Real-time processing method of automobile data based on artificial intelligence

Abstract

The present invention discloses a real-time processing method for automobile data based on artificial intelligence, and specifically relates to the technical field of data processing. By adopting an improved Transformer model, combined with a self-attention mechanism and cross-modal feature fusion technology, cross-context sentiment association analysis is realized, and the recognition accuracy of users' real emotional tendencies is improved. Based on accurate sentiment recognition results, the system predicts sentiment trends through a time series analysis model, and automatically sends an early warning to the enterprise management end when an abnormal public opinion event occurs. By using knowledge graphs and reinforcement learning, combined with historical data and user sentiment trends, the optimal public opinion response strategy is automatically generated, and a visual analysis report is used to assist enterprises in making accurate market decisions. The present invention realizes deep cross-modal sentiment understanding, real-time public opinion trend prediction, intelligent early warning and accurate decision-making optimization, and effectively improves the public opinion monitoring, brand management and market competitiveness of enterprises in the automotive industry.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a real-time processing method for automobile data based on artificial intelligence. Background Art

[0002] In recent years, with the rapid development of the automobile industry and the popularization of Internet information dissemination, the market dynamics, user feedback and brand public opinion information of the automobile industry have shown exponential growth. Social media, news media, short video platforms, forums and blogs have become important platforms for consumers to express their opinions, share experiences and discuss hot issues. However, due to the fragmentation and diversification of information and the huge amount of data, traditional manual monitoring and analysis methods can no longer meet the needs of enterprises for market trend identification, consumer sentiment analysis and competitor dynamic tracking. Therefore, the big data insight system for the automobile industry based on artificial intelligence came into being. With natural language processing (NLP), deep learning and big data analysis technology as the core, it realizes efficient collection, accurate analysis and intelligent early warning of public opinion data in the automobile industry, thereby providing scientific decision-making support for enterprises.

[0003] The prior art has the following deficiencies: In the process of big data analysis of public opinion in the automotive industry, the problem of cross-context emotion recognition of multimodal data is particularly serious and there are few effective solutions. Since automotive public opinion involves multimodal data such as text (news, social media comments), voice (user feedback), images (accident photos, advertisements), and videos (evaluations, live broadcasts), data of different modalities have huge differences in context understanding, and emotional expression methods vary. For example, when a user posts a post containing text and pictures on social media, the text may appear to be a positive comment when analyzed alone, but when combined with pictures (such as photos of vehicle damage), the actual emotion may be negative. However, existing natural language processing (NLP) and computer vision (CV) technologies are difficult to accurately integrate these heterogeneous data, making it impossible for the system to accurately identify real user emotions, resulting in distorted public opinion analysis results, which in turn affects the company's brand management and market strategy formulation. Summary of the invention

[0004] The purpose of the present invention is to provide a real-time processing method for automobile data based on artificial intelligence to solve the shortcomings of the background technology.

[0005] In order to achieve the above object, the present invention provides the following technical solution: a real-time processing method of automobile data based on artificial intelligence, comprising the following steps: S1: Real-time acquisition of automotive-related multimodal data from multiple data sources, including text, voice, images, and videos; S2: Extract features from the collected multimodal data, including word segmentation, denoising and sentiment word recognition for text data, speech recognition and conversion of speech data into text, and object detection and sentiment feature extraction for image and video data; S3: Based on the improved Transformer deep learning model, combined with the self-attention mechanism and multimodal feature fusion technology, cross-context correlation analysis is performed on different features extracted from multimodal data, and the accuracy of identifying the user's true emotional tendency is evaluated based on the analysis results; S4: For the accurate identification of the user's true emotional tendency, a real-time emotional trend prediction mechanism is built through a time series analysis model, and when an abnormal event is detected, an early warning signal is automatically sent to the preset enterprise management end; S5: Utilize knowledge graphs and reinforcement learning technology, combined with historical data and the accuracy of identifying users’ real emotional tendencies, to generate response strategies for public opinion events and provide visual analysis reports to assist companies in formulating accurate market response plans.

[0006] Preferably, the S1 multimodal data collection adopts a combination of Web crawlers, API interface calls and streaming data collection to obtain real-time and comprehensive automotive industry-related data from social media, news platforms, video platforms, user forums, e-commerce platforms and government regulatory agencies.

[0007] Preferably, the S2 multimodal data feature extraction includes joint processing of text, speech, image and video data, wherein: text data is processed using BERT-WordPiece word segmentation, denoising and sentiment dictionary matching; speech data is processed for sentiment recognition using Wav2Vec2.0 and CNN+LSTM models; image data is processed for target detection and OCR recognition using YOLO, Faster R-CNN and ResNet; video data is processed for key frame extraction and semantic sentiment analysis using TimeSformer and CNN-LSTM.

[0008] Preferably, the S3 is based on cross-context sentiment analysis of the improved Transformer model, extracts correlation features among text, speech, images and videos through a self-attention mechanism, and maps data of different modalities to a unified sentiment representation space through modal alignment technology to improve the accuracy of cross-modal sentiment recognition.

[0009] Preferably, the sentiment drift index is used to calculate the sentiment change trend of users in different time periods, and the context contradiction degree is used to evaluate the sentiment consistency of multimodal data.

[0010] Preferably, the calculation method of the sentiment drift index is: collect the user's comment data on the same brand / product at multiple time points, use the BERT+LSTM time series model to calculate the sentiment score at each time point , the expression is: ;in, is the user text at time t, f is the output function of the LSTM model; the sentiment change rate at adjacent time points is calculated as the sentiment drift index: ; where Δt is the time interval, is the sentiment drift index; The calculation method of contextual contradiction degree CDI is as follows: perform emotion recognition on multimodal data, calculate text emotion ST, voice emotion SV, image emotion SI, and video emotion SF respectively, and the score range of each emotion is [-1,1]; calculate emotion consistency EC, the expression is: ; Calculate the contextual contradiction degree CDI, the expression is: .

[0011] Preferably, the Transformer encoding layer is used to jointly encode text, speech, image, and video features; the modality alignment technology is used to map different modal data to the same semantic space; and the self-attention mechanism is used to calculate the weights of different modal information. , calculation formula: ; where Q, K, and V are query, key, and value matrices, respectively. is the dimension of the key, T is the matrix transpose; Measuring the changing relationship between SDC and CDI: If Corr(SDC, CDI)>0.5, it means that the emotion drift is large, and the emotion contradiction between the modes is serious, and there are implicit negative emotions or sarcastic tones; if Corr(SDC, CDI)<-0.5, it means that the emotion changes are stable, the emotion expression is consistent, and the user's evaluation is credible.

[0012] Preferably, the matching degree between the emotion labels predicted by the calculation model and the manually annotated emotion labels is calculated, that is, the emotion recognition accuracy SPA is calculated: ; Where Predicted Sentiment is the sentiment prediction value of the model for the sample, ranging from [-1, 1]; Ground Truth is the true sentiment label manually annotated, ranging from [-1, 1], N is the total number of samples, if SPA is lower than the threshold, the model is optimized; Adopt reinforcement learning training strategy and optimize emotion recognition model through reward mechanism: ; Use policy gradient optimization methods to adjust the parameters of the Transformer model to improve the accuracy of future predictions.

[0013] Preferably, the S4 time-series sentiment trend prediction mechanism is based on an improved Transformer structure, combined with a long short-term memory network, to construct a real-time public opinion change prediction model, and automatically sends a warning signal to the enterprise management end when it detects that the abnormal sentiment fluctuation value exceeds the set threshold.

[0014] In the above technical solution, the technical effects and advantages provided by the present invention are: 1. Through multimodal data fusion, improved Transformer deep learning model, self-attention mechanism, time series analysis and reinforcement learning technology, the present invention can collect automotive industry-related data from social media, news, video, voice comments and other channels in real time, and perform joint feature extraction on text, voice, image and video data to accurately identify the user's true emotional tendency. At the same time, through the calculation of the sentiment drift index (SDC) and contextual contradiction (CDI), the system can analyze the trend of emotional changes across modalities, detect potential negative emotions or sarcastic information, and avoid the misjudgment problem of single modal analysis.

[0015] 2. The present invention also predicts the development trend of public opinion through a time series analysis model (TimeSformer + LSTM), and automatically sends an early warning signal to the enterprise management end when abnormal events (such as a surge in negative emotions, breaking news) are detected to ensure that the enterprise can respond quickly. In addition, the knowledge graph and reinforcement learning technology are used to combine historical data and user emotion change patterns to intelligently generate the optimal public opinion response strategy, and visual analysis reports are used to assist enterprises in making accurate market decisions. The present invention not only improves the accuracy of cross-modal emotion recognition, but also optimizes the company's brand management, market public relations and risk control, and improves the company's public opinion management capabilities and user trust in the highly competitive automotive market. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0017] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] For examples, see Figure 1 As shown, the real-time processing method of automobile data based on artificial intelligence in this embodiment includes the following steps: S1: Real-time acquisition of automotive-related multimodal data from multiple data sources, including text, voice, images, and videos; S2: Extract features from the collected multimodal data, including word segmentation, denoising and sentiment word recognition for text data, speech recognition and conversion of speech data into text, and object detection and sentiment feature extraction for image and video data; S3: Based on the improved Transformer deep learning model, combined with the self-attention mechanism and multimodal feature fusion technology, cross-context correlation analysis is performed on different features extracted from multimodal data, and the accuracy of identifying the user's true emotional tendency is evaluated based on the analysis results; S4: For the accurate identification of the user's true emotional tendency, a real-time emotional trend prediction mechanism is built through a time series analysis model, and when an abnormal event is detected, an early warning signal is automatically sent to the preset enterprise management end; S5: Utilize knowledge graphs and reinforcement learning technology, combined with historical data and the accuracy of identifying users’ real emotional tendencies, to generate response strategies for public opinion events and provide visual analysis reports to assist companies in formulating accurate market response plans.

[0020] In S1, based on multiple heterogeneous data sources, the collected information is ensured to be comprehensive and accurate, covering the following types: Social media platforms (such as Weibo, WeChat, TikTok, and Kuaishou): Obtain car-related comments, short videos, live broadcast content, voice discussions, etc. posted by users.

[0021] News media (such as automobile portals, industry news websites): collect information about automobile market trends, industry policies, brand evaluations, recalls, etc.

[0022] Video platforms (such as YouTube and Bilibili): obtain video content such as car reviews, user sharing, and fault feedback, and analyze them in combination with automatic subtitle / voice analysis.

[0023] User forums and communities (such as Dongchedi, Autohome, Reddit): capture users’ long-form car-buying experiences, problem consultations, discussion posts, and other text content.

[0024] E-commerce and after-sales platforms (such as Tmall, JD.com, and car service apps): obtain users’ car purchase evaluations, repair feedback, and parts quality comments, etc.

[0025] Government and industry regulatory agencies (such as the State Administration for Market Regulation and the Standardization Administration of China): obtain data such as policies and regulations, vehicle recall announcements, and safety inspection reports.

[0026] In order to collect the above-mentioned multimodal data in real time and efficiently, the present invention adopts the following technical means: Web crawler technology: using crawler programs to regularly or in real time capture news, forums, social media and other web page data. API interface call: obtain real-time content through the official API interface provided by social media and video platforms, such as Weibo API, WeChat applet API, YouTubeData API, etc. Streaming data collection: using streaming data processing frameworks such as Kafka and Flink to realize the continuous collection of real-time published content, such as Weibo hot searches, Douyin live broadcast comments, etc. Speech to text (ASR, Automatic Speech Recognition): automatic speech recognition of voice content in audio or video, and transcription of speech into text for further sentiment analysis. Image and video analysis (OCR, target detection, sentiment analysis): perform target recognition (such as brand logo, vehicle model recognition), OCR text extraction, video summary analysis, etc. on the acquired pictures and video frames.

[0027] The collected multimodal data will be pre-processed through the ETL (Extract, Transform, Load) process and stored in a distributed database (such as Elasticsearch, MongoDB) or a big data platform (such as Hadoop, Spark) for subsequent deep learning modeling and public opinion analysis.

[0028] Since multimodal data may contain a large amount of redundant or irrelevant information, the present invention ensures the accuracy and validity of the data through technologies such as text deduplication algorithms (such as SimHash), spam filtering (such as blacklists, stop words), and format standardization (such as timestamp conversion).

[0029] S2: Perform feature extraction on the collected multimodal data, including word segmentation, denoising and sentiment word recognition processing on text data, speech recognition and conversion of speech data into text, and target detection and sentiment feature extraction on image and video data.

[0030] Text data mainly comes from social media, news websites, user comments, forum posts, policy announcements, etc. The processing process includes: Since Chinese text has no natural word boundaries, a word segmentation algorithm is needed to split sentences into independent words to improve the model's understanding ability: rule-based word segmentation (such as regular expression matching), statistical word segmentation (such as TF-IDF weight calculation), deep learning word segmentation (such as BERT-WordPiece, LSTM word segmenter). Example: Original text: "Tesla Model 3 recently adjusted its price, with a large price reduction." Word segmentation result: ["Tesla", "Model 3", "recently", "price", "adjustment", "price reduction", "range", "large"] In order to improve data quality, invalid information needs to be removed: HTML tags, special symbols, advertising content (such as , @User, #Topic#), stop words (such as meaningless words like "de", "le", "ne", etc.), repeated character detection (e.g., "aaaaaa" → "a").

[0031] Use the method of combining sentiment dictionary and sentiment analysis model to identify positive, negative or neutral sentiment in the text: keyword matching (such as "positive review", "fault", "complaint"), dependency syntactic analysis (e.g., "vehicle + [has a fault]" indicates negative sentiment), and combine with the deep learning BERT sentiment analysis model for sentiment classification.

[0032] The voice data mainly comes from user voice comments, telephone customer service recordings, video and audio, etc. The processing process includes: using automatic speech recognition (ASR) technology to convert speech into text. Common methods include: traditional HMM-GMM (Hidden Markov Model + Gaussian Mixture Model); modern deep learning models (such as DeepSpeech, Wav2Vec2.0). Example: User voice: "This car is really too fuel-consuming. It's simply unaffordable to drive!" Converted text: "This car is really too fuel-consuming. It's simply unaffordable to drive!" Analyze the emotional characteristics of the voice (such as anger, satisfaction, doubt). Methods include: Mel Frequency Cepstral Coefficients (MFCC) to extract emotional features; pitch, volume, and speech rate analysis (e.g., high pitch and fast speech rate when angry); neural network sentiment classification (such as CNN + LSTM); finally, the emotional tendency (positive / negative / neutral) and emotion category (such as angry / happy / anxious) can be obtained.

[0033] The image data mainly comes from social media pictures, news photos, accident photos, vehicle advertisements, etc. The processing process includes: used to identify the car brand, model, license plate, damaged parts, etc. in the picture. Methods include: CNN (Convolutional Neural Network) + Faster R-CNN; YOLO (You Only Look Once) series; lightweight models such as ResNet, MobileNet. Example: Input picture → Detection result (identify "Tesla Model 3") Identify the text information in the image, such as license plate number, announcement, user evaluation screenshot, etc. Adopt: Tesseract OCR (open source), PaddleOCR (high precision), Google Vision OCR API; Example: Input: A screenshot of a recall announcement, OCR result: Identify "Due to a battery management system fault, this batch of vehicles will be recalled for repair." Analyze the main body of the image (such as accident scene, product promotion picture), color emotion analysis (e.g., red tends to be warning, blue tends to be safe), deep learning expression recognition (such as whether the user shows anger / satisfaction in the picture).

[0034] Video data mainly comes from car reviews, user experience sharing, accident recordings, advertising, etc. The processing flow includes: extracting key frames using frame difference method, optical flow method, and CNN-LSTM method, and removing redundant frames. For example, filter out key frames involving vehicle damage from "user complaint videos". Detect vehicle brands, models, road conditions, etc. through algorithms such as YOLO and Faster R-CNN; for example, automatically identify "Tesla Model Y" in the review video. Perform OCR recognition on subtitles or annotation information in the video to extract text information; for example, extract vehicle parameters from "car advertising videos". Extract voice content in the video and convert it into text using ASR; perform sentiment analysis on the comments of the commentator or user.

[0035] S3: Based on the improved Transformer deep learning model, combined with the self-attention mechanism and multimodal feature fusion technology, cross-context correlation analysis is performed on the different features extracted from multimodal data, and the accuracy of identifying the user's true emotional tendency is evaluated based on the analysis results.

[0036] Collect user-generated automotive industry-related data from different channels in real time, including: text (social media, forums, news comments, etc.); voice (telephone customer service, user video comments); image (user-uploaded accident photos, product advertisements); video (car reviews, news reports, user experience sharing).

[0037] Data preprocessing and feature extraction, text: word segmentation (BERT-WordPiece), denoising, sentiment word recognition; speech: automatic speech recognition (ASR), speech emotion feature extraction (Mel frequency cepstral coefficient, MFCC); image: object detection (YOLO, Faster R-CNN), OCR text extraction; video: key frame extraction (CNN-LSTM), speech sentiment analysis.

[0038] Perform cross-context correlation analysis on different features extracted from multimodal data, including: Measure the changing trend of users' emotions towards the same brand / event over different time periods. Traditional sentiment analysis is usually based on static text or sentiment snapshots at a single point in time. By introducing the sentiment drift index, the rate of change of sentiment tendency over time is calculated to identify long-term impacts or short-term fluctuations.

[0039] Among them, the calculation method of the sentiment drift index is: Collect user comments on the same brand / product at multiple time points (such as Weibo posts, forum discussions, news reports, etc.). Use the BERT+LSTM time series model to calculate the sentiment score at each time point. , the expression is: ;in, is the user text at time t, and f is the output function of the LSTM model.

[0040] Calculate the emotion change rate at adjacent time points as the emotion drift index: ; where Δt is the time interval, It is the sentiment drift index.

[0041] Negative SDC (emotional downward trend): If the user's emotion drift coefficient drops sharply in a short period of time (for example, from 0.8 to -0.5), it indicates that the brand / product may have a sudden negative event (such as vehicle recall, accident exposure).

[0042] Positive SDC (sentiment upward trend): If SDC grows rapidly, it means that user sentiment is turning positive, which may be related to factors such as corporate public relations activities and product upgrades.

[0043] Stable SDC (low rate of change): Sentiment remains stable, indicating that brand reputation is more durable.

[0044] It measures the consistency or contradiction of multimodal information such as text, speech, image and video in emotional expression. Existing sentiment analysis usually analyzes a certain modality separately. By obtaining the contextual contradiction index (CDI), it can detect the conflict between modalities and avoid misjudgment.

[0045] The calculation method of contextual contradiction CDI is: Perform emotion recognition on multimodal data and calculate text emotion ST, speech emotion SV, image emotion SI, and video emotion SF respectively. The score range of each emotion is [-1, 1].

[0046] Calculate the emotional consistency EC, the expression is: ; in, Values ​​close to 1 indicate high consistency. Values ​​close to 0 indicate a high level of conflict.

[0047] Calculate the contextual contradiction degree CDI, the expression is: ; The higher the CDI value, the higher the degree of emotional contradiction between different modes.

[0048] High CDI (high contradiction): indicates that the information expressed by the user has emotional conflict. For example, if the user's comment is "This car is great!", but the accompanying picture is a photo of a broken car, it means that the user may be sarcastic or have potential negative emotions.

[0049] Low CDI (low contradiction): indicates that the multimodal data is emotionally consistent, indicating that the user's emotional expression is authentic and credible.

[0050] The Transformer Encoder is used to jointly encode text, speech, image, and video features.

[0051] Modality Alignment technology is used to map different modal data into the same semantic space.

[0052] The self-attention mechanism is used to calculate the weights of different modal information to ensure that the most relevant information receives higher attention.

[0053] Calculation formula (self-attention calculation): ; Among them, Q, K, and V are query, key, and value matrices respectively. is the dimension of the key and T is the matrix transpose.

[0054] Calculate the SDC-Correlation to measure the changing relationship between SDC and CDI: If Corr(SDC, CDI)>0.5, it means that the emotion drift is large, and the emotion contradiction between the modes is serious, and there is implicit negative emotion or sarcastic tone. If Corr(SDC, CDI)<-0.5, it means that the emotion change is stable, the emotion expression is consistent, and the user's evaluation is credible.

[0055] Calculate the matching degree between the emotion labels predicted by the model and the manually annotated emotion labels, that is, calculate the emotion recognition accuracy SPA: ; In the formula, is the sentiment prediction value of the model for the sample, ranging from [-1, 1] (-1 means strongly negative, 0 means neutral, and 1 means strongly positive). GroundTruth is the true sentiment label annotated manually, also ranging from [-1, 1], and N is the total number of samples. If SPA is lower than the threshold (such as 85%), the model is optimized.

[0056] Adopt reinforcement learning (RL) training strategy and optimize the emotion recognition model through reward mechanism (Reward Function): ; Use the Policy Gradient Optimization method to adjust the parameters of the Transformer model to improve the accuracy of future predictions.

[0057] If it is detected that SDC decreases and CDI increases, it will trigger a corporate public opinion warning, reminding the brand to take measures. Combined with historical data, the knowledge graph is used to generate automated response strategies, such as: issuing official statements in advance, adjusting marketing strategies, and improving product design.

[0058] S4: For the accurate identification of users’ real emotional tendencies, a real-time emotional trend prediction mechanism is constructed through a time series analysis model. When an abnormal event is detected, an early warning signal is automatically sent to the preset enterprise management terminal.

[0059] Sentiment trend prediction requires a continuous time-based data stream, which mainly includes the following information: Text Sentiment (ST): social media comments, news, forum posts (BERT / BART calculates sentiment scores); Voice Sentiment (SV): user voice feedback, complaint calls (Wav2Vec2.0 calculates sentiment scores); Image Sentiment (SI): accident photos and promotional advertisements uploaded by users (ResNet + OCR to recognize visual emotions); Video Sentiment (SF): user reviews, news reports, short videos (TimeSformer calculates video sentiment trends).

[0060] These sentiment data are stored in time series and normalized: Among them, S(t) represents the comprehensive emotional state at time t.

[0061] Construct a time series sentiment dataset: aggregate data by time window (such as 1 hour, 1 day, 1 week), and calculate the average sentiment value, sentiment fluctuation amplitude, and abnormal sentiment deviation. Traditional time series models (such as ARIMA) cannot effectively capture nonlinear and long-term dependent sentiment changes, so this method uses the Transformer prediction model: For large-scale time series data, this method uses TimeSformer (Time Series Transformer) for prediction and uses the self-attention mechanism to model long-term emotional changes: ; Among them, Q, K, and V represent the query, key, and value matrices of sentiment data respectively, and softmax calculates the correlation to ensure that important information receives higher attention.

[0062] Abnormal events are usually manifested as sudden and drastic changes in emotion values. Therefore, this method calculates the emotion anomaly score (AE, Anomaly Score) to determine whether to trigger an early warning: ; Where: S(t) is the current sentiment score, is the normal emotional trend predicted by the model, is the historical emotional fluctuation range; if >θ (set threshold, such as 2.5), it is considered that abnormal emotional fluctuations have occurred, and an early warning signal is automatically sent to the preset enterprise management end.

[0063] S5: Utilize knowledge graphs and reinforcement learning technology, combined with historical data and the accuracy of identifying users’ real emotional tendencies, to generate response strategies for public opinion events and provide visual analysis reports to assist companies in formulating accurate market response plans.

[0064] Knowledge graphs are used to represent the multi-level semantic relationships of public opinion events, facilitating situational analysis, risk assessment, and response strategy reasoning.

[0065] The knowledge graph construction includes: entities (Nodes): brands (Tesla, BYD), models (Model Y, Song PLUS), events (spontaneous combustion, recall), markets (China, the United States), etc.; relationships (Edges): vehicle failure → recall, product upgrade → user praise, negative news; using natural language processing (NLP) + relationship extraction, extract triples from news, social media, and industry reports to construct a dynamic knowledge base.

[0066] Reinforcement Learning (RL) is used to train intelligent policy models so that they can automatically optimize corporate response measures in different public opinion situations. The core Agent-Environment interaction architecture of reinforcement learning: Agent: public opinion management system; State: current public opinion situation (event type, sentiment trend, market impact, etc.); Action: possible corporate response strategies (such as official statements, product recalls, price adjustments, etc.). Reward: measure the effectiveness of each strategy based on historical data, such as user sentiment recovery and brand trust improvement.

[0067] Reinforcement learning optimization strategy process: State Representation: Input current public opinion information (negative sentiment index, event severity, etc.); Policy Selection: Use Q-learning / DDPG / PPO and other algorithms to select the best response measures; Feedback Learning (Reward Calculation): Update strategies based on user sentiment recovery and market impact analysis; Continuous Optimization (Policy Improvement): Optimize the strategy recommendation system through multiple rounds of interactive training.

[0068] Generate accurate public opinion response strategies through reinforcement learning historical decision backtracking + real user sentiment data. Historical similar event mining: Based on knowledge graph query, find similar public opinion events in the past. Effect evaluation: Calculate the public opinion recovery speed (Recovery Rate, RR) and user sentiment change (SDC, CDI) of historical strategies. Automatically learn the optimal strategy: The reinforcement learning model compares the effects of different strategies and selects the optimal solution. Example: In 2022, a certain electric vehicle brand caused spontaneous combustion public opinion due to battery failure, and finally reduced the negative impact through official recall + compensation measures. At present, a certain brand has also encountered similar public opinion due to battery problems. The system recommends optimizing by learning from the best historical strategy.

[0069] Analysis of real user emotions: Text analysis (social media, forum comments): extraction of negative emotional keywords (such as "out of control" and "recall"); Voice emotion analysis (customer service, complaint recordings): extraction of anger / anxiety emotional features; Multimodal emotion fusion (pictures / videos): detection of emotional signals such as user expressions and accident photos; The system adjusts the response plan based on the user's real emotional data: If the user is mainly concerned about safety → strengthen technical transparency and release safety improvement measures; If the user's anger is dominant → give priority to compensation / recall strategies; If public opinion has not yet spread completely → respond in a low-key manner to avoid expanding negative impacts. Finally, the system displays public opinion dynamics through data visualization and provides intelligent decision-making support. Generate public opinion visualization reports, including: public opinion heat trends (negative news & changes in social discussion volume); emotional tendency changes (user praise rate, complaint rate changes); competitor comparative analysis (Does the competing product also encounter similar public opinion? How to deal with it?).

[0070] Sample visualization content: Timeline: shows the whole process of public opinion outbreak, spread and decline; Sentiment Trend: changes in users' positive and negative emotions over time; Geo-Impact Map: analyzes the distribution of public opinion among users in different regions; Scenario Simulation: the expected impact of different response strategies.

[0071] Generating intelligent response strategy suggestions includes: optimal strategy recommendation (based on knowledge graph + reinforcement learning to calculate the best solution); automatic generation of public relations manuscripts (combined with NLP to generate official response text); marketing adjustment suggestions (such as adjusting advertising to reduce losses).

[0072] In this embodiment, first, the system collects multi-source data from social media, news, voice comments, images and videos in real time, and performs feature extraction, including text segmentation, voice transcription, image / video target detection and sentiment analysis. Subsequently, an improved Transformer deep learning model is used, combined with the self-attention mechanism and multimodal feature fusion technology, to perform cross-context association analysis to improve the accuracy of identifying the user's true emotional tendency. Based on the accurately identified emotional data, the system predicts emotional trends through a time series analysis model, and automatically sends an early warning to the enterprise management end when an abnormal event is detected. Finally, using knowledge graphs and reinforcement learning technology, combined with historical data and user emotional trends, the optimal public opinion response strategy is automatically generated, and a visual analysis report is provided to help companies make accurate market decisions. The system can improve the company's public opinion monitoring, risk warning and brand management capabilities, and enhance market competitiveness.

[0073] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0074] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented by software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state hard disk.

[0075] It should be understood that the term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there may be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone, where A and B can be singular or plural. In addition, the character " / " in this article generally indicates that the objects associated with each other are in an "or" relationship, but it may also indicate an "and / or" relationship, which can be understood by referring to the context. A person of ordinary skill in the art can appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this article can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0076] The above description is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A real-time processing method for automobile data based on artificial intelligence, characterized in that: The following steps are involved: S1: Real-time acquisition of automotive-related multimodal data from multiple data sources, including text, voice, images, and videos; S2: Extract features from the collected multimodal data, including word segmentation, denoising and sentiment word recognition for text data, speech recognition and conversion of speech data into text, and object detection and sentiment feature extraction for image and video data; S3: Based on the improved Transformer deep learning model, combined with the self-attention mechanism and multimodal feature fusion technology, cross-context correlation analysis is performed on different features extracted from multimodal data, and the accuracy of identifying the user's true emotional tendency is evaluated based on the analysis results; S4: For the accurate identification of the user's true emotional tendency, a real-time emotional trend prediction mechanism is built through a time series analysis model, and when an abnormal event is detected, an early warning signal is automatically sent to the preset enterprise management end; S5: Utilize knowledge graphs and reinforcement learning technology, combined with historical data and the accuracy of identifying users’ real emotional tendencies, to generate response strategies for public opinion events and provide visual analysis reports to assist companies in formulating accurate market response plans.

2. The method for real-time processing of automobile data based on artificial intelligence according to claim 1, characterized in that: The S1 multimodal data collection adopts a combination of Web crawlers, API interface calls and streaming data collection to obtain real-time and comprehensive automotive industry-related data from social media, news platforms, video platforms, user forums, e-commerce platforms and government regulatory agencies.

3. The method for real-time processing of automobile data based on artificial intelligence according to claim 1, characterized in that: The S2 multimodal data feature extraction includes the joint processing of text, speech, image and video data, wherein: text data is processed by BERT-WordPiece word segmentation, denoising and sentiment dictionary matching; speech data is processed by Wav2Vec2.0 and CNN+LSTM model for sentiment recognition; image data is processed by YOLO, Faster R-CNN and ResNet for target detection and OCR recognition; video data is processed by TimeSformer and CNN-LSTM for key frame extraction and semantic sentiment analysis.

4. The method for real-time processing of automobile data based on artificial intelligence according to claim 1, characterized in that: The S3 is based on cross-context sentiment analysis of the improved Transformer model, extracts correlation features among text, speech, images and videos through a self-attention mechanism, and maps data of different modalities to a unified sentiment representation space through modal alignment technology to improve the accuracy of cross-modal sentiment recognition.

5. The method for real-time processing of automobile data based on artificial intelligence according to claim 4 is characterized in that: The sentiment drift index is used to calculate the sentiment change trend of users in different time periods, and the contextual contradiction degree is used to evaluate the sentiment consistency of multimodal data.

6. The method for real-time processing of automobile data based on artificial intelligence according to claim 5 is characterized in that: in, The calculation method of sentiment drift index is as follows: collect users’ comments on the same brand or product at multiple time points, use BERT+LSTM time series model to calculate the sentiment score at each time point , the expression is: ;in, is the user text at time t, f is the output function of the LSTM model; the sentiment change rate at adjacent time points is calculated as the sentiment drift index: ; where Δt is the time interval, is the sentiment drift index; The calculation method of contextual contradiction degree CDI is as follows: perform emotion recognition on multimodal data, calculate text emotion ST, voice emotion SV, image emotion SI, and video emotion SF respectively, and the score range of each emotion is [-1,1]; calculate emotion consistency EC, the expression is: ; Calculate the contextual contradiction degree CDI, the expression is: .

7. The method for real-time processing of automobile data based on artificial intelligence according to claim 6 is characterized in that: The Transformer encoding layer is used to jointly encode text, speech, image, and video features; the modality alignment technology is used to map different modal data to the same semantic space; the self-attention mechanism is used to calculate the weights of different modal information , calculation formula: ; where Q, K, and V are query, key, and value matrices, respectively. is the dimension of the key, T is the matrix transpose; Measuring the changing relationship between SDC and CDI: If Corr(SDC, CDI) > 0.5, it means that the emotion drift is large, and the emotion contradiction between the modes is serious, and there is implicit negative emotion or sarcastic tone; if Corr(SDC, CDI)< -0.5, it means that the emotion change is stable, the emotion expression is consistent, and the user's evaluation is credible.

8. The method for real-time processing of automobile data based on artificial intelligence according to claim 7 is characterized in that: Calculate the matching degree between the emotion labels predicted by the model and the manually annotated emotion labels, that is, calculate the emotion recognition accuracy SPA: ; Where Predicted Sentiment is the sentiment prediction value of the model for the sample, ranging from [-1, 1]; Ground Truth is the true sentiment label manually annotated, ranging from [-1, 1], N is the total number of samples, if SPA is lower than the threshold, the model is optimized; Adopt reinforcement learning training strategy and optimize emotion recognition model through reward mechanism: ; Use policy gradient optimization methods to adjust the parameters of the Transformer model to improve the accuracy of future predictions.

9. The method for real-time processing of automobile data based on artificial intelligence according to claim 1, characterized in that: The S4 time-series sentiment trend prediction mechanism is based on an improved Transformer structure and combined with a long short-term memory network to construct a real-time public opinion change prediction model. When it detects that the abnormal sentiment fluctuation value exceeds the set threshold, it automatically sends a warning signal to the enterprise management end.

Images