Travel emergency state judgment method and device based on semantic analysis and storage medium

Abstract

The invention discloses a travel emergency state judgment method and device based on semantic analysis and a storage medium, and relates to the technical field of classification models. Comprising the following steps: extracting speech characteristic parameters such as speech speed, volume, speech signal spectrum entropy and the like by collecting speech data of an independent user with known emotion and target intention, and calculating a speech judgment coefficient; and based on the voice text data, performing voice recognition by using a knowledge perception network model, and generating a danger feature comprehensive vector. And using the vector to train a neural network model, taking the danger feature comprehensive vector as input, and predicting a corresponding emotion and a target intention. And monitoring voice data of the independent user in real time, obtaining a corresponding danger feature comprehensive vector, calculating a comprehensive emergency state judgment index by combining the voice judgment coefficient and the environment data, comparing the index with a risk threshold value, and outputting an emergency state instruction. The method can effectively predict the emergency situation of the user in travel, and provides support for safety emergency.

Description

Technical Field

[0001] The present invention relates to the technical field of classification models, and in particular to a travel emergency state judgment method, device and storage medium based on semantic analysis. Background Art

[0002] With the development of society, the travel safety of users, especially those traveling alone, has increasingly attracted social attention. When traveling in public places or remote areas, solo users are easily targeted by criminals, and their safety issues are often ignored in daily life. Although traditional safety precautions, such as alarm systems and monitoring equipment, can play a warning role to a certain extent, these measures often rely on the intervention of the external environment, and in emergencies, the user's own emotional changes and the degree of danger on the scene are not fully considered. Therefore, how to conduct effective emergency status assessment and response in a smarter and more real-time manner in the early stages of a crisis has become a technical problem that needs to be solved urgently.

[0003] At present, technologies based on sentiment analysis and speech recognition are being widely used, especially in smart assistants and voice control systems. However, existing speech recognition systems mainly focus on text transcription and simple sentiment analysis, and still have certain limitations in judging complex emotional states and potential dangerous situations. Especially in the field of emergency response, how to accurately capture the user's emotional changes, analyze potential dangerous scenarios, and combine them with surrounding environmental data to issue corresponding emergency instructions in real time is a problem that has not yet been fully solved.

[0004] In traditional sentiment analysis technology, there is a lack of in-depth understanding of complex emotions and intention recognition. The identification of dangerous scenes is more dependent on environmental perception devices, which are often limited by factors such as distance and viewing angle, and it is difficult to accurately judge all potential dangers. Therefore, traditional emergency status judgment methods have defects such as untimely response and poor accuracy. Especially in emotionally unstable or complex emergencies, the system often cannot recognize and respond to the emotional changes of travelers in a timely manner, resulting in delayed emergency response and inability to effectively ensure the safety of users' travel.

[0005] Therefore, how to use semantic analysis technology, combined with voice data and surrounding environment data, to achieve accurate emergency status judgment in emergency situations, especially personalized emotion analysis based on user voice characteristics, has become a technical problem that urgently needs to be solved.

[0006] In the prior art, the publication number CN114530146A discloses an intelligent judgment method, device, equipment and medium for emergency status, the method includes: receiving the initial voice information input by the user and extracting the audio coefficient information, further identifying the audio coefficient information to obtain the corresponding text information, judging whether the text information meets the image acquisition condition, if it meets the condition, obtaining the user image information corresponding to the initial voice information and extracting the corresponding posture feature information from it, judging the text information and posture feature information according to the judgment model to obtain the judgment result of whether it is in an emergency state. However, in this scheme, if the user is in a poorly lit environment, moving or in a scene with limited visual range, it may be difficult to obtain clear image information. Even if the image is obtained, the posture feature extraction may be affected by occlusion, shooting angle or human posture complexity, resulting in misjudgment. And the posture feature extraction consumes a lot of computing resources. In emergency scenarios with high real-time requirements, delays and performance bottlenecks may occur. Therefore, this scheme will reduce the real-time and effectiveness of monitoring.

[0007] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0008] The purpose of the present invention is to provide a travel emergency status judgment method, device and storage medium based on semantic analysis to solve the problems raised in the above background technology.

[0009] To achieve the above object, the present invention provides the following technical solutions:

[0010] A method for judging travel emergency status based on semantic analysis, the specific steps include:

[0011] Collecting a number of speech data with known emotions and target intentions, extracting corresponding speech feature parameters from the speech data, decoding the speech data to obtain speech text data, and calculating a speech judgment coefficient based on the speech feature parameters, wherein the speech feature parameters include the user's speech speed, volume, and speech signal spectral entropy;

[0012] Based on the obtained speech text data, the speech text data is recognized by using the knowledge perception network model, the data features of the dangerous scene description are extracted and enhanced, and the input speech text data is converted and output into a comprehensive vector of dangerous features through the knowledge perception network model, wherein the comprehensive vector of dangerous features includes speech emotion and text description data of the dangerous scene;

[0013] Based on the comprehensive vector of dangerous features, a neural network learning model is established, and the obtained comprehensive vector of dangerous features is used as the input of the neural network learning model. The neural network learning model is trained with the known corresponding emotions and target intentions as labels, and the input is the comprehensive vector of dangerous features, and the output is the emotion analysis prediction model of the corresponding emotions and target intentions.

[0014] The voice data of the monitored user in different time windows are collected, and the collected voice data are identified and separated to obtain the voice data of the monitored user and the voice data of the interlocutor. Based on the two types of voice data, the voice feature parameters are extracted to calculate the voice judgment coefficient, and the voice judgment coefficient is converted into voice text data and input into the knowledge perception network model to obtain the comprehensive vector of danger features. The comprehensive vector of danger features is input into the trained emotion analysis prediction model, and the emotion analysis prediction model outputs the corresponding emotion and target intention.

[0015] The corresponding emotions and target intentions outputted are numerically encoded and converted into specific numerical values. The comprehensive emergency state judgment index of the monitored user and the interlocutor is generated by combining the speech judgment coefficient and the surrounding environment data. The average emergency state judgment index is calculated based on the comprehensive emergency state judgment index of the monitored user and the interlocutor in adjacent time windows. The average emergency state judgment index is compared with the risk judgment threshold, and different emergency state instructions are issued according to different comparison results.

[0016] Furthermore, voice data of several users with known emotions and target intentions are collected, and voice feature parameters of the corresponding users are extracted from the voice data, wherein the method used to extract the voice feature parameters of the corresponding users is MFCC processing, and the specific steps include voice signal segmentation, window function processing, FFT conversion, Mel filter and logarithmic operation to obtain voice feature parameters including user's speech speed, volume and voice signal spectral entropy;

[0017] The specific method for decoding the voice data to obtain the voice text data is as follows: input the voice data into the Wav2Vec2.0 model, and use the CTC decoding method to convert the voice data of the lone user into text data.

[0018] Furthermore, a speech judgment coefficient is calculated based on the speech feature parameters, wherein the formula for calculating the speech judgment coefficient is:

[0019]

[0020] Where C is the speech judgment coefficient, V is the user's speaking speed, S is the volume, G represents the speech signal spectral entropy, β is the adjustment factor, specifically a constant greater than 1, α, γ and δ are the weight coefficients of the user's speaking speed, volume and speech signal spectral entropy, respectively, where δ>γ≥α and α, γ and δ are all greater than 0.

[0021] Furthermore, based on the obtained speech text data, the knowledge-aware network model is used to identify the speech text data, extract and enhance the data features of the dangerous scene description, wherein the specific steps of extracting and enhancing the data features of the dangerous scene description include: knowledge embedding, input representation, feature extraction and attention mechanism and feature output processing text, and knowledge embedding is achieved by constructing an emotional vocabulary library and a dangerous scene vocabulary library, wherein the emotional vocabulary library is specifically represented as:

[0022] V emotion = {v e1 ,v e2 ,…,v ep ,…,v en}

[0023] Where V emotion represents the sentiment vocabulary, v ep represents the pth sentiment word, where p is the index of the sentiment word, p = 1, 2, ..., n, and n is the total number of sentiment words;

[0024] The dangerous scene vocabulary library is specifically expressed as:

[0025] V danger = {v d1 ,v d2 ,…,v dl ,…,v dq}

[0026] Where V danger represents the dangerous scene vocabulary, v dl is the lth dangerous scene word, where l is the index of the dangerous scene word, l = 1, 2, ..., q, and q is the total number of dangerous scene words;

[0027] Each word in the sentiment vocabulary library and the dangerous scene vocabulary library is represented by a vector generated by the Kolmogorov-Arnold representation theorem and is recorded as a knowledge vector.

[0028] Furthermore, each word in the knowledge vector is converted into a word vector, and the dimension of the word vector model used is set to z. The input representation of the entire speech text is constructed based on the word vector, which is specifically represented as an m×z matrix, where m represents the number of words in the speech text;

[0029] The data features of the dangerous scene description extracted and enhanced by the knowledge perception network model are based on: the dangerous features of the scene are extracted by calculating the hidden layer of the knowledge perception network model, wherein the output of the hidden layer of the knowledge perception network model is specifically based on the formula:

[0030]

[0031] In the formula, H ij is the output of the jth hidden layer node of the word input of the i-th speech text, with a dimension of m×z, where j is the index of the hidden layer node, j=1,2,…,h, h represents the total number of hidden layer nodes, φ jk is a parameterized activation function, Represents the kth element in the word vector, w jk is the weight between the jth hidden layer node and the kth word vector element, k represents the index of the word vector element, where k = 1, 2, …, z;

[0032] The weight w between the jth hidden layer node and the kth word vector element is jk is the element in the weight matrix W from the input layer to the hidden layer. The dimension of the weight matrix W is z×h.

[0033] Furthermore, the input speech text data is converted into a comprehensive vector of dangerous features through the knowledge perception network model, wherein the specific formula for calculating the comprehensive vector of dangerous features is:

[0034]

[0035] In the formula, is the comprehensive vector of risk characteristics, α ij represents the attention weight, where the attention weight α ij The calculation is based on the formula:

[0036]

[0037] In the formula, e ij represents the matching degree between the word input of the i-th speech text at the j-th hidden layer node and the attention query vector, where e ij The calculation is based on the formula:

[0038]

[0039] Where W a is the weight matrix of the attention mechanism, with a dimension of h×h, b a is the bias term, is the attention query vector set with dimension h.

[0040] Further, based on the comprehensive vector of risk characteristics, a neural network learning model is established, wherein the neural network learning model is an LSTM model;

[0041] The corresponding emotions and target intentions of the output are numerically encoded. The logic for converting the corresponding emotions and target intentions into specific values ​​is as follows: set the corresponding values ​​according to the intensity of the emotions. Let the emotion category value set F = {f1, f2, …, f M} and the intention recognition set B = {b1,b2,…,b N}, the data in the set correspond to the numerical codes of different emotions and target intentions;

[0042] Collecting real-time voice data of the user to be monitored, the specific collection method is: collecting real-time voice data of several time windows, setting the boundary of each time window, the boundary includes the start time and the end time, setting the collection interval, and acquiring the real-time voice data according to the collection interval and the time window boundary;

[0043] At the same time, if the real-time voice data collected in the time window contains conversation content, the real-time voice data of the user to be monitored and the real-time voice data of the interlocutor are obtained through voice recognition and separation, and the real-time voice data of the two are analyzed separately, among which whether to separate is determined by whether there is a difference in pitch and timbre in the collected voice data.

[0044] Furthermore, the voice judgment coefficient and the surrounding environment data are combined to calculate and generate a comprehensive emergency state judgment index, wherein the surrounding environment data includes the number and density of surrounding commercial buildings and the distance to the nearest public facilities. The formula for calculating the comprehensive emergency state judgment index is:

[0045] AH=ω1*C+ω2*(f+b) 2 +ω3*R(M bui ,ρ bui ,L bui )

[0046] Where AH is the comprehensive emergency state judgment index, f is the emotion code of the user to be monitored, b is the target intention code of the user to be monitored, R(M bui ,ρ bui ,L bui ) is the surrounding environment influence function, M bui , bui and L bui Respectively represent the number and density of surrounding commercial buildings and the distance to the nearest public facilities, ω1, ω2 and ω3 represent the weight coefficients of speech judgment coefficient, emotion and target intention encoding and surrounding environment influence function, where ω2>ω1>ω3 and ω1, ω2 and ω3 are all greater than 0;

[0047] Surrounding environment influence function R(M bui ,ρ bui ,L bui )The specific expression is:

[0048]

[0049] Among them, the method for calculating the comprehensive emergency state judgment index of the interlocutor is consistent with the method for calculating the comprehensive emergency state judgment index of the monitored user;

[0050] The average emergency state judgment index is generated based on the comprehensive emergency state judgment index of different time windows, where the formula for calculating the average emergency state judgment index is:

[0051]

[0052] In the formula, AH′ is the average emergency state judgment index, AH f-1 and AH f AH″ respectively represents the comprehensive emergency state judgment index corresponding to the f-1th time window and the fth time window of the monitored user. f-1 and AH″ f They represent the comprehensive emergency state judgment index corresponding to the interlocutor’s f-1th time window and fth time window respectively;

[0053] The average emergency state judgment index is compared with the risk judgment threshold, and different emergency state instructions are issued according to different comparison results. The specific logic is as follows:

[0054] When AH′≥1.0*yz, an emergency command is issued, indicating that the monitored user is currently in a dangerous situation;

[0055] When 0.4*yz≤AH′<1.0*yz, a risk instruction is issued, indicating that the monitored user is currently at risk of encountering danger;

[0056] When 0≤AH′<0.4*yz, a safety instruction is issued, indicating that the current situation of the monitored user is safe.

[0057] The present invention also provides a travel emergency state judgment device based on semantic analysis, and the travel emergency state judgment device based on semantic analysis is used in a travel emergency state judgment method based on semantic analysis, comprising:

[0058] A voice feature data extraction module is used to collect voice data of a number of known emotions and target intentions, extract corresponding voice feature parameters from the voice data, decode the voice data to obtain voice text data, and calculate the voice judgment coefficient based on the voice feature parameters, wherein the voice feature parameters include the user's speech speed, volume and voice signal spectral entropy;

[0059] A text feature description module is used to identify the speech text data based on the obtained speech text data using a knowledge perception network model, extract and enhance the data features of the dangerous scene description, and convert the input speech text data into a comprehensive vector of dangerous features through the knowledge perception network model. The comprehensive vector of dangerous features includes speech emotion and text description data of the dangerous scene;

[0060] The prediction model training module is used to establish a neural network learning model based on the comprehensive vector of dangerous features, use the obtained comprehensive vector of dangerous features as the input of the neural network learning model, and use the known corresponding emotions and target intentions as labels to train the neural network learning model, so that the input is the comprehensive vector of dangerous features, and the output is the emotion analysis prediction model of the corresponding emotions and target intentions;

[0061] The real-time speech prediction module is used to collect speech data of the monitored user in different time windows, identify and separate the collected speech data, obtain the speech data of the monitored user and the speech data of the interlocutor, extract speech feature parameters based on the two types of speech data, calculate the speech judgment coefficient, and convert them into speech text data to input into the knowledge perception network model to obtain the comprehensive vector of danger features, input the comprehensive vector of danger features into the trained emotion analysis prediction model, and the emotion analysis prediction model outputs the corresponding emotion and target intention;

[0062] The emergency situation judgment module is used to numerically encode the output corresponding emotions and target intentions, convert the corresponding emotions and target intentions into specific numerical values, combine the voice judgment coefficient and the surrounding environment data, generate the comprehensive emergency state judgment index of the monitored user and the interlocutor, and calculate the average emergency state judgment index based on the comprehensive emergency state judgment index of the monitored user and the interlocutor in adjacent time windows. The average emergency state judgment index is compared with the risk judgment threshold, and different emergency state instructions are issued according to different comparison results.

[0063] The present invention also provides a travel emergency state judgment storage medium based on semantic analysis, on which a computer program is stored. When the program is executed by a processor, the travel emergency state judgment method based on semantic analysis as described above is implemented.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] First, a technical solution combining speech feature parameters with emotion analysis is introduced. By extracting the features of speech speed, volume and speech signal spectral entropy from the user's speech data, and combining it with the analysis of speech text data, it can capture the emotional changes in the user's voice in real time, and promptly identify the psychological and emotional fluctuations of the user when facing different situations. Secondly, by combining the knowledge perception network model and the neural network learning model, the system can perform deep learning and analysis in dangerous scenarios. By constructing a comprehensive vector of dangerous features, the system can extract potential dangerous scenario information from the speech data and enhance the description of the dangerous scene, greatly improving the accuracy and sensitivity of emergency judgment. In addition, through the training of known emotions and target intentions, the emotion analysis prediction model can accurately predict the emotions and target intentions of the monitored user based on the input comprehensive vector of dangerous features. Finally, combined with real-time environmental data, it can more intelligently judge the emergency state of travelers, thereby providing customized and personalized protection for user travel safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a schematic diagram of the overall method flow of the present invention;

[0067] Figure 2 It is a schematic diagram of the overall structure of the device of the present invention. DETAILED DESCRIPTION

[0068] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.

[0069] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0070] Example:

[0071] See also Figure 1 , the present invention provides a technical solution:

[0072] A method for judging travel emergency status based on semantic analysis, the specific steps include:

[0073] Step 1: Collect some speech data with known emotions and target intentions, extract corresponding speech feature parameters from the speech data, decode the speech data to obtain speech text data, and calculate the speech judgment coefficient based on the speech feature parameters, wherein the speech feature parameters include the user's speaking speed, volume and speech signal spectral entropy.

[0074] Collect a number of user voice data with known emotions and target intentions, and extract the voice feature parameters of the corresponding user from the voice data. The method used to extract the voice feature parameters of the corresponding user is MFCC processing. The specific steps include segmenting the voice signal to divide the continuous voice signal into short time frames. Usually, the length of these frames is between 20ms and 40ms, and the overlapping part is generally taken as 50% of the frame length; window function processing, applying a window function (such as Hamming window or Hanning window) to each frame to reduce the boundary effect. The window function can help smooth the signal and reduce the frequency domain leakage phenomenon; FFT conversion, fast Fourier transform (FFT) is performed on the frame processed by the window function to convert the time domain signal into the frequency domain signal. This step can obtain the spectrum information of the signal; Mel filter, apply Mel frequency filter group to the FFT result, Mel scale is a way to simulate the human ear's perception of different frequencies. The filter divides the spectrum into multiple bandpass filters, so that the low-frequency part has a higher resolution; logarithmic operation, take the logarithm of the output of the Mel filter. This step is to simulate the perception of the sound amplitude by the human auditory system. Finally, the speech feature parameters obtained include the user's speaking speed, volume and speech signal spectral entropy;

[0075] The specific method for decoding the voice data to obtain the voice text data is as follows: input the voice data into the Wav2Vec2.0 model, and use the CTC decoding method to convert the voice data of the lone user into text data.

[0076] Wav2Vec2.0 is a self-supervised learning speech model developed by Facebook AI (now Meta), which is suitable for low-resource speech recognition tasks. CTC decoding is a decoding method widely used in sequence-to-sequence tasks such as speech recognition and handwriting recognition. It calculates the probability distribution of the output characters and minimizes the difference between the target text sequence and the true label to output the correct character sequence. The Wav2Vec2.0 model and CTC decoding method are both mature existing technologies, and the specific operation process is not repeated here.

[0077] The speech judgment coefficient is calculated based on the speech feature parameters, wherein the formula for calculating the speech judgment coefficient is:

[0078]

[0079] Where C is the speech judgment coefficient, V is the user's speaking speed, S is the volume, G represents the speech signal spectral entropy, β is the adjustment factor, specifically a constant greater than 1, α, γ and δ are the weight coefficients of the user's speaking speed, volume and speech signal spectral entropy, respectively, where δ>γ≥α and α, γ and δ are all greater than 0.

[0080] The user's speech rate V represents the speech flow rate per unit time, which is usually related to the tension of emotions. A faster speech rate may be associated with excitement, anxiety or anger, while a slower speech rate may indicate calmness or fatigue. Therefore, in an emergency, the speech rate will generally increase. A logarithmic function is used to adjust the effect of volume on the judgment coefficient. The relationship between speech rate and emotion is not linear, especially within a larger range of speech rate changes. The use of a logarithmic function can smooth this relationship and avoid excessive volume leading to overly extreme results.

[0081] Volume S, volume indicates the intensity of the voice of the speaker. When calling for help or asking for help, the volume will increase, and when nervous and afraid, the volume will generally decrease. Therefore, the volume changes in both directions, so the design part uses the square term to reflect the impact of volume changes on emotions. Since the impact of speech speed on emotions is nonlinear within a certain range, the square is used to emphasize that the impact is greater when the volume is higher. At the same time, the adjustment factor β is introduced. When the volume is less than a certain range, the volume S is inversely proportional to the voice judgment coefficient C. When the volume S is greater than the adjustment factor β, the voice judgment coefficient C increases as the volume S increases.

[0082] The spectral entropy G of speech signal measures the complexity or uncertainty of speech signal. A higher spectral entropy may indicate that the speaker is emotionally unstable or anxious, while a lower spectral entropy may indicate a stable and calm emotion. The spectral entropy of speech signal uses an exponential function to represent the effect of spectral entropy on emotional fluctuations. The higher the spectral entropy, the higher the complexity and uncertainty of speech signal, and the more unstable the emotion, so e is used. -G To reflect the relationship between spectral entropy and judgment coefficient. A higher spectral entropy value will lead to a lower coefficient, indicating an emergency or abnormal state when the emotion is unstable.

[0083] Since the speech signal spectral entropy G directly reflects the fluctuation of emotions, the weight coefficient of the speech signal spectral entropy is the largest. At the same time, since the ability of speech rate to reflect emotions is greater than volume, it is set that δ>γ≥α and α, γ and δ are all greater than 0.

[0084] Step 2: Based on the obtained speech text data, the knowledge-aware network model is used to identify the speech text data, extract and enhance the data features of the dangerous scene description, and the input speech text data is converted and output into a comprehensive vector of dangerous features through the knowledge-aware network model. The comprehensive vector of dangerous features includes speech emotion and text description data of the dangerous scene.

[0085] Based on the obtained speech text data, the knowledge-aware network model is used to identify the speech text data, extract and enhance the data features of the dangerous scene description, wherein the specific steps of extracting and enhancing the data features of the dangerous scene description include: knowledge embedding, input representation, feature extraction and attention mechanism and feature output processing text, and knowledge embedding is achieved by constructing an emotional vocabulary library and a dangerous scene vocabulary library, wherein the emotional vocabulary library is specifically represented as:

[0086] V emotion = {v e1 ,v e2 ,…,v ep ,…,v en}

[0087] Where V emotion represents the sentiment vocabulary, v ep represents the pth sentiment word, where p is the index of the sentiment word, p = 1, 2, ..., n, and n is the total number of sentiment words;

[0088] The dangerous scene vocabulary library is specifically expressed as:

[0089] V danger = {v d1 ,v d2 ,…,v dl ,…,v dq}

[0090] Where V danger represents the dangerous scene vocabulary, v dl is the lth dangerous scene word, where l is the index of the dangerous scene word, l = 1, 2, ..., q, and q is the total number of dangerous scene words;

[0091] Each word in the sentiment vocabulary library and the dangerous scene vocabulary library is represented by a vector generated by the Kolmogorov-Arnold representation theorem and is recorded as a knowledge vector.

[0092] The words in these vocabulary libraries are selected based on in-depth analysis of voice data related to danger of single users in a large number of actual scenarios. The emotional vocabulary library covers the words that single users may use when they feel uneasy, fearful, nervous, etc., such as "afraid", "panic", "terrified", etc.; the dangerous scene vocabulary library contains scene description words that often appear in dangerous events, such as "remote alleys", "dark corners", "unmanned parking lots", etc. Each word is given a vector representation generated based on the Kolmogorov-Arnold representation theorem, which can accurately capture the semantic and emotional characteristics of the word in the danger warning context.

[0093] Convert each word in the knowledge vector into a word vector. The dimension of the word vector model used is set to z. Based on the word vector, the input representation of the entire speech text is constructed. Specifically, it is represented as an m×z matrix, where m represents the number of words in the speech text.

[0094] The data features of the dangerous scene description extracted and enhanced by the knowledge perception network model are based on: the dangerous features of the scene are extracted by the hidden layer calculation of the knowledge perception network model. When the words expressing fear appear in the knowledge vector of a single user, the dangerous features of the scene are extracted by the hidden layer calculation of the KAN model. The specific formula for the output of the hidden layer of the knowledge perception network model is:

[0095]

[0096] In the formula, H ij is the output of the jth hidden layer node of the word input of the i-th speech text, with a dimension of m×z, where j is the index of the hidden layer node, j=1,2,…,h, h represents the total number of hidden layer nodes, φ jk is a parameterized activation function, Represents the kth element in the word vector, w jk is the weight between the jth hidden layer node and the kth word vector element, k represents the index of the word vector element, where k = 1, 2, …, z; this step converts the speech text into a mathematical form that can be processed by a computer, preparing for subsequent operations such as feature extraction.

[0097] The weight w between the jth hidden layer node and the kth word vector element is jk is the element in the weight matrix W from the input layer to the hidden layer. The dimension of the weight matrix W is z×h.

[0098] Parameterized activation functions are those whose behavior is adjusted by parameters (such as weights and biases), where φ jk For the parameterized activation function, a specific function expression is set based on the Kolmogorov-Arnold representation theorem.

[0099] The input speech text data is converted into a comprehensive vector of dangerous features through the knowledge perception network model, and the hidden layer output is weighted and summed according to the attention weight. The specific formula for calculating the comprehensive vector of dangerous features is:

[0100]

[0101] In the formula, is the comprehensive vector of risk characteristics, α ijRepresents the attention weight. The comprehensive vector of dangerous features integrates the importance information of different words in the speech text and can highlight the features related to danger.

[0102] The attention weight α ij The calculation is based on the formula:

[0103]

[0104] In the formula, e ij represents the matching degree between the word input of the i-th speech text at the j-th hidden layer node and the attention query vector, where e ij The calculation is based on the formula:

[0105]

[0106] Where W a is the weight matrix of the attention mechanism, with a dimension of h×h, b a is the bias term, is the attention query vector set with dimension h.

[0107] The attention mechanism calculates the weight of each element in the voice text and allocates attention resources based on the semantics, sentiment, and contextual information of the voice text of a single user. This enables the model to accurately focus on key information when processing the voice text of a single user, thereby improving the accuracy of hazard identification.

[0108] Assume the attention query vector is The dimension is h), and its initial value is scientifically set according to the type of information that single users usually pay attention to in dangerous scenarios.

[0109] Step 3: Based on the comprehensive vector of dangerous features, a neural network learning model is established. The obtained comprehensive vector of dangerous features is used as the input of the neural network learning model, and the neural network learning model is trained with the known corresponding emotions and target intentions as labels. The input is the comprehensive vector of dangerous features, and the output is the emotion analysis prediction model of the corresponding emotions and target intentions.

[0110] Based on the comprehensive vector of risk characteristics, a neural network learning model is established, wherein the neural network learning model is an LSTM model;

[0111] Taking the feature vector containing rich information such as voice emotion, dangerous scenarios and geographical environment processed by the knowledge-aware network model (KAN) as input, LSTM processes the conversation situations encountered by single users in real time. It has the ability to capture long-term dependencies and dynamically analyze dangerous scenarios, significantly improving the accuracy of user emotion analysis and dangerous situation judgment, accurately reflecting the dynamic changes in the degree of danger, and providing a reliable basis for the system to assess danger in real time.

[0112] Step 4: Collect voice data of the user to be monitored in different time windows, identify and separate the collected voice data, obtain the voice data of the user to be monitored and the voice data of the interlocutor, extract voice feature parameters based on the two types of voice data, calculate the voice judgment coefficient, and convert it into voice text data and input it into the knowledge perception network model to obtain a comprehensive vector of dangerous features, and input the comprehensive vector of dangerous features into the trained emotion analysis prediction model, which outputs the corresponding emotion and target intention.

[0113] The three gating mechanisms of LSTM control the inflow, retention, and outflow of information respectively. The forget gate determines which information needs to be forgotten from the memory cell; the input gate determines which new information will be stored in the memory cell; and the output gate controls how much information in the memory cell at the current moment is used as the output of LSTM.

[0114] The forget gate determines in real time whether to retain or discard cell state information. When a single user describes irrelevant information in daily life, the forget gate will gradually forget such information. When statements related to danger appear, the forget gate can identify and retain key danger information and focus on danger signals; the input gate determines in real time whether new information can enter the cell state. When a single user encounters a new situation, it determines its importance and allows it to enter the cell state if it is important. The cell state is updated in real time. If a single user continues to be in a dangerous situation and constantly expresses related feelings, it accumulates danger information, accurately reflects changes in the degree of danger, and provides a basis for the system to assess danger in real time. The output gate determines the output of cell state information in real time.

[0115] LSTM uses a gating mechanism to update cells and hidden states based on information at each time step, capture changes in emotions and dangerous intentions, associate distribution information, and identify changes in emotions from initial doubt to fear, as well as potential dangerous intentions, by updating cell states and hidden states.

[0116] Step 5: Numerically encode the output corresponding emotions and target intentions, convert the corresponding emotions and target intentions into specific numerical values, and generate the comprehensive emergency state judgment index of the monitored user and the interlocutor in combination with the voice judgment coefficient and the surrounding environment data. According to the comprehensive emergency state judgment index of the monitored user and the interlocutor in adjacent time windows, the average emergency state judgment index is calculated, and the average emergency state judgment index is compared with the risk judgment threshold. Different emergency state instructions are issued according to different comparison results.

[0117] The corresponding emotions and target intentions of the output are numerically encoded. The logic for converting the corresponding emotions and target intentions into specific values ​​is as follows: set the corresponding values ​​according to the intensity of the emotions. Let the emotion category value set F =

[0118] {f1,f2,…,f M} and the intention recognition set B = {b1,b2,…,b N}, the data in the set correspond to the numerical codes of different emotions and target intentions;

[0119] Set the corresponding value according to the intensity of the emotion. Emotion categories can include: calm, joy, tension, disgust, fear, etc. Each emotion category corresponds to a numerical code. Each emotion can be represented by an integer or a real number. The intensity increases in turn, and the corresponding value also increases in turn.

[0120] The intensity of intent is usually a quantitative representation of the clarity or urgency of a certain goal. Intent categories include querying information, expressing suggestions, and requesting help, where the corresponding values ​​increase in sequence.

[0121] Collect real-time voice data of the user to be monitored. The specific collection method is: collect real-time voice data of several time windows, set the boundary of each time window, the boundary includes the start time and the end time, set the collection interval, and obtain real-time voice data according to the collection interval and the time window boundary; wherein each time window boundary is set, the duration of each time window is generally 2 to 5 minutes, and the collection interval is generally set to between 30 seconds and 2 minutes.

[0122] At the same time, if the real-time voice data collected in the time window contains conversation content, the real-time voice data of the user to be monitored and the real-time voice data of the interlocutor are obtained through voice recognition and separation, and the real-time voice data of the two are analyzed respectively;

[0123] Among them, by analyzing the characteristics of the voice signal, such as pitch and timbre, it is determined whether there is any conversation content. If there are obvious differences in pitch and timbre in the collected real-time voice data, it is determined that there is a conversation content, and voice recognition and separation are performed. Otherwise, voice recognition and separation are not performed.

[0124] Separation is usually combined with speech recognition. This technology can automatically identify and separate speech segments of different speakers. It includes two steps: speaker segmentation and speaker recognition. The real-time speech data of the two are analyzed separately, specifically including calculating the comprehensive emergency status judgment index of the monitored user and the interlocutor in the corresponding time window.

[0125] The comprehensive emergency state judgment index is calculated by combining the voice judgment coefficient and the surrounding environment data, wherein the surrounding environment data includes the number and density of surrounding commercial buildings and the distance to the nearest public facilities. The formula for calculating the comprehensive emergency state judgment index is:

[0126] AH=ω1*C+ω2*(f+b) 2 +ω3*R(Mbui ,ρ bui ,L bui )

[0127] Where AH is the comprehensive emergency state judgment index, f is the emotion code of the user to be monitored, b is the target intention code of the user to be monitored, R(M bui ,ρ bui ,L bui ) is the surrounding environment influence function, M bui , bui and L bui Respectively represent the number and density of surrounding commercial buildings and the distance to the nearest public facilities, ω1, ω2 and ω3 represent the weight coefficients of speech judgment coefficient, emotion and target intention encoding and surrounding environment influence function, where ω2>ω1>ω3 and ω1, ω2 and ω3 are all greater than 0;

[0128] Among them, the method for calculating the comprehensive emergency state judgment index of the interlocutor is consistent with the method for calculating the comprehensive emergency state judgment index of the user to be monitored.

[0129] The comprehensive emergency status judgment index AH indicates the urgency of the current situation. The larger the AH is, the more dangerous the current environment is. The larger the voice judgment coefficient C is, the more excited the user to be monitored is, and the greater the possibility of encountering an emergency. Therefore, the voice judgment coefficient C is proportional to the comprehensive emergency status judgment index AH.

[0130] The emotion code f of the monitored user and the target intention code b of the monitored user are set according to the intensity of the emotion. The larger the corresponding emotion code f is, the stronger the emergency emotion expressed is. The larger the target intention code b is, the greater the willingness to ask for help expressed is, which means the greater the probability of encountering an emergency situation is. Therefore, the emotion code f and the target intention code b are both proportional to the comprehensive emergency state judgment index AH. At the same time, through the square sum method (f+b) 2 , to indicate a significant response to an emergency situation.

[0131] Surrounding environment influence function R(M bui ,ρ bui ,L bui )The specific expression is:

[0132]

[0133] Among them, the surrounding environment influence function R(M bui ,ρ bui ,L bui) is inversely proportional to the number and density of surrounding commercial buildings. Specifically, the higher the number and density of surrounding commercial buildings, the greater the population density nearby, and the lower the probability of a forced emergency. bui ,ρ bui ,L bui ) is proportional to the comprehensive emergency state judgment index AH, so the number and density of surrounding commercial buildings are proportional to R(M bui ,ρ bui ,L bui ), and adopt The logarithmic form of indicates that the number of surrounding commercial buildings has a significant impact on safety.

[0134] Distance to the nearest public facility L bui The larger it is, the more desolate the current location is, that is, the higher the probability of a forced emergency situation. Therefore, it is proportional to the comprehensive emergency state judgment index AH.

[0135] The specific method for obtaining the number of surrounding commercial buildings is: using online map services (such as Baidu Maps, Amap, etc.) to view the target area, and counting the commercial building information within a radius of 1 km; the density is expressed by the ratio of the number of surrounding commercial buildings to the area within a radius of 1 km; the distance to the nearest public facilities, finding the nearest public facilities, and calculating the distance, where the public facilities include hospitals, police stations, schools, etc.

[0136] The average emergency state judgment index is generated based on the comprehensive emergency state judgment index of different time windows, where the formula for calculating the average emergency state judgment index is:

[0137]

[0138] In the formula, AH′ is the average emergency state judgment index, AH f-1 and aH f AH″ respectively represents the comprehensive emergency state judgment index corresponding to the f-1th time window and the fth time window of the monitored user. f-1 and AH″ f They represent the comprehensive emergency state judgment index corresponding to the interlocutor’s f-1th time window and fth time window respectively;

[0139] Among them, AH f-1 and AH″ f The calculation formula is the same as the calculation formula for the comprehensive emergency status judgment index of the monitored user, except that the data is replaced by the characteristic data of the interlocutor.

[0140] The average emergency state judgment index is compared with the risk judgment threshold, and different emergency state instructions are issued according to different comparison results. The specific logic is as follows:

[0141] When AH′≥1.0*yz, an emergency command is issued, indicating that the monitored user is currently in a dangerous situation;

[0142] When 0.4*yz≤aH′<1.0*yz, a risk instruction is issued, indicating that the monitored user is currently at risk of encountering danger;

[0143] When 0≤AH′<0.4*yz, a safety instruction is issued, indicating that the current situation of the monitored user is safe.

[0144] The present invention also provides a travel emergency state judgment device based on semantic analysis, and the travel emergency state judgment device based on semantic analysis is used in a travel emergency state judgment method based on semantic analysis, comprising:

[0145] A voice feature data extraction module is used to collect voice data of a number of known emotions and target intentions, extract corresponding voice feature parameters from the voice data, decode the voice data to obtain voice text data, and calculate the voice judgment coefficient based on the voice feature parameters, wherein the voice feature parameters include the user's speech speed, volume and voice signal spectral entropy;

[0146] A text feature description module is used to identify the speech text data based on the obtained speech text data using a knowledge perception network model, extract and enhance the data features of the dangerous scene description, and convert the input speech text data into a comprehensive vector of dangerous features through the knowledge perception network model. The comprehensive vector of dangerous features includes speech emotion and text description data of the dangerous scene;

[0147] The prediction model training module is used to establish a neural network learning model based on the comprehensive vector of dangerous features, use the obtained comprehensive vector of dangerous features as the input of the neural network learning model, and use the known corresponding emotions and target intentions as labels to train the neural network learning model, so that the input is the comprehensive vector of dangerous features, and the output is the emotion analysis prediction model of the corresponding emotions and target intentions;

[0148] The real-time speech prediction module is used to collect speech data of the monitored user in different time windows, identify and separate the collected speech data, obtain the speech data of the monitored user and the speech data of the interlocutor, extract speech feature parameters based on the two types of speech data, calculate the speech judgment coefficient, and convert them into speech text data to input into the knowledge perception network model to obtain the comprehensive vector of danger features, input the comprehensive vector of danger features into the trained emotion analysis prediction model, and the emotion analysis prediction model outputs the corresponding emotion and target intention;

[0149] The emergency situation judgment module is used to numerically encode the output corresponding emotions and target intentions, convert the corresponding emotions and target intentions into specific numerical values, combine the voice judgment coefficient and the surrounding environment data, generate the comprehensive emergency state judgment index of the monitored user and the interlocutor, and calculate the average emergency state judgment index based on the comprehensive emergency state judgment index of the monitored user and the interlocutor in adjacent time windows. The average emergency state judgment index is compared with the risk judgment threshold, and different emergency state instructions are issued according to different comparison results.

[0150] The present invention also provides a travel emergency state judgment storage medium based on semantic analysis, on which a computer program is stored. When the program is executed by a processor, the travel emergency state judgment method based on semantic analysis as described above is implemented.

[0151] 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.

[0152] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product. Those skilled in the art may appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software methods depends on the specific application and design constraints of the technical solution.

[0153] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0154] 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 method for judging travel emergency status based on semantic analysis, characterized in that: The specific steps include: Collecting a number of speech data with known emotions and target intentions, extracting corresponding speech feature parameters from the speech data, decoding the speech data to obtain speech text data, and calculating a speech judgment coefficient based on the speech feature parameters, wherein the speech feature parameters include the user's speech speed, volume, and speech signal spectral entropy; Based on the obtained speech text data, the speech text data is recognized by using the knowledge perception network model, the data features of the dangerous scene description are extracted and enhanced, and the input speech text data is converted and output into a comprehensive vector of dangerous features through the knowledge perception network model, wherein the comprehensive vector of dangerous features includes speech emotion and text description data of the dangerous scene; Based on the comprehensive vector of dangerous features, a neural network learning model is established, and the obtained comprehensive vector of dangerous features is used as the input of the neural network learning model. The neural network learning model is trained with the known corresponding emotions and target intentions as labels, and the input is the comprehensive vector of dangerous features, and the output is the emotion analysis prediction model of the corresponding emotions and target intentions. The voice data of the monitored user in different time windows are collected, and the collected voice data are identified and separated to obtain the voice data of the monitored user and the voice data of the interlocutor. Based on the two types of voice data, the voice feature parameters are extracted to calculate the voice judgment coefficient, and the voice judgment coefficient is converted into voice text data and input into the knowledge perception network model to obtain the comprehensive vector of danger features. The comprehensive vector of danger features is input into the trained emotion analysis prediction model, and the emotion analysis prediction model outputs the corresponding emotion and target intention. The corresponding emotions and target intentions outputted are numerically encoded and converted into specific numerical values. The comprehensive emergency state judgment index of the monitored user and the interlocutor is generated by combining the speech judgment coefficient and the surrounding environment data. The average emergency state judgment index is calculated based on the comprehensive emergency state judgment index of the monitored user and the interlocutor in adjacent time windows. The average emergency state judgment index is compared with the risk judgment threshold, and different emergency state instructions are issued according to different comparison results.

2. The method for judging travel emergency status based on semantic analysis according to claim 1, characterized in that: Collecting a number of user voice data with known emotions and target intentions, and extracting the voice feature parameters of the corresponding user from the voice data, wherein the method used to extract the voice feature parameters of the corresponding user is MFCC processing, and the specific steps include voice signal segmentation, window function processing, FFT conversion, Mel filter and logarithmic operation to obtain voice feature parameters including the user's speaking speed, volume and voice signal spectral entropy; The specific method for decoding the voice data to obtain the voice text data is as follows: input the voice data into the Wav2Vec2.0 model, and use the CTC decoding method to convert the voice data of the lone user into text data.

3. The method for judging travel emergency status based on semantic analysis according to claim 2, characterized in that: The speech judgment coefficient is calculated based on the speech feature parameters, wherein the formula for calculating the speech judgment coefficient is: Where C is the speech judgment coefficient, V is the user's speaking speed, S is the volume, G represents the speech signal spectral entropy, β is the adjustment factor, specifically a constant greater than 1, α, γ and δ are the weight coefficients of the user's speaking speed, volume and speech signal spectral entropy, respectively, where δ>γ≥α and α, γ and δ are all greater than 0.

4. The method for judging travel emergency status based on semantic analysis according to claim 1, characterized in that: Based on the obtained speech text data, the knowledge-aware network model is used to identify the speech text data, extract and enhance the data features of the dangerous scene description, wherein the specific steps of extracting and enhancing the data features of the dangerous scene description include: knowledge embedding, input representation, feature extraction and attention mechanism and feature output processing text, and knowledge embedding is achieved by constructing an emotional vocabulary library and a dangerous scene vocabulary library, wherein the emotional vocabulary library is specifically represented as: V emotion ={v e1 ,v e2 ,…,v ep ,…,v en } Where V emotion represents the sentiment vocabulary, v ep represents the pth sentiment word, where p is the index of the sentiment word, p = 1, 2, ..., n, n is the total number of sentiment words; The dangerous scene vocabulary library is specifically expressed as: V danger ={v d1 ,v d2 ,…,v dl ,…,v dq } Where V danger represents the dangerous scene vocabulary, v dl is the lth dangerous scene word, where l is the index of the dangerous scene word, l = 1, 2, ..., q, and q is the total number of dangerous scene words; Each word in the sentiment vocabulary library and the dangerous scene vocabulary library is represented by a vector generated by the Kolmogorov-Arnold representation theorem and is recorded as a knowledge vector.

5. The method for judging travel emergency status based on semantic analysis according to claim 4, characterized in that: Convert each word in the knowledge vector into a word vector. The dimension of the word vector model used is set to z. Based on the word vector, the input representation of the entire speech text is constructed. Specifically, it is represented as an m×z matrix, where m represents the number of words in the speech text. The data features of the dangerous scene description extracted and enhanced by the knowledge perception network model are based on: the dangerous features of the scene are extracted by calculating the hidden layer of the knowledge perception network model, wherein the output of the hidden layer of the knowledge perception network model is specifically based on the formula: In the formula, H ij is the output of the jth hidden layer node of the word input of the i-th speech text, with a dimension of m×z, where j is the index of the hidden layer node, j=1,2,…,h, h represents the total number of hidden layer nodes, φ jk is a parameterized activation function, Represents the kth element in the word vector, w jk is the weight between the jth hidden layer node and the kth word vector element, k represents the index of the word vector element, where k = 1, 2, …, z; The weight w between the jth hidden layer node and the kth word vector element is jk is the element in the weight matrix W from the input layer to the hidden layer. The dimension of the weight matrix W is z×h.

6. The method for determining travel emergency status based on semantic analysis according to claim 5, characterized in that: The input speech text data is converted into a comprehensive vector of dangerous features through the knowledge perception network model. The specific formula for calculating the comprehensive vector of dangerous features is as follows: In the formula, is the comprehensive vector of risk characteristics, α ij represents the attention weight, where the attention weight α ij The calculation is based on the formula: In the formula, e ij represents the matching degree between the word input of the i-th speech text at the j-th hidden layer node and the attention query vector, where e ij The calculation is based on the formula: Where W a is the weight matrix of the attention mechanism, with a dimension of h×h, b a is the bias term, is the attention query vector set with dimension h.

7. The method for judging travel emergency status based on semantic analysis according to claim 2, characterized in that: Based on the comprehensive vector of risk characteristics, a neural network learning model is established, wherein the neural network learning model is an LSTM model; The corresponding emotions and target intentions of the output are numerically encoded. The logic for converting the corresponding emotions and target intentions into specific values ​​is as follows: set the corresponding values ​​according to the intensity of the emotions. Let the emotion category value set F = {f1, f2, …, f M } and the intention recognition set B = {b1,b2,…,b N }, the data in the set correspond to the numerical codes of different emotions and target intentions; Collecting real-time voice data of the user to be monitored, the specific collection method is: collecting real-time voice data of several time windows, setting the boundary of each time window, the boundary includes the start time and the end time, setting the collection interval, and acquiring the real-time voice data according to the collection interval and the time window boundary; At the same time, if the real-time voice data collected in the time window contains conversation content, the real-time voice data of the user to be monitored and the real-time voice data of the interlocutor are obtained through voice recognition and separation, and the real-time voice data of the two are analyzed separately, among which whether to separate is determined by whether there is a difference in pitch and timbre in the collected voice data.

8. The method for judging travel emergency status based on semantic analysis according to claim 7, characterized in that: The comprehensive emergency state judgment index is calculated by combining the voice judgment coefficient and the surrounding environment data, wherein the surrounding environment data includes the number and density of surrounding commercial buildings and the distance to the nearest public facilities. The formula for calculating the comprehensive emergency state judgment index is: AH=ω1*C+ω2*(f+b) 2 +ω3*R(M bui ,r bui ,L bui ) Where AH is the comprehensive emergency state judgment index, f is the emotion code of the user to be monitored, b is the target intention code of the user to be monitored, r(M bui ,ρ bui ,L bui ) is the surrounding environment influence function, M bui , bui and L bui Respectively represent the number and density of surrounding commercial buildings and the distance to the nearest public facilities, ω1, ω2 and ω3 represent the weight coefficients of speech judgment coefficient, emotion and target intention encoding and surrounding environment influence function, where ω2>ω1>ω3 and ω1, ω2 and ω3 are all greater than 0; Surrounding environment influence function R(M bui ,ρ bui ,L bui )The specific expression is: Among them, the method for calculating the comprehensive emergency state judgment index of the interlocutor is consistent with the method for calculating the comprehensive emergency state judgment index of the monitored user; The average emergency state judgment index is generated based on the comprehensive emergency state judgment index of different time windows, where the formula for calculating the average emergency state judgment index is: In the formula, AH′ is the average emergency state judgment index, AH f-1 and AH f AH″ respectively represents the comprehensive emergency state judgment index corresponding to the f-1th time window and the fth time window of the monitored user. f-1 and AH″ f They represent the comprehensive emergency state judgment index corresponding to the interlocutor’s f-1th time window and fth time window respectively; The average emergency state judgment index is compared with the risk judgment threshold, and different emergency state instructions are issued according to different comparison results. The specific logic is as follows: When AH′≥1.0*yz, an emergency command is issued, indicating that the monitored user is currently in a dangerous situation; When 0.4*yz≤AH′<1.0*yz, a risk instruction is issued, indicating that the monitored user is currently at risk of encountering danger; When 0≤AH′<0.4*yz, a safety instruction is issued, indicating that the current situation of the monitored user is safe.

9. A travel emergency status judgment device based on semantic analysis, characterized in that: The travel emergency state judgment device based on semantic analysis is used to execute the travel emergency state judgment method based on semantic analysis according to any one of claims 1 to 8, comprising: A voice feature data extraction module is used to collect voice data of a number of known emotions and target intentions, extract corresponding voice feature parameters from the voice data, decode the voice data to obtain voice text data, and calculate the voice judgment coefficient based on the voice feature parameters, wherein the voice feature parameters include the user's speech speed, volume and voice signal spectral entropy; A text feature description module is used to identify the speech text data based on the obtained speech text data using a knowledge perception network model, extract and enhance the data features of the dangerous scene description, and convert the input speech text data into a comprehensive vector of dangerous features through the knowledge perception network model. The comprehensive vector of dangerous features includes speech emotion and text description data of the dangerous scene; The prediction model training module is used to establish a neural network learning model based on the comprehensive vector of dangerous features, use the obtained comprehensive vector of dangerous features as the input of the neural network learning model, and use the known corresponding emotions and target intentions as labels to train the neural network learning model, so that the input is the comprehensive vector of dangerous features, and the output is the emotion analysis prediction model of the corresponding emotions and target intentions; The real-time speech prediction module is used to collect speech data of the monitored user in different time windows, identify and separate the collected speech data, obtain the speech data of the monitored user and the speech data of the interlocutor, extract speech feature parameters based on the two types of speech data, calculate the speech judgment coefficient, and convert them into speech text data to input into the knowledge perception network model to obtain the comprehensive vector of danger features, input the comprehensive vector of danger features into the trained emotion analysis prediction model, and the emotion analysis prediction model outputs the corresponding emotion and target intention; The emergency situation judgment module is used to numerically encode the output corresponding emotions and target intentions, convert the corresponding emotions and target intentions into specific numerical values, combine the voice judgment coefficient and the surrounding environment data, generate the comprehensive emergency state judgment index of the monitored user and the interlocutor, and calculate the average emergency state judgment index based on the comprehensive emergency state judgment index of the monitored user and the interlocutor in adjacent time windows. The average emergency state judgment index is compared with the risk judgment threshold, and different emergency state instructions are issued according to different comparison results.

10. A storage medium for judging travel emergency status based on semantic analysis, on which a computer program is stored, characterized in that: When the program is executed by a processor, the travel emergency status judgment method based on semantic analysis as described in any one of claims 1 to 8 is implemented.

Images