Personal safety automatic alarm system based on emotion analysis

Abstract

The invention provides a personal safety automatic alarm system based on sentiment analysis. The system converts an externally input sound signal into a characteristic frequency spectrum conforming to a human ear hearing law; continuously training a recognition model by using the characteristic spectrum, inputting a voice signal by the trained recognition model, and outputting the judgment probability of an extreme emotion and an alarm secret signal; comprehensively judging whether to trigger alarm or not according to the judgment probability and by combining the characteristics of the original sound signal; and the audio data and the position information are verified, alarm information is packaged and sent, and the purpose of alarming is achieved. Compared with the prior art, the voice signal detection precision under complex background noise is improved through short-time Fourier transform and an improved Me filtering group, the extreme emotion recognition capability is improved through multi-task network training, and a rapid and hidden safety alarm function is realized through automatic trigger alarm and automatic environmental information collection. The method has remarkable technical advantages in the aspect of improving the user safety protection efficiency and reliability.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and public safety technology, and is particularly suitable for rapid early warning and emergency response of complex emergencies, and relates to an automatic personal safety alarm system based on sentiment analysis. Background Art

[0002] With the development of artificial intelligence technology, sentiment analysis and speech recognition are increasingly used in the field of public safety. However, the current alarm mechanism mostly relies on active triggering by the user, such as making an emergency call or pressing an alarm button, which is difficult to implement in extremely dangerous situations where the victim is restrained, closely monitored or controlled. The current speech emotion recognition has insufficient recognition accuracy in extreme emotional states (such as panic, fear, excitement), the existing code recognition has poor robustness in complex noise scenes, and the alarm response mechanism fails to achieve fast and accurate triggering.

[0003] Specifically: in cases such as kidnapping, human trafficking, and home robbery, victims are often unable to move freely or are under close surveillance, making it impossible to use traditional alarm methods, resulting in a rapid loss of rescue opportunities; in complex environments, factors such as noise and non-standard voices (such as victims being tied up and unable to speak) can significantly interfere with the accuracy of emotion recognition and codeword recognition, resulting in a high false alarm rate; existing alarm mechanisms generally lack the ability to analyze emotional states, and are unable to buy victims valuable rescue time;

[0004] Existing technologies fail to effectively integrate voice codes and emotion recognition technologies, making it difficult to implement a covert and accurate security alarm mechanism.

[0005] In summary, the existing technology mainly has the following problems:

[0006] (1) The alarm trigger mechanism is too dependent on user initiative and lacks concealment and automation;

[0007] (2) lack of efficient identification capabilities based on the victim’s extreme emotional state;

[0008] (3) The detection accuracy of voice codes is not high under complex background noise. Summary of the invention

[0009] The present invention proposes an automatic personal safety alarm system based on sentiment analysis to solve the problems in the prior art of inaccurate extreme emotion recognition by the alarm device, insufficient robustness in voice code recognition, and the technical defects of the alarm mechanism lacking concealment and automation.

[0010] The technical solution of the present invention is achieved in this way:

[0011] An automatic personal safety alarm system based on sentiment analysis, comprising:

[0012] Abnormal speech extraction module, used to convert external input sound signals into characteristic spectrum that conforms to the laws of human hearing;

[0013] The extreme emotion and alarm code recognition module uses the characteristic spectrum to continuously train the recognition model. The trained recognition model inputs the voice signal and outputs the judgment probability of extreme emotions and alarm codes.

[0014] The personal threat alarm trigger module comprehensively determines whether to trigger the alarm based on the probability of extreme emotions and alarm codes, combined with the characteristics of the original sound signal;

[0015] The alarm information sending module verifies the audio data and location information, encapsulates the alarm information and sends it;

[0016] The location information module provides location data for the alarm information sending module;

[0017] The parameter configuration module provides the configuration interface for the alarm protocol format and sending parameters.

[0018] Preferably, the abnormal speech extraction module includes two processes: abnormal speech data processing and feature extraction. The abnormal speech data processing process converts the sound signal into a spectrum sequence indexed by a time window through short-time Fourier transform. The two processes are specifically:

[0019] 1) According to the characteristics of the frequency distribution of the sound signal, the signal is converted into the time-frequency domain through short-time Fourier transform, and the Hanning window is used to capture the time-varying characteristics of the sound signal and identify subtle changes in abnormal speech;

[0020] 2) According to the high-frequency characteristics of abnormal speech signals, a non-uniform Mel filter bank is designed by increasing the mapping bandwidth of high-frequency sound information to generate a non-uniform Mel spectrum that is more sensitive to abnormal speech;

[0021] 3) According to the differences in sound recording by different devices and the impact of environmental noise on abnormal speech recognition, a characteristic spectrum with a uniform range is obtained through normalization processing.

[0022] Preferably, the extreme emotion and alarm signal recognition module includes two processes, training and running. The training process is a multi-task objective function that includes extreme emotion classification, alarm signal recognition and alignment constraints. Specifically, according to the characteristic that extreme emotions have a low probability distribution in the signal, the weight of the low probability distribution data is amplified by the weighted focal loss function and hyperparameters to increase the recognition probability of minority emotion samples; according to the characteristics of the binary classification of alarm signals, the confidence logarithmic loss function is adopted to optimize the confidence of signal recognition; according to the characteristics that there are conflicts in recognition results in the multi-task network, the maximum mean difference feature alignment constraint is adopted to measure and optimize the differences in the distribution of emotion and signal features, and dynamically adjust the loss weights of each task through the weight hyperparameters to reduce conflicts in multi-task learning. The running process runs the model of extreme emotion classification and alarm signal recognition output by the training process, and obtains the probability values ​​of extreme emotions and alarm signals by inputting feature spectrum data into the model.

[0023] Preferably, the personal threat alarm trigger module calculates a judgment value based on extreme emotions and alarm code probability values ​​combined with original sound signal characteristics, and decides whether to trigger an alarm based on the judgment value. The personal threat alarm trigger module also requires that the judgment value be compared with a threshold, and the threshold used for comparison is configurable.

[0024] Preferably, the location information module provides current location data via a geographic information sensor.

[0025] Preferably, the alarm protocol format provides a standardized alarm protocol format specification including the alarm person, user location information and on-site recording data; the configurable parameters in the sending parameters include but are not limited to emergency contacts and alarm methods, and the alarm methods include but are not limited to mobile phone text messages and voice dialing.

[0026] Preferably, the alarm information sending module includes four processes: audio data verification, position data verification, alarm information encapsulation and alarm information sending. The audio data verification is used to check the quality and integrity of the original audio signal; the position data verification is used to check the position data output by the position information module; the alarm information encapsulation uses the alarm protocol format output by the parameter configuration module to package the alarm information into a message; the alarm information sending sends the packaged message to the emergency contact configured in the parameter configuration module.

[0027] Preferably, the extreme emotion classification adopts the weighted focal loss method to define the emotion classification loss function; the alarm signal recognition adopts the confidence logarithmic loss method to define the signal detection loss function; the alignment constraint defines the alignment constraint function according to the maximum mean difference method of the results of extreme emotion classification and alarm signal recognition.

[0028] Preferably, the location data includes but is not limited to timestamp, signal strength, and location information validity.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] The present invention is based on the personal safety automatic alarm system of emotion analysis, which improves the voice signal detection accuracy under complex background noise through short-time Fourier transform and improved Mel filter group, improves the extreme emotion recognition ability through multi-task network training, and realizes fast and covert safety alarm function through automatic trigger alarm and automatic environmental information collection. Specifically, it can accurately extract emotion and code features in speech through the multi-layer feature extraction and fusion mechanism of deep learning model; introduces attention mechanism and adversarial learning module to improve the recognition performance of the system in complex noise environment; the system can be embedded in smart wearable devices or run as independent hardware to adapt to various application scenarios; and ensures the integrity and transmission security of alarm information through secondary verification and encrypted message sending. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a structural diagram of the composition of the automatic personal safety alarm system based on sentiment analysis of the present invention. DETAILED DESCRIPTION

[0032] The present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.

[0033] like Figure 1 As shown, an automatic personal safety alarm system based on sentiment analysis of the present invention comprises:

[0034] An abnormal speech extraction module 1, the abnormal speech extraction module 1 is used to convert an external input sound signal into a characteristic spectrum that conforms to the hearing law of the human ear;

[0035] Extreme emotion and alarm code recognition module 2, the extreme emotion and alarm code recognition module 2 uses the characteristic spectrum to continuously train the recognition model, the trained recognition model inputs the voice signal, and outputs the judgment probability of extreme emotion and alarm code;

[0036] A personal threat alarm triggering module 3, which comprehensively determines whether to trigger an alarm based on the judgment probability of extreme emotions and alarm codes and the characteristics of the original sound signal;

[0037] An alarm information sending module 4, which verifies the audio data and location information, encapsulates the alarm information and sends it;

[0038] A location information module 5, which provides location data for the alarm information sending module;

[0039] The parameter configuration module 6 provides a configuration interface for the alarm protocol format and the sending parameters.

[0040] The abnormal speech extraction module 1 includes two processes: abnormal speech data processing and feature extraction. The abnormal speech data processing process converts the sound signal into a spectrum sequence indexed by a time window through short-time Fourier transform; the feature extraction process uses a non-uniform Mel filter group with high-frequency resolution amplification to filter the spectrum sequence, and generates a feature spectrum after processing.

[0041] Abnormal voice data processing includes:

[0042] According to the characteristics of the frequency distribution of the sound signal, the signal is converted into the time-frequency domain through the short-time Fourier transform (STFT), and the Hanning window is used to more accurately capture the time-varying characteristics of the sound signal, which helps to identify subtle changes in abnormal speech;

[0043] According to the high-frequency characteristics of abnormal speech signals, the Mel filter bank is improved to simulate the auditory characteristics of the human ear, especially to enhance the mapping bandwidth of high-frequency sound information, so that the generated Mel spectrum can better reflect the perception of abnormal speech, thereby improving the sensitivity of abnormal speech detection;

[0044] Based on the differences in sound recording between different devices and the impact of environmental noise on abnormal speech recognition, normalization is performed to make the feature spectrum have a uniform range, effectively improving the robustness and generalization ability of the system.

[0045] The extreme emotion and alarm code recognition module 2 includes two processes: training and running. The training process includes a multi-task objective function including extreme emotion classification, alarm code recognition and alignment constraints. The emotion classification task and the code detection task are jointly optimized by a weighted method. The running process runs the extreme emotion classification and alarm code recognition model output by the training process. The probability value of extreme emotion and alarm code occurrence is obtained by inputting characteristic spectrum data into the model.

[0046] The multi-task objective function is specifically designed as:

[0047] According to the characteristic that extreme emotions have a low probability distribution in the signal, a weighted focal loss function is designed. The weight of low probability distribution data is amplified by the hyperparameter γ, which can effectively deal with the problem of unbalanced distribution of extreme emotion samples and improve the accuracy of emotion classification, especially the recognition of minority emotion samples.

[0048] According to the characteristics of binary classification of alarm codes, a confidence logarithmic loss function is designed. By optimizing the confidence of code recognition, the model is made more reliable in code recognition and the number of false positives and false negatives is reduced.

[0049] Based on the fact that multi-task networks may have conflicting recognition results, a maximum mean difference feature alignment constraint is designed. By measuring and optimizing the differences in the distribution of emotion and code features, and dynamically adjusting the loss weights of each task through weight hyperparameters, the conflicts in multi-task learning are reduced and the overall performance of the model is improved.

[0050] The personal threat alarm trigger module 3 calculates the judgment value based on the extreme emotion and alarm code probability value combined with the original sound signal characteristics, and decides whether to trigger the alarm according to the judgment value. The personal threat alarm trigger module 3 also needs to compare the judgment value with the threshold, and the threshold used for comparison is configurable. According to the characteristics of false alarms caused by possible sound signal noise interference, the judgment value is calculated by combining the original sound signal quality, thereby improving the accuracy of alarm triggering and avoiding the occurrence of misjudgment. .

[0051] The location information module 5 provides current location data through a geographic information sensor.

[0052] The alarm protocol format provides a standardized alarm protocol format specification including the alarm person, user location information and on-site recording data; the configurable parameters in the sending parameters include but are not limited to emergency contacts and alarm methods, and the alarm methods include but are not limited to mobile phone text messages and voice dialing.

[0053] The alarm information sending module 4 includes four processes: audio data verification, position data verification, alarm information encapsulation and alarm information sending. The audio data verification is used to check the quality and integrity of the original audio signal; the position data verification is used to check the position data output by the position information module; the alarm information encapsulation uses the alarm protocol format output by the parameter configuration module 6 to package the alarm information into a message; the alarm information sending sends the packaged message to the emergency contact configured in the parameter configuration module 6.

[0054] The extreme emotion classification adopts the weighted focal loss method to define the emotion classification loss function; the alarm code recognition adopts the confidence logarithmic loss method to define the code detection loss function; the alignment constraint defines the alignment constraint function according to the maximum mean difference method of the results of extreme emotion classification and alarm code recognition.

[0055] The location data includes but is not limited to timestamp, signal strength, and location information validity.

[0056] Example:

[0057] The specific implementation steps are as follows:

[0058] Step 1: Prepare a training speech dataset containing extreme emotions and coded speech. The dataset covers a variety of speech scenarios, including quiet, noisy, normal communication, extreme emotions, and coded speech. The sample space is no less than 5,000.

[0059] Step 2: Prepare a training server, including a speech dataset for training, an abnormal speech extraction module, and an extreme emotion and alarm code recognition module. The extreme emotion and alarm code recognition module only needs to install the training model part.

[0060] Step 3: Loop to extract sound signals from the speech data set and input them into the abnormal speech extraction module;

[0061] Step 4: The abnormal speech extraction module uses short-time Fourier transform (STFT) to process the sound signal x(t) and convert the sound signal into the time-frequency domain. The specific formula is:

[0062]

[0063] Wherein, w(t) is the window function, using the Hanning window, the window width τ is 20ms, and f is the frequency component;

[0064] Step 5: Set up the Mel filter bank, which contains a total of 40 triangular filters h 1… h 40 , set the high frequency bandwidth to twice the low frequency bandwidth;

[0065] Step 6: Use the Mel filter bank to process the time-frequency domain information, and take the logarithm of the result to generate the Mel spectrum to represent the frequency range that the human ear is sensitive to in speech. The specific formula is:

[0066]

[0067] Where hm(fk) is the frequency response function of the Mel filter, m is the index of the Mel filter bank with a maximum value of 40, K is the number of frequency components, and ε is 0.001 to prevent numerical underflow in logarithmic operations;

[0068] Step 7: Normalize the Mel spectrum to generate a characteristic spectrum, so that the mean of the characteristic value is 0 and the standard deviation is 1, and unify the characteristic range to reduce noise interference and device differences:

[0069]

[0070] Among them, μ and σ are the characteristic mean and standard deviation respectively;

[0071] Step 8: Extreme Emotion and Alarm Signal Recognition Module Using Feature Spectrum The extreme emotion classification neural network is trained with a weighted focal loss function to solve the problem of unbalanced distribution of extreme emotion samples. The specific formula is:

[0072]

[0073] Among them, wi is the sample weight, γ is the hyperparameter for adjusting the difficulty of the sample, which is initially set to 0.5, and π is the emotion classification probability of sample i;

[0074] Step 9: Extreme Emotion and Alarm Signal Recognition Module Using Feature Spectrum Train the neural network for extremely secret code recognition. To improve the confidence of secret code recognition, design the logarithmic confidence loss function. The specific formula is:

[0075]

[0076] Among them, ki is the true label of whether sample i contains the secret code, and σ is the Sigmoid activation function;

[0077] Step 10: The extreme emotion and alarm signal recognition module reduces the conflict between multiple tasks by aligning the distribution of emotion features and signal features. The feature alignment constraint is defined as:

[0078] CT align =MMD(H emotion , H keyword ),

[0079] Among them, MMD stands for Maximum Mean Discrepancy, which is used to measure the difference in feature distribution between the two tasks, Hemotion and Hkeyword are emotion and code feature vectors;

[0080] Step 11: Combine the multi-task loss function with the alignment constraint, update the model parameters, and complete the model training. The specific formula is:

[0081] L target =α1·FL emotion +α2·RL keyword +β·CT align ,

[0082] in:

[0083] FLemotion is a sentiment classification loss based on dynamic sentiment weights;

[0084] RLkeyword is the weighted confidence logarithmic loss for cryptographic recognition;

[0085] CTalign is the alignment constraint between emotion features and secret code features, which is used to optimize the shared features of the two tasks;

[0086] α1, α2, and β are weight hyperparameters, which are dynamically adjusted to balance the loss weights in multi-task learning and are initialized to 1 at the beginning of training;

[0087] Step 12: Prepare a smart watch and pre-install the personal safety automatic alarm system based on sentiment analysis, of which the extreme emotion and alarm code recognition module only needs to install the running model part.

[0088] Step 13: Download the trained extreme emotion classification model and alarm code recognition model to the smart watch. If the model already exists, overwrite it directly;

[0089] Step 14: Repeat steps 1 to 10 to update the model in the smart watch;

[0090] Step 15: The user accesses the parameter configuration module interface to configure the alarm protocol format, and directly uses the built-in protocol format template. The protocol format specification is described as follows:

[0091] |Name of the person who called the police (20 Bytes)|ID card of the person who called the police (18 Bytes)|Location (16 Bytes)|Length of the recording data N (4 Bytes)|Recording data (N Bytes)|,

[0092] Step 16: The user accesses the parameter configuration module interface to configure the sending parameters, such as emergency contact and contact information (mobile phone number), and this information is stored in the smart watch;

[0093] Step 17: Collect user voice signals in real time through the microphone module of the smart watch;

[0094] Step 18: Input the sound signal to the abnormal speech extraction module, execute steps 3 to 6, and obtain the characteristic spectrum

[0095] Step 19: The extreme emotion and alarm signal recognition module will generate the characteristic spectrum The input is given to the extreme emotion classification model to obtain the extreme emotion probability Pemotion, which is then output to the personal threat alarm trigger module;

[0096] Step 20: The extreme emotion and alarm signal recognition module will generate the characteristic spectrum Input to the alarm code recognition model, obtain the code detection probability Pkeyword, and output it to the personal threat alarm trigger module;

[0097] Step 21: The personal threat alarm trigger module obtains the original sound signal and calculates the context feature score, which is calculated comprehensively based on the audio background noise level and the duration of the voice. The specific formula is:

[0098]

[0099] Among them, SNR is signal-to-noise ratio, duration is the duration of speech, which is 20ms, a and b are adjustment parameters, with initial values ​​of a=1; b=10;

[0100] Step 22: The personal threat alarm trigger module uses the probability fusion model to calculate the alarm trigger score Strigger. When Strigger>θ, the alarm is triggered. The specific formula is:

[0101] S trigger =λ1·P emotion +λ2·P keyword +λ3·F context ,

[0102] Among them, λ1, λ2, and λ3 are weight coefficients, and their initial values ​​are set to 1. They need to be fine-tuned according to the effect in the subsequent implementation. The initial value of θ is set to 2.2;

[0103] Step 23: Simulate calling for help in dangerous situations in various scenarios such as at home, outdoors, and at a station;

[0104] Step 24: When an alarm is triggered, the system calls the alarm information sending module;

[0105] Step 25: The alarm information sending module collects real-time sound signals and verifies the signal-to-noise ratio and integrity of the sound signals. If the verification passes, the sound signals are stored as recordings;

[0106] Step 26: The alarm information sending module collects the user's location data, and verifies the timestamp, signal strength and validity of the location data. If the verification passes, the location information is stored;

[0107] Step 27: The alarm information sending module fills in data using the alarm protocol format configured in step 13;

[0108] Step 28: The alarm information sending module uses the contact information configured in step 14 to send the alarm information to the emergency contact.

[0109] Based on the structure and specific implementation process of the present invention, the personal safety automatic alarm system based on emotion analysis of the present invention improves the detection accuracy of speech signals under complex background noise through short-time Fourier transform and improved Mel filter group, improves the ability to recognize extreme emotions through multi-task network training, and realizes fast and hidden safety alarm function through automatic trigger alarm and automatic environmental information collection. By integrating emotion analysis and code recognition technology, the device effectively improves the intelligence level of public safety and is suitable for application deployment in police systems, security companies and urban monitoring networks;

[0110] The present invention can also be used for personal safety protection, and is suitable for wearable devices, smart phones or smart home systems. By real-time monitoring of user emotions and voice codes, the alarm mechanism is automatically triggered to effectively respond to situations that may threaten personal safety, especially to provide safety protection when the user is restricted or unable to actively report an alarm.

[0111] The present invention can also be used for security monitoring in daily environments such as schools, families, and communities. For example, through real-time emotion analysis and specific language code detection, it can prevent school bullying, domestic violence, and community security risks, and implement intelligent and personalized security protection measures. At the same time, the technology of the present invention can also be integrated into existing security products to form a multi-level security monitoring network to meet the needs of multiple scenarios.

Claims

1. A personal safety automatic alarm system based on sentiment analysis, characterized in that: include: Abnormal speech extraction module, used to convert external input sound signals into characteristic spectrum that conforms to the laws of human hearing; The extreme emotion and alarm code recognition module uses the characteristic spectrum to continuously train the recognition model. The trained recognition model inputs the voice signal and outputs the judgment probability of extreme emotions and alarm codes. The personal threat alarm trigger module comprehensively determines whether to trigger the alarm based on the probability of extreme emotions and alarm codes, combined with the characteristics of the original sound signal; The alarm information sending module verifies the audio data and location information, encapsulates the alarm information and sends it; The location information module provides location data for the alarm information sending module; The parameter configuration module provides the configuration interface for the alarm protocol format and sending parameters.

2. The personal safety automatic alarm system based on sentiment analysis as claimed in claim 1, characterized in that: The abnormal speech extraction module includes three processes: abnormal speech data processing, feature extraction and normalization. The abnormal speech data processing process converts the sound signal into a spectrum sequence indexed by the time window through short-time Fourier transform. The three processes are specifically as follows: 1) According to the characteristics of the frequency distribution of the sound signal, the signal is converted into the time-frequency domain through short-time Fourier transform plus Hanning window; 2) According to the high frequency characteristics of abnormal speech signals, a non-uniformly distributed Mel filter bank is set by increasing the mapping bandwidth of high frequency sound information; 3) A characteristic spectrum with a uniform range is obtained through normalization processing.

3. The personal safety automatic alarm system based on sentiment analysis as claimed in claim 1, characterized in that: The extreme emotion and alarm sign recognition module includes two processes: training and running. The training process includes a multi-task objective function containing extreme emotion classification, alarm sign recognition and alignment constraints, which is used to train and output aligned extreme emotion classification and alarm sign recognition models. The running process runs the extreme emotion classification and alarm sign recognition model output by the training process, and obtains the probability value of the occurrence of extreme emotions and alarm signs by inputting characteristic spectrum data into the model.

4. The personal safety automatic alarm system based on sentiment analysis as claimed in claim 1, characterized in that: The personal threat alarm trigger module calculates a judgment value based on the probability value of extreme emotions and alarm codes combined with the noise characteristics of the original sound signal, and decides whether to trigger an alarm based on the judgment value. The personal threat alarm trigger module also requires that the judgment value be compared with a threshold, and the threshold used for comparison is configurable.

5. The personal safety automatic alarm system based on sentiment analysis as claimed in claim 1, characterized in that: The location information module provides current location data through a geographic information sensor.

6. The personal safety automatic alarm system based on sentiment analysis as claimed in claim 1, characterized in that: The alarm protocol format provides a standardized alarm protocol format specification including the alarm person, user location information and on-site recording data; the configurable parameters in the sending parameters include but are not limited to emergency contacts and alarm methods, and the alarm methods include but are not limited to mobile phone text messages and voice dialing.

7. The personal safety automatic alarm system based on sentiment analysis as claimed in claim 6, characterized in that: The alarm information sending module includes four processes: audio data verification, position data verification, alarm information encapsulation and alarm information sending. The audio data verification is used to check the quality and integrity of the original audio signal; the position data verification is used to check the position data output by the position information module; the alarm information encapsulation uses the alarm protocol format output by the parameter configuration module to package the alarm information into a message; the alarm information sending sends the packaged message to the emergency contact configured in the parameter configuration module.

8. The personal safety automatic alarm system based on sentiment analysis as claimed in claim 3, characterized in that: The multi-task objective function is specifically as follows: amplifying the weight of the distribution of extreme emotional signals through a weighted focal loss function and hyperparameters; optimizing the confidence of binary classification recognition of alarm codes through a confidence logarithmic loss function; aligning the conflicts of multi-task learning through maximum mean difference feature alignment constraints and dynamically adjusting loss weights.

9. The personal safety automatic alarm system based on sentiment analysis as claimed in claim 7, characterized in that: The location data includes but is not limited to timestamp, signal strength, and location information validity.

Images