image segmentation algorithm and a target extraction algorithm to extract the boundaries of buildings or scenic spots in the POI data in the region, distinguishes the specific boundaries of each region of the POI data in the region by identifying the boundary collision areas between different regions, and generates a POI region. By identifying the boundary collision areas between regions, the specific boundaries of each POI region can be better distinguished;

[0085] The result visualization unit is used to visualize the results of the POI area and the red line processing unit, generate a vector boundary map of the building or scenic area and refine it into a GIS service.

[0086] The regional management and redline processing module can comprehensively analyze the appropriate boundaries and redline boundaries of the POI area, provide accurate regional boundary analysis results and related information, and provide a scientific basis for triggering regional broadcasts at travel area points to reduce the risk of distant or random broadcasts.

[0087] The positioning calibration machine learning algorithm application module includes a data preprocessing unit, a feature extraction unit, and a model training unit;

[0088] The data preprocessing unit is used to clean and process the geographic environment and positioning data entered into the data platform. Specifically, before the geographic environment and positioning data are entered into the data platform, duplicate data are removed through data uniqueness verification in the programming method, and missing data are supplemented according to the date sequence. At the same time, outliers are avoided and processed according to the business scope, including date format problems, too large or too small data volume, and then the data format is uniformly converted to meet the input requirements of the machine learning algorithm for unified data format, corresponding data labels, same time series and same dimension;

[0089] The feature extraction unit screens features related to the POI area based on computer graphics knowledge to reduce feature area coverage or chaos, specifically: using statistical methods, signal processing technology and geographic information system tools to convert raw data into feature representations that can be used by machine learning algorithms, including extracting statistical features, frequency domain features or regional features, and reducing the dimension of high-dimensional features or combining features through principal component analysis, linear discriminant analysis, and feature embedding methods to reduce redundant information and enhance feature expression capabilities. The raw data is the joint data of POI data and traffic network data in the region, and is time-effective;

[0090] The model training unit uses K-Means clustering training and an autoencoder model to extract features of user positioning data and POI areas, divides the original data into a training set and a validation set, and uses the training set to train the autoencoder model. The K-Means clustering training is used to determine the number K of clusters to be clustered, clusters the original data using the K-Means algorithm, assigns the original data points to the nearest cluster center, and repeats the iteration until the cluster center no longer changes significantly or a predetermined number of iterations is reached;

[0091] The autoencoder model includes an encoder and a decoder part, and the network structure of the autoencoder is designed, specifically: the number of nodes is estimated according to the maximum number of the input original data set, the activation function ReLU is used to act on the hidden layer, the sigmoid function is used to act on the output layer, and then the encoded dimension is determined; the training set is used to train the autoencoder so that the autoencoder can learn the effective representation of the data in the training set, the network weights are adjusted by minimizing the reconstruction error, and the autoencoder model parameters are adjusted by optimization algorithm and cross-validation method to achieve the best autoencoder model performance.

[0092] The data preprocessing unit, feature extraction unit, and model training unit improve the accuracy and generalization ability of the autoencoder model;

[0093] The positioning calibration machine learning algorithm application module also includes a model verification and adjustment unit;

[0094] The model verification and adjustment unit includes a model evaluation subunit and a model optimization subunit;

[0095] The model evaluation subunit uses a cross-validation technique to first divide the data set into K subsets of equal size; each subset is used as a test set in turn, and the remaining K-1 subsets are used as evaluation training sets; then in each iteration, the evaluation training set is used to train the autoencoder model, and the performance of the autoencoder model is evaluated on the test set; the accuracy, precision, recall and F1 score of each iteration are calculated, and the F1 score is an indicator used to measure the accuracy of a binary classification model; finally, the results of all iterations are averaged to obtain the average accuracy, precision, recall and F1 score of the autoencoder model; by calculating the accuracy, precision, recall and F1 score indicators of the autoencoder model (which are inherent parts of machine learning), the performance and generalization ability of the autoencoder model are evaluated;

[0096] The model optimization subunit adjusts the hyperparameters of the autoencoder model, improves feature engineering or adjusts the sample weight method according to the evaluation results, uses cross-validation, confusion matrix, ROC curve, precision, and recall rate indicators to evaluate the current performance of the autoencoder model, and then finds the optimal hyperparameter combination through random search; sets a noise threshold, continues to remove irrelevant features or features with noise greater than the noise threshold, creates new features, and then combines existing features or uses the positioning domain knowledge in the existing positioning model; uses feature selection technology to reduce dimensions and retain the most informative vector features. During this period, if the data set is unbalanced, use oversampling technology, i.e., Oversampling. mpling to balance categories; use sample weights to give more attention to minority classes by setting weight vector values, generally quantified as 1.35 vector values, and then apply L1 regularization combined with Dropout technology to reduce overfitting, where L1 is the minimum absolute value deviation. Finally, use Stacking ensemble learning method to improve the robustness of the autoencoder model, record the configuration and results of each experiment to track progress and compare different methods; analyze the errors of the autoencoder model, understand on which types of samples the autoencoder model performs poorly, and use learning curves and validation curves to diagnose whether the autoencoder model has bias or variance problems.

[0097] The positioning calibration machine learning algorithm application module also includes a model analysis result and visualization unit, a model update and maintenance unit, and an encryption and decryption processing unit;

[0098] The model analysis result and visualization unit uses the autoencoder model trained by the model training unit to perform regional red line analysis on the original data of the new date, and outputs the preprocessing result of the red line of the irregular area; specifically: necessary cleaning and conversion are performed on the original data of the new date, including scaling the red line area to a uniform scale, normalizing the red line area label content, and encoding the red line key label weights, so that the format of the newly input red line area data is consistent with the original data format used by the model training unit; then the trained autoencoder model is loaded to ensure that the parameters and structure of the autoencoder model are consistent with those of the model training unit; the preprocessed original data of the new date is input into the autoencoder model for prediction; according to the output of the autoencoder model, the red line of the irregular area is identified; a clear red line boundary is obtained, and then contour detection and morphological operations in image processing technology are applied to optimize the representation result of the red line; finally, the processed red line result is output in an appropriate format and saved as an image, vector map or coordinate data set; the prediction result is visualized, including a raw POI area red line map, a POI area vector map or a map in a geographic information system; so that users can intuitively understand the spatial distribution and degree of the red lines in the POI area and its surrounding areas.

[0099] The model update and maintenance unit is used for raw data appending and auto-encoding model retraining, specifically: regularly adding raw data of new dates, and using incremental learning and online learning methods to update the model to adapt to changes in data distribution, regularly collecting raw data samples of new dates, including red line data of scenic spots and important points, ensuring that data samples can represent the current data distribution, pre-processing the newly collected red line areas, user location data, and traffic network model data to make them consistent with the data format used in auto-encoding model training, adding mini-batch learning, and using new data samples to incrementally update the auto-encoding model instead of starting from scratch. Start retraining the model, establish an online learning mechanism, join the online learning system, process real-time data and update the autoencoder model in real time. At the same time, regularly evaluate the performance of the model to ensure that incremental learning or online learning does not lead to a decline in the overall recognition performance of the autoencoder model. Use the original data of the new date for verification to monitor the accuracy and robustness of the autoencoder model. Finally, in incremental learning, introduce a forgetting mechanism to avoid the impact of outdated data on the model. Use weight decay to balance the impact of new and old data. This is achieved by adding a regularization term to the loss function, which is usually the L2 norm of the model weight. The L2 norm represents the minimum square error, and the formula is:

[0100]

[0101] Among them, L is the total loss, L0 is the original loss function, γ is the regularization coefficient, and w iis the model weight, which penalizes large weight values ​​to encourage the model to learn simpler patterns and reduce its dependence on old data;

[0102] The encryption and decryption processing unit encrypts the collected geographic environment and positioning data according to unified industry standards, establishes a unified encryption and decryption process, first generates a random key, and then uses the AES algorithm and CBC mode to encrypt in sequence; the encrypted geographic environment and positioning data are converted into Base64 format for easy transmission and storage; the decryption process is the opposite of the encryption process, using the same key and initialization vector IV to restore the geographic environment and positioning data before encryption; finally, records logs of all encryption and decryption operations for auditing and tracking; regularly checks logs to ensure the security of the encryption and decryption process; this can ensure the integrity, accuracy and security of user data, and provide convenient data support for the user data security analysis module.

[0103] The TTS speech synthesis module includes a speech synthesis unit and a transfer learning training unit;

[0104] The speech synthesis unit is used to learn the correspondence between speech and text, optimizes using the Deep Voice model, converts text into frequency spectrum, and then synthesizes speech through a neural network;

[0105] The transfer learning training unit is used to improve the performance of the Deep Voice model. It uses the knowledge of the already trained speech synthesis models, namely the Tacotron model and the WaveNet model, to improve the performance of the current task. The knowledge of the Tacotron model and the WaveNet model is transferred to the Deep Voice model to accelerate the training of the Deep Voice model and improve its performance. First, the weights of the pre-trained Deep Voice model are loaded, and then the weights of the pre-trained Deep Voice model are transferred to the Deep Voice model. In this process, it is necessary to ensure that the structures of the corresponding layers in the Tacotron model and the WaveNet model are compatible, including filtering out unmatched weights, updating the weights of the Deep Voice model at this time, and finally loading the updated weights so that the Deep Voice model already has some pre-trained knowledge and can continue to be trained or fine-tuned.

[0106] Preferably, the speech synthesis unit specifically comprises the following steps:

[0107] S1. Increase the amount of training data based on the existing language data. The training data includes speech samples with different accents, including Cantonese and Sichuanese, and recordings in different environments, including recordings in noisy backgrounds and different room acoustic conditions.

[0108] S2. Use recurrent neural network (RNN) to improve the performance and accuracy of the Deep Voice model. Since RNN has a memory function, it can take into account the input of the previous time step, so that it can better capture the long-term dependencies in the sequence and increase the model's ability to model the speech sequence, so that it can better synthesize natural and fluent speech.

[0109] S3. Introducing the attention mechanism into the Deep Voice model to improve the performance of the neural network, so that the model pays more attention to the important parts of the text and improves the accuracy and naturalness of the synthesis;

[0110] S4. Use the Adam optimizer to update the Deep Voice model parameters, including:

[0111] S41. In each iteration, the mini-batch gradient descent method is used to calculate the gradient of the loss function with respect to the model parameters; according to the gradient algorithm:

[0112]

[0113] in represents the gradient change value, ω and b are the optimal solution parameters, and the optimal solution parameters can be any value; Representation function The directional derivative along the x-axis at point x is Any point x of the set of functions on the x-axis, Represents the objective function The partial derivative of parameter b reflects the function The rate of change in the b direction; according to the gradient results, the parameters are updated:

[0114]

[0115]

[0116] Among them, α in the formula k is a learning rate that changes with the number of iterations k. It can be set according to different strategies, such as fixed, decreasing, and adaptive. This allows the step size to be dynamically adjusted during the optimization process to achieve better results. Both ω and b can be determined by α k Calculated by the above formula; finally, when the gradient changes The absolute value approaches the threshold δ grad After that, the final result parameters after training are returned: ω and b, δ grad The calculation formula is as follows:

[0117]

[0118] Among them, the symbol (||…||) represents the norm of the vector, which is used to measure the size of the gradient; the colon is a mathematical symbol that represents a ratio or proportion, which means the relative size between two numbers or two quantities; the lowercase i in the formula is an index, which represents the parameter of the i-th neuron or the i-th layer. It is a variable that can take different values ​​such as 1, 2, 3,…. Its function is to allow us to use a general formula to represent the gradients of different neurons or different layers. For example, It represents the weight of the i-th neuron or i-th layer (w i ) and bias (b i ), if i = 1, it means the gradient of the first neuron or the first layer, if i = 2, it means the gradient of the second neuron or the second layer, and so on;

[0119] At this point, we can get the optimal parameters ω and b; α in the formula k Is a parameter representing the learning rate, which determines the step size at each iteration in the optimization algorithm. The learning rate is an important hyperparameter that affects the convergence speed and effect of the loss function. Generally speaking, the larger the learning rate, the larger the step size, the faster the convergence, but it may also cause excessive jumps or oscillations. The smaller the learning rate, the smaller the step size, the slower the convergence, but it may also cause falling into a local optimum or stagnation. Therefore, choosing an appropriate learning rate is a key issue, and sometimes it needs to be adjusted according to actual conditions.

[0120] S42, the learning rate, beta1 and beta2 values ​​are adjusted according to the centering amount of each returned result; the learning rate, beta1 and beta2 are important components of the Adam optimizer, and they have a significant impact on the training process and final performance of the model; specifically: first, with respect to learning rate decay, as the training progresses, the learning rate is gradually reduced, and exponential decay is used; at the same time, the adaptive learning rate of the Adam optimizer is improved, and the learning rate is automatically adjusted according to the historical information of the gradient; in the Adam optimizer, beta1 and beta2 are the decay rates that control the first-order moment estimate, i.e. momentum, and the second-order moment estimate, i.e. the unbiased variance of the gradient, respectively; beta1 is usually set to a value close to 1, such as 0.9, which is used to calculate the exponential moving average of the gradient, thereby smoothing the gradient and providing momentum; beta2 is also usually set to a value close to 1, such as 0.999, which is used to calculate the exponential moving average of the square of the gradient, thereby stabilizing the variance of the gradient; then, the learning rate scheduler Learning Rate Scheduler, which automatically adjusts the learning rate during training based on the validation performance of the model; uses the ReduceLROnPlateau callback function to reduce the learning rate when the validation performance of the model stops improving; and finally, uses Bayesian optimization to find the best hyperparameter combination;

[0121] S43, repeating steps S41 and S42 until the preset number of iterations is reached or other stopping conditions are met. The preset number of iterations is used to ensure that the model has enough time to learn data features and optimize parameters. Usually, the number of iterations is determined according to the size of the data set, the complexity of the model and the training goal. Too few preset iterations may result in insufficient model training, while too many preset iterations may result in overfitting or waste of computing resources. Currently, according to the size of the training data set, the number of iterations is set to 500, which is only an intermediate value. Other stopping conditions include that the performance on the validation set is no longer improved. If the performance of the model on the validation set is not significantly improved in several consecutive iterations, it can be considered that the model has converged and training can be stopped. In addition, when the value of the loss function is lower than the preset threshold, It can also be considered that the model has achieved sufficient accuracy. These conditions are to ensure that the model can achieve optimal performance without overfitting, so that the TTS speech synthesis module can convert the input text into natural language speech output, support multiple languages ​​and timbre selections, and customize parameters such as volume and speaking speed. Furthermore, by processing the user's location information, the user is reminded to enter important area POIs or surrounding scenic spots based on the regional red lines. At the same time, the collected POI information and relevant data of the scenic spot, including attraction introductions, ticket information, travel routes, catering and accommodation, etc., are sent to the TTS module after preprocessing. After receiving the data, the TTS language module will use speech synthesis technology to convert the text information into speech signals and generate voice files or real-time voice broadcasts.

[0122] The user data security analysis module includes a data protection unit, a data backup and recovery unit, a data access control unit, a data audit and monitoring unit, a security incident response unit, a security policy management unit, a risk assessment and reporting unit, a data classification and marking unit, an anomaly detection and alarm unit, and an integration and linkage unit;

[0123] The user data security analysis module can effectively prevent the loss and leakage of user privacy data, help ensure the security and confidentiality of user data, protect the core assets of the enterprise, and improve the overall information security level.

[0124] The data protection unit is used to encrypt and securely store the user location data to ensure the security and confidentiality of the user location data;

[0125] The data backup and recovery unit is used to regularly back up user location data and restore data when necessary to ensure the integrity and availability of user location data;

[0126] The data access control unit is used to perform access control on user location data, and only authorized users can access sensitive data;

[0127] The data audit and monitoring unit is used to monitor and audit user location data in real time, and promptly detect and handle abnormal operations and potential security threats;

[0128] The security incident response unit is used to respond immediately when a security incident is detected and take corresponding measures to prevent the further spread of the security incident;

[0129] The security policy management unit is used to manage and configure security policies to ensure that only users who comply with the security policies can access user location data;

[0130] The risk assessment and reporting unit is used to perform risk assessment on user location data, identify potential security risks, and generate corresponding risk reports;

[0131] The data classification and marking unit classifies and marks the user location data according to the importance and sensitivity of the user data, so as to better manage and protect the user location data;

[0132] The anomaly detection and alarm unit is used to monitor user location data in real time, and once abnormal behavior or potential threats are found, an alarm is immediately issued and corresponding measures are taken;

[0133] The integration and linkage unit is used to integrate and link with other units of the user data security analysis module, supervise the working conditions of other units of the user data security analysis module, and form a comprehensive security protection system.

[0134] The user data security analysis module is an important part of the intelligent broadcasting technology. It ensures the privacy of user data by encrypting the user location data. At the same time, decryption calculation is required to effectively avoid information security and privacy protection issues.

[0135] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. The protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. Artificial intelligence broadcasting system based on Beidou satellite positioning and TTS framework, It is characterized in that The system includes a positioning acquisition and calibration module, an area management and redline processing module, a positioning calibration machine learning algorithm application module, a TTS speech synthesis module, and a user data security analysis module; The positioning acquisition and calibration module is used to obtain geographic environment and positioning data, and integrate and calibrate the geographic environment and positioning data, wherein the geographic environment and positioning data include user positioning data, POI data in the area, and vehicle operation data in the area; The regional management and redline processing module is used to perform regional redline processing on POI data in the region; The positioning calibration machine learning algorithm application module is used to analyze the training geographic environment and positioning data and output visualization results; The TTS speech synthesis module is used to convert the geographical environment and positioning data into speech signals and broadcast them in real time; The user data security analysis module is used to encrypt user location data.

2. According to the artificial intelligence broadcasting system based on Beidou satellite positioning and TTS framework according to claim 1, It is characterized in that The positioning acquisition and calibration module includes a geographic environment and positioning data acquisition submodule, and the geographic environment and positioning data acquisition submodule includes a user positioning data acquisition unit, a POI data acquisition unit in the area, and a vehicle data acquisition unit in the area; The user positioning data acquisition unit is used to acquire user positioning data and other basic information data. The user positioning data is acquired through the GPS positioning device at the user end, and the basic information data is acquired through the Beidou satellite data interface; The POI data acquisition unit in the area is used to collect POI data in the relevant area, and the operating vehicle data acquisition unit in the area is used to acquire operating vehicle data in the area and analyze traffic operation conditions in the area.

3. According to the artificial intelligence broadcasting system based on Beidou satellite positioning and TTS framework according to claim 2, It is characterized in that The positioning acquisition and calibration module also includes a data integration management unit and a data calibration unit; The data integration management unit is used to integrate the collected geographic environment and positioning data, establish a data platform, and store the integrated geographic environment and positioning data in the data platform; The data calibration unit is used to calibrate the user positioning data through machine learning and intersatellite link technology, and adopts a traffic road network model to offset the coordinates of the road network where the user positioning data is located, and offset it to the correct road network symmetric information coordinate range.

4. According to claim 1, the artificial intelligence broadcasting system based on Beidou satellite positioning and TTS framework, It is characterized in that The area management and redline processing module includes a redline preprocessing unit, an area boundary extraction unit and a result visualization unit; The redline preprocessing unit is used to preprocess the map vector obtained from the third-party data interface according to the OGC standard, and preprocess the map vector according to the GB / T17791-2007 Geographic Information Highway Network Data Exchange Format Standard Specification, and the preprocessing includes removing noise around the map vector elements and geographic information, geometric correction and radiation correction; The region boundary extraction unit uses an image segmentation algorithm and a target extraction algorithm to extract the boundaries of buildings or scenic spots in the POI data in the region, distinguishes the specific boundaries of each region of the POI data in the region by identifying the boundary collision area between different regions, and generates a POI region; The result visualization unit is used to visualize the results of the POI area and the red line processing unit, generate a vector boundary map of the building or scenic area and refine it into a GIS service.

5. According to claim 1, the artificial intelligence broadcasting system based on Beidou satellite positioning and TTS framework, It is characterized in that The positioning calibration machine learning algorithm application module includes a data preprocessing unit, a feature extraction unit, and a model training unit; The data preprocessing unit is used to clean and process the geographical environment and positioning data entered into the data platform; The feature extraction unit is used to filter features related to the POI area based on computer graphics knowledge; The model training unit is used to extract features of user positioning data and POI areas by using K-Means clustering training and an autoencoder model, divide the original data into a training set and a validation set, and use the training set to train the autoencoder model. The K-Means clustering training is used to determine the number K of clusters to be clustered, cluster the original data using the K-Means algorithm, assign the original data points to the nearest cluster center, and repeat iterations until the cluster center no longer changes significantly or a predetermined number of iterations is reached; The autoencoder model includes an encoder and a decoder part. The network structure of the autoencoder is designed, specifically: the number of nodes is estimated according to the maximum number of the input original data set, the activation function ReLU is used to act on the hidden layer, the sigmoid function is used to act on the output layer, and then the encoded dimension is determined; Use the training set to train the autoencoder so that the autoencoder can learn the effective representation of the data in the training set, adjust the network weights by minimizing the reconstruction error, and then adjust the autoencoder model parameters through optimization algorithms and cross-validation methods to achieve the best autoencoder model performance.

6. According to claim 5, the artificial intelligence broadcasting system based on Beidou satellite positioning and TTS framework, It is characterized in that The positioning calibration machine learning algorithm application module also includes a model verification and adjustment unit; The model verification and adjustment unit includes a model evaluation subunit and a model optimization subunit; The model evaluation subunit uses a cross-validation technique to first divide the data set into K subsets of equal size; each subset is used as a test set in turn, and the remaining K-1 subsets are used as evaluation training sets; then in each iteration, the evaluation training set is used to train the autoencoder model, and the performance of the autoencoder model is evaluated on the test set; the accuracy, precision, recall and F1 score of each iteration are calculated, and the F1 score is an indicator used to measure the accuracy of a binary classification model; finally, the results of all iterations are averaged to obtain the average accuracy, precision, recall and F1 score of the autoencoder model; by calculating the accuracy, precision, recall and F1 score indicators of the autoencoder model, the performance and generalization ability of the autoencoder model are evaluated; The model optimization subunit adjusts the hyperparameters of the autoencoder model, improves feature engineering or adjusts the sample weight method according to the evaluation results, uses cross-validation, confusion matrix, ROC curve, precision, and recall rate indicators to evaluate the current performance of the autoencoder model, and then finds the optimal hyperparameter combination through random search; Set a noise threshold, continue to remove irrelevant features or features with noise greater than the noise threshold, create new features, and then combine existing features or use the positioning domain knowledge in the existing positioning model; Use feature selection techniques to reduce dimensions and retain the most informative vector features. If the dataset is unbalanced, use oversampling to balance the categories. Sample weights are used to give more importance to minority classes by setting weight vector values, and then L1 regularization combined with Dropout technology is applied to reduce overfitting, where L1 is the minimum absolute deviation. Finally, Stacking ensemble learning method is used to improve the robustness of the autoencoder model.

7. According to claim 5, the artificial intelligence broadcasting system based on Beidou satellite positioning and TTS framework, It is characterized in that The positioning calibration machine learning algorithm application module also includes a model analysis result and visualization unit, a model update and maintenance unit, and an encryption and decryption processing unit; The model analysis result and visualization unit uses the auto-encoding model trained by the model training unit to perform regional red line analysis on the original data of the new date, and outputs the irregular regional red line preprocessing result; The model updating and maintenance unit is used for appending original data and retraining the autoencoder model; The encryption and decryption processing unit encrypts the collected geographic environment and positioning data according to unified industry standards and establishes a unified encryption and decryption process.

8. The artificial intelligence broadcasting system based on Beidou satellite positioning and TTS framework according to claim 1, It is characterized in that The TTS speech synthesis module includes a speech synthesis unit and a transfer learning training unit; The speech synthesis unit is used to learn the correspondence between speech and text, optimizes using the Deep Voice model, converts text into frequency spectrum, and then synthesizes speech through a neural network; The transfer learning training unit is used to improve the performance of the Deep Voice model. It uses the knowledge of the already trained speech synthesis models, namely the Tacotron model and the WaveNet model, to improve the performance of the current task. The knowledge of the Tacotron model and the WaveNet model is transferred to the Deep Voice model to accelerate the training of the Deep Voice model and improve its performance.

9. According to claim 8, the artificial intelligence broadcasting system based on Beidou satellite positioning and TTS framework, It is characterized in that The speech synthesis unit specifically comprises the following steps: S1. Increase the amount of training data based on the existing language data. The training data includes speech samples with different accents and recordings in different environments. S2. Use recurrent neural network (RNN) to improve the performance and accuracy of the Deep Voice model. S3. Introducing the attention mechanism into the Deep Voice model to improve the performance of the neural network; S4. Use the Adam optimizer to update the Deep Voice model parameters, including: S41. In each iteration, the mini-batch gradient descent method is used to calculate the gradient of the loss function with respect to the model parameters. According to the gradient algorithm, we can get: in Indicates the gradient change value. and is the optimal solution parameter; Representation function The directional derivative along the x-axis at point x is Any point x of the set of functions on the x-axis, Represents the objective function Parameters b The partial derivative of the function exist b The rate of change in direction; according to the gradient results, the parameters are updated: Among them, the formula is a learning rate that changes with the number of iterations k, and b Can be Calculated by the above formula; finally, when the gradient changes The absolute value approaches the threshold After that, the final result parameters after training are returned: and b, The calculation formula is as follows: Among them, the symbol (||…||) represents the norm of the vector, which is used to measure the size of the gradient; : is a mathematical symbol that represents a ratio or proportion, which means the relative size between two numbers or two quantities; i is an index that represents the parameter of the ith neuron or the ith layer, which is a variable. If i = 1, it means the gradient of the 1st neuron or the 1st layer; if i = 2, it means the gradient of the 2nd neuron or the 2nd layer, and so on; At this point, the optimal parameters can be obtained and b The value of S42, the learning rate, beta1 and beta2 values ​​are adjusted according to the center amount of each returned result; specifically: first, for learning rate decay, the learning rate is gradually reduced as the training progresses, using exponential decay; at the same time, the adaptive learning rate of the Adam optimizer is increased, and the learning rate is automatically adjusted according to the historical information of the gradient; in the Adam optimizer, beta1 and beta2 are the decay rates of the unbiased variance that controls the first-order moment estimate, i.e. momentum, and the second-order moment estimate, i.e. gradient; beta1 is used to calculate the exponential moving average of the gradient; beta2 is used to calculate the exponential moving average of the square of the gradient; then, the learning rate scheduler is used to automatically adjust the learning rate according to the verification performance of the model during training; the ReduceLROnPlateau callback function is used to reduce the learning rate when the verification performance of the model stops improving; finally, Bayesian optimization is used to find the best hyperparameter combination; S43. Repeat steps S41 and S42 until the preset number of iterations is reached or other stopping conditions are met; other stopping conditions include that the performance on the validation set no longer improves. If the performance of the model on the validation set does not improve significantly in several consecutive iterations, the model is considered to have converged and training is stopped. In addition, when the value of the loss function is lower than a preset threshold, the model can also be considered to have achieved sufficient accuracy.

10. The artificial intelligence broadcasting system based on Beidou satellite positioning and TTS framework according to claim 1, It is characterized in that The user data security analysis module includes a data protection unit, a data backup and recovery unit, a data access control unit, a data audit and monitoring unit, a security incident response unit, a security policy management unit, a risk assessment and reporting unit, a data classification and marking unit, an anomaly detection and alarm unit, and an integration and linkage unit; The data protection unit is used to encrypt and securely store user location data; The data backup and recovery unit is used to regularly back up user location data and restore data when necessary; The data access control unit is used to perform access control on user location data, and only authorized users can access sensitive data; The data audit and monitoring unit is used to monitor and audit user location data in real time, and promptly detect and handle abnormal operations and potential security threats; The security incident response unit is used to respond immediately when a security incident is detected and take corresponding measures to prevent the further spread of the security incident; The security policy management unit is used to manage and configure security policies to ensure that only users who comply with the security policies can access user location data; The risk assessment and reporting unit is used to perform risk assessment on user location data, identify potential security risks, and generate corresponding risk reports; The data classification and labeling unit classifies and labels the user location data according to the importance and sensitivity of the user data; The anomaly detection and alarm unit is used to monitor user location data in real time, and once abnormal behavior or potential threats are found, an alarm is immediately issued and corresponding measures are taken; The integration and linkage unit is used to integrate and link with other units of the user data security analysis module, supervise the working conditions of other units of the user data security analysis module, and form a comprehensive security protection system.

Images