A method and system for early warning of risk of non-suicidal self-injury in adolescents

Abstract

The application discloses a kind of early warning method and system of adolescent non-suicidal self-injury risk, first acquisition adolescent voice information and carry out information preprocessing, subsequently by three independent modules, feature extraction is carried out to the voice signal after preprocessing, first module extracts emotional tendency feature, second module extracts intonation feature, third module extracts speech rate feature.Next, based on the information extracted by three modules, use artificial intelligence algorithm, predict adolescent non-suicidal self-injury behavior probability.And non-suicidal self-injury probability is classified, and relevant intervention measures are given.The method and system of the application, by time-sequenced voice information acquisition, multidimensional feature hierarchical extraction, feature correlation modeling driven by self-attention mechanism, combined with dialogue object classification analysis, can effectively capture the potential signal of adolescent non-suicidal self-injury behavior, significantly improve the accuracy and timeliness of risk early warning, and provide reliable basis for early intervention.

Description

Technical Field

[0001] This invention relates to the technical field of mental health monitoring, and in particular to a method and system for early warning of non-suicidal self-harm risks in adolescents. Background Technology

[0002] Nonsuicidal self-injury (NSSI) is one of the most common mental health and behavioral problems among adolescents. As maladaptive behaviors (such as anxiety, depression, interpersonal difficulties, and academic struggles), NSSI is not socially accepted and not only harms the physical and mental health of adolescents but also increases the risk of suicide. The prevalence and harmfulness of NSSI have made it an increasingly urgent public health issue, attracting increasing attention from research.

[0003] Given the highly concealed nature of non-suicidal self-harm, many cases of NSSI occurring in adolescents are difficult to detect, hindering early identification and intervention. Since NSSI exhibits core linguistic characteristics that significantly contrast with everyday states—such as angry / irritable language signals, depressed / low-spirited language signals, and anxious / tense language signals—this invention monitors negative emotions in adolescents through their non-suicidal self-harm language characteristics. This includes multimodal information such as emotional tendency, tone of voice, and speech rate, enabling graded early warning and intervention for non-suicidal self-harm behaviors in adolescents. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for early warning of non-suicidal self-harm risks in adolescents, which can effectively predict non-suicidal self-harm behaviors in adolescents and thus intervene in advance to reduce the adverse consequences caused by them.

[0005] The technical solution of this invention is: A method for early warning of non-suicidal self-harm risks in adolescents includes the following steps: S1. Collect the voice information of teenagers within a set period, convert the voice information into text, and then classify and statistically analyze it according to the dialogue object. The dialogue object includes six categories: talking to oneself, talking to parents, talking to teachers, talking to friends, talking to relatives, and talking to strangers. S2. The collected adolescent speech information is preprocessed and finally segmented into short frames of fixed length for subsequent feature extraction. S3. Multi-dimensional feature extraction is performed on the collected adolescent speech information. Three modules are used: the first module extracts the emotional features in the speech through the emotion tendency extraction module, and the emotional features include positive, negative and neutral. The second module extracts the fundamental frequency and its dynamic change features through the intonation extraction module. The third module calculates the number of syllables or the effective speech ratio per unit time through the speech rate extraction module to form a multi-dimensional speech feature set. S4. Input the multi-dimensional speech features of emotion, tone, and speech rate extracted in step S3 into the self-attention mechanism neural network model. By focusing on the correlation between features through the self-attention mechanism, adaptive weight allocation and deep feature fusion are performed on the input features. Then, through inference calculation, the probability of non-suicidal self-harm behavior in adolescents is output, and finally, accurate early warning of non-suicidal self-harm risk in adolescents is achieved. S5. Provide corresponding intervention plans based on the warning risk level.

[0006] Preferably, in step S1, a portable miniature sensor is used to collect voice information. The portable miniature sensor is suitable for daily wear scenarios for teenagers, so as to achieve continuous and stable collection of voice information.

[0007] Preferably, in step S2, when preprocessing the extracted speech signal, firstly, blank segments in the signal are removed by endpoint detection, then a noise reduction algorithm is used to suppress environmental noise interference, and finally, speech separation technology is used to separate the speech components of other people. Finally, the processed clean speech continuous signal is divided into short frames of fixed length for subsequent feature extraction.

[0008] Preferably, in step S3, the sentiment tendency extraction module of the first module uses a convolutional neural network (CNN) or a long short-term memory network (LSTM) as the core algorithm architecture to perform feature learning and classification on the preprocessed speech data, and finally outputs the sentiment tendency results in the speech of teenagers, providing core input features of the sentiment dimension for the subsequent risk warning model.

[0009] Preferably, in step S3, the intonation extraction module of the second module uses preset gender classification logic to perform targeted processing on the speech signals of different genders by adopting extraction parameters adapted to their fundamental frequency range, so as to improve the accuracy and stability of intonation feature extraction and provide more reliable intonation dimension features for subsequent risk warning models.

[0010] Preferably, in step S3, the speech rate extraction module of the third module distinguishes between male and female speech, uses the ratio of the number of pronunciation units in the effective speech segment to the effective duration as the core logic for speech rate quantification, and outputs a value reflecting the speech rate characteristics of teenagers, providing input data for the speech rate dimension of the subsequent risk warning model.

[0011] Preferably, in step S4, the self-attention mechanism neural network model learns the correlation between features, automatically assigns differentiated weights to the three-dimensional features, and outputs the risk probability of non-suicidal self-harm behavior in adolescents after deep feature fusion and inference calculation, and outputs the source of negative emotions, thereby realizing early warning driven by multi-dimensional features.

[0012] Preferably, in step S5, the probability of non-suicidal self-harm among adolescents is divided into four levels, namely: A probability of less than 40% indicates almost no risk and no risk of self-harm. A probability of 40-60% indicates a risk, with the possibility of self-harm. A probability of 60-80% indicates a relatively serious condition, with self-harming behavior. A probability of 80-100% indicates a very serious condition; the self-harm behavior is severe and requires immediate medical attention.

[0013] Preferably, in step S5, based on the analysis results of risk level and source of negative emotions, online or offline intervention is provided, wherein online intervention involves setting relevant psychological intervention strategies within the system.

[0014] A risk warning system for non-suicidal self-harm in adolescents includes: The information acquisition module collects the voice information of teenagers within a set period through a sound-to-electricity conversion sensor and converts the voice information into text. The dialogue object classification module categorizes the collected voice information according to the dialogue object. The dialogue objects include six categories: talking to oneself, talking to parents, talking to teachers, talking to friends, talking to relatives, and talking to strangers. The signal preprocessing module removes blank segments from the signal, reduces signal noise, separates other people's speech, and finally divides the processed clean speech continuous signal into short frames of fixed length. A multimodal speech signal analysis module, connected to the signal preprocessing module, is configured to extract multi-dimensional features from the collected adolescent speech information, specifically including: The sentiment extraction submodule is used to extract sentiment features from speech information, including positive sentiment, negative sentiment, neutral sentiment, etc. The intonation extraction submodule is used to extract the fundamental frequency of speech and its dynamic variation features; The speech rate extraction submodule is used to calculate the number of syllables or the percentage of effective speech per unit time to characterize speech rate features; The multimodal speech signal analysis module integrates the above-mentioned emotional tendency, intonation, and speech rate features to form a multi-dimensional speech feature set; The risk prediction module, connected to the multimodal speech signal analysis module, is configured as follows: The system receives the multi-dimensional speech features as input; it processes the input features using a self-attention mechanism neural network model, focusing on the correlation between features through the self-attention mechanism to achieve adaptive weight allocation and deep feature fusion of the input features; and through model inference calculation, it outputs the risk probability of adolescents exhibiting non-suicidal self-harm behavior, thus completing accurate risk warning.

[0015] An intervention plan generation module, connected to the risk prediction module, is configured to automatically match and generate corresponding intervention plans based on the early warning risk level output by the risk prediction module.

[0016] A computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the aforementioned method for early warning of non-suicidal self-harm risks in adolescents. Beneficial effects

[0017] The method and system of this invention, through time-sequential speech information acquisition, multi-dimensional feature hierarchical extraction, and feature association modeling driven by a self-attention mechanism, combined with dialogue object classification analysis, can effectively capture potential signals of non-suicidal self-harm behaviors in adolescents, significantly improving the accuracy and timeliness of risk warnings and providing a reliable basis for early intervention. Specifically, it includes the following advantages: 1. This invention uses voice information within a set period as input data to predict the risk of non-suicidal self-harm behavior in adolescents. Considering the limitations of voice information collected from adolescents within a short period (it is difficult to fully reflect the true state) and its insufficient representativeness, this invention avoids the limitations of a single period by collecting voice data from multiple periods (such as one day, one week, one month, or one year), thereby improving the accuracy of the prediction results.

[0018] 2. This invention fuses three dimensions of features in adolescent speech information: emotional features, intonation features, and speech rate features. By adding these information dimensions, the predictive model more closely reflects the practical experience of professionals or institutions in assessing non-suicidal self-harm behaviors.

[0019] 3. The self-attention mechanism used in the prediction of non-suicidal self-harm behavior in adolescents in this invention greatly improves the accuracy of the model in assessing non-suicidal self-harm behavior. Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the method for early warning of non-suicidal self-harm risks in adolescents according to the present invention; Figure 2 A flowchart for multimodal speech signal analysis; Figure 3 A structural diagram of a risk warning system for non-suicidal self-harm behaviors in adolescents.

[0021] Figure 3 The module consists of: 1. Information acquisition module; 2. Dialogue object classification module; 3. Signal preprocessing module; 4. Multimodal speech signal analysis module; 4a. Sentiment extraction submodule; 4b. Intonation extraction submodule; 4c. Speech rate extraction submodule; 5. Risk prediction module; and 6. Intervention plan generation module. Detailed Implementation Example

[0022] like Figure 1 As shown, this embodiment provides a method for early warning of non-suicidal self-harm behavior risks in adolescents, including the following steps: S1. Collect the voice information of teenagers within a set period, convert the voice information into text, and then classify and statistically analyze it according to the dialogue object. The dialogue object includes six categories: talking to oneself, talking to parents, talking to teachers, talking to friends, talking to relatives, and talking to strangers. It uses a portable miniature sensor to collect voice information, which is suitable for teenagers to wear in daily life, and can achieve continuous and stable collection of voice information.

[0023] Set a period of one day, one week, one month, or one year. By comparing the changes every day, week, month, and year, the prediction results become more reliable, and the changes in emotions over time are captured. Pack a bag each day, record a set of evaluation results, form a time series chart, and finally summarize the data over a period of time for a comprehensive evaluation.

[0024] S2. The collected adolescent speech information is preprocessed and finally segmented into short frames of fixed length for subsequent feature extraction. When preprocessing the extracted speech signal, firstly, blank segments in the signal are removed by endpoint detection, then noise reduction algorithm is used to suppress environmental noise interference, and finally, speech separation technology is used to separate the speech components of other people. Finally, the processed clean speech continuous signal is divided into short frames of fixed length for subsequent feature extraction.

[0025] S3. Multi-dimensional feature extraction is performed on the collected adolescent speech information. Three modules are used: the first module extracts the emotional features in the speech through the emotion tendency extraction module, which includes positive, negative and neutral emotions; the second module extracts the fundamental frequency and its dynamic change features through the intonation extraction module; and the third module calculates the number of syllables or the proportion of effective speech per unit time through the speech rate extraction module, forming a multi-dimensional speech feature set. The sentiment extraction module of the first module uses a convolutional neural network (CNN) or a long short-term memory network (LSTM) as the core algorithm architecture to learn features and classify preprocessed speech data, and finally output the sentiment results of adolescent speech, providing core input features of the sentiment dimension for subsequent risk warning models.

[0026] The second module, the intonation extraction module, is responsible for intonation extraction. The core of intonation extraction is to obtain the fundamental frequency of the speech signal. Given that there are significant differences in the fundamental frequency distribution range between males and females in the adolescent group (males are usually 80-250Hz, and females are 160-500Hz), this module uses preset gender classification logic to apply extraction parameters adapted to the fundamental frequency range of the speech signals of different genders for targeted processing, so as to improve the accuracy and stability of intonation feature extraction and provide more reliable intonation dimension features for subsequent risk warning models.

[0027] The speech rate extraction module of the third module is responsible for speech rate extraction. Given the difference in speech rate between men and women, it distinguishes between male and female speech and uses the ratio of the number of pronunciation units in the effective speech segment to the effective duration as the core logic for speech rate quantification. It outputs a value that reflects the speech rate characteristics of teenagers and provides input data for the speech rate dimension of the subsequent risk warning model.

[0028] S4, such as Figure 2 As shown, the multi-dimensional speech features of emotion, tone, and speech rate extracted in step S3 are input into the self-attention mechanism neural network model. The self-attention mechanism focuses on the relationship between features, such as the coupling between negative emotions and slowed speech rate. The input features are adaptively weighted and fused with deep features. Then, the probability of non-suicidal self-harm behavior in adolescents is output through inference calculation, and finally, the accurate early warning of non-suicidal self-harm risk in adolescents is achieved. The self-attention mechanism neural network model learns the correlation between features and automatically assigns differentiated weights to the three-dimensional features. After deep feature fusion and inference calculation, it outputs the risk probability of non-suicidal self-harm behavior in adolescents and outputs the source of negative emotions, such as talking to oneself or communicating with others, so as to achieve early warning driven by multi-dimensional features.

[0029] S5. Based on the warning risk level, provide corresponding intervention plans. The probability of non-suicidal self-harm among adolescents is divided into four levels: A probability of less than 40% indicates almost no risk and no risk of self-harm. A probability of 40-60% indicates a risk, with the possibility of self-harm. A probability of 60-80% indicates a relatively serious condition, with self-harming behavior. A probability of 80-100% indicates a very serious condition; the self-harm behavior is severe and requires immediate medical attention.

[0030] Based on the analysis results of risk level and sources of negative emotions, online or offline interventions are provided, with online interventions involving the setting of relevant psychological intervention strategies within the system. Example

[0031] like Figure 3As shown, this embodiment provides a risk warning system for non-suicidal self-harm behavior in adolescents, which includes the following modules.

[0032] The system comprises six modules: information acquisition module 1, dialogue object classification module 2, signal preprocessing module 3, multimodal speech signal analysis module 4, risk prediction module 5, and intervention plan generation module 6. Through collaborative division of labor and efficient linkage, these modules achieve the goal of intelligently generating risk warnings and intervention plans. The specific process is as follows: The acoustic-to-electric conversion sensor used in information acquisition module 1 must meet the design requirements of being small and lightweight to fit the form factor limitations of wearable devices such as wristbands, while also possessing the function of continuously acquiring voice information within the device. Selection requirements: size: diameter ≤ 6mm, thickness ≤ 3mm (to fit the confined space of a wristband); power consumption: operating current ≤ 1mA (to avoid frequent charging and support continuous acquisition); sensitivity: -38dBV~-42dBV (to ensure clear pickup of voice signals at close range (30-50cm), such as a microelectromechanical system (MEMS) microphone.

[0033] When categorizing conversational participants, voice information is collected at a 10kHz sampling rate, with four sampling periods: one day, one week, one month, and one year. Daily voice samples are packaged into a single audio file named "Teen's Name - Year, Month, Day," and stored in the wristband.

[0034] Dialogue object classification module 2 performs speech-to-text processing on daily audio files: It adopts the Whisper-large-v2 model based on the Transformer encoder-decoder architecture to convert speech into text sequences. At the same time, it identifies dialogue objects through voiceprint recognition technology (such as a CNN-based voiceprint classification model). The voiceprint feature database of teenagers, parents, teachers, friends, relatives, and strangers is pre-entered. By calculating the cosine similarity between the real-time voiceprint and the features in the database (with a threshold set to 0.85), the speech is classified into 6 categories: "talking to oneself" (no other voiceprint matching), "talking to parents", "talking to teachers", "talking to friends", "talking to relatives", and "talking to strangers". Basic data such as daily speech duration and dialogue frequency are separately counted for each category and stored in a relational database (such as MySQL).

[0035] Signal preprocessing module 3 performs refined preprocessing on the acquired original speech signal according to the progressive process of "noise suppression - interference separation - frame processing" to provide high-quality data support for subsequent feature extraction. (1) Endpoint detection and blank segment removal: The dual threshold method (based on the energy and zero-crossing rate characteristics of the speech signal) is used to detect the endpoints, identify and remove silent blank segments in the speech signal (including background blanks without speech and equipment idle noise segments), retain the speech intervals with effective semantics, and reduce the interference of invalid data on subsequent processing; (2) Environmental noise adaptive suppression: The spectrum subtraction combined with the wavelet threshold noise reduction algorithm is used. First, pure noise samples (environmental noise in blank segments) are extracted through the speech activity detection (VAD) module to construct a noise spectrum model. Then, the spectrum subtraction operation is performed on the effective speech interval to suppress steady-state noise (such as equipment operation sound and background human voice) and non-steady-state noise (such as sudden interference sound) in the environment. At the same time, the detailed features of the speech signal are retained through wavelet threshold processing. (e.g., intonation changes) to ensure the integrity of speech information after noise reduction; (3) Multi-speaker speech separation: Conv-TasNet speech separation model is introduced, the time-frequency features of mixed speech are extracted by the encoder, and the differences in speech features of different speakers (such as voiceprint, intonation, speech rate) are distinguished by the attention mechanism, and a dedicated speech mask is generated to separate the speech components of the target teenager from the speech components of other speakers, so as to achieve accurate extraction of single-target speech; (4) Speech framing processing: Hanning window function is used to perform framing operation on the separated pure speech continuous signal, and the continuous speech is divided into short frame sequences of fixed length, which not only ensures the stability of the speech signal within a single frame, but also retains the temporal correlation features between frames through frame overlap, laying the foundation for subsequent extraction of multi-dimensional features such as emotional tendency, intonation, and speech rate.

[0036] Multimodal speech signal analysis module 4 receives the audio file and classification results output by the information acquisition and classification module, and performs multi-dimensional feature extraction: The sentiment extraction submodule 4a uses a pre-trained speech sentiment recognition model (such as the SER model based on LSTM) as input, inputs pre-processed speech information, outputs a sentiment sentiment probability distribution (positive, negative, neutral, etc.), takes the category with the highest probability as the sentiment label of the speech segment, and counts the proportion of negative sentiment in each category scenario (such as conversations with parents) every day. Intonation extraction submodule 4b: Extracts the fundamental frequency (F0) sequence of audio using the Praat speech analysis tool, calculates the mean, standard deviation, maximum, minimum and dynamic rate of change of the fundamental frequency (such as the difference of F0 within adjacent 500ms), and characterizes the stability and fluctuation features of the intonation; Speech rate extraction submodule 4c: Combines speech information to count the number of effective syllables per unit time (1 minute) (syllables after removing pauses and interjections), and calculates the proportion of effective speech duration to the total collection time (effective speech refers to speech segments containing semantics, identified by the speech activity detection VAD model), as speech rate features; the above features are integrated into a multi-dimensional feature vector (such as [proportion of negative emotions, fundamental frequency standard deviation, number of effective syllables per minute, and dialogue object category encoding]) for subsequent model input.

[0037] Risk prediction module 5 employs a neural network model based on a self-attention mechanism (such as a Transformer encoder). Its specific structure includes an input layer, an embedding layer, four self-attention encoding layers, and an output layer. The input layer receives multi-dimensional feature vectors; the embedding layer converts the classification features (discrete values) of the dialogue object into 256-dimensional embedding vectors, which are then concatenated with continuous features such as emotion, tone, and speech rate (after standardization) to form a 512-dimensional feature matrix. The self-attention encoding layer calculates the association weights between features (such as the coupling coefficient between "negative emotions in conversations with strangers" and "slowed speech rate") through a multi-head attention mechanism (8 heads), adaptively strengthens the influence of key features (such as increasing the weight of the negative emotion + slow speech rate combination by 20%), and stabilizes the training through residual connections and Layer Normalization. The output layer uses the Sigmoid activation function to output a risk probability value between 0 and 1 (0 represents no risk, and 1 represents extremely high risk), and is divided into 4 risk levels according to the probability value (below 40% is almost no risk, 40-60% is risky, 60-80% is relatively serious, and 80-100% is very serious). Model training phase: A voice dataset containing 5,000 teenagers (12-18 years old) (including labels of non-suicidal self-harm behavior) was used. The training set and test set were divided in a 7:3 ratio. The model parameters were optimized using the cross-entropy loss function. The final AUC value of the test set was ≥0.92 to ensure prediction accuracy.

[0038] Intervention plan generation module 6 has a built-in risk level-intervention plan mapping library, which outputs corresponding plans for different risk levels. Almost no risk (below 40%): Generate a "daily attention plan" that recommends parents have an in-depth conversation with their teenager once a week and record their emotional changes; Risk (40%-60%): Generates a "Family Guidance Plan" which includes 3 sets of parent-child interactive activities (such as joint exercise and interest cultivation) and tips for emotional management; More serious cases (60%-80%): Develop a "school-family collaboration plan", notify the homeroom teacher to pay attention to the teenager's performance at school, and work with the school counselor or mental health promotion center to conduct counseling once every 1-2 weeks to dynamically assess their physical and mental condition.

[0039] Very serious (80%-100%): Initiate the "emergency intervention process", simultaneously notify parents, schools and mental health promotion centers, coordinate professional assessment and intervention within 24 hours, and seek medical attention promptly.

[0040] This embodiment achieves full automation of the entire process from voice information collection, multi-dimensional feature extraction, intelligent risk prediction to precise intervention plan generation through the collaborative work of the above modules. It can effectively capture potential signals of non-suicidal self-harm behavior in adolescents and provide a scientific basis for early intervention.

[0041] This invention also provides a terminal, the terminal comprising: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the adolescent non-suicidal self-harm risk warning method as described in any of the above embodiments.

[0042] This invention also provides a computer-readable storage medium storing a computer program thereon, characterized in that, when executed by a processor, the program implements the adolescent non-suicidal self-harm risk warning method as described in any of the above embodiments.

[0043] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All modifications made according to the spirit and essence of the main technical solution of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for early warning of non-suicidal self-harm risks in adolescents, characterized in that, Includes the following steps: S1. Collect the voice information of teenagers within a set period, convert the voice information into text, and then classify and statistically analyze it according to the dialogue object. The dialogue object includes six categories: talking to oneself, talking to parents, talking to teachers, talking to friends, talking to relatives, and talking to strangers. S2. The collected adolescent speech information is preprocessed and finally segmented into short frames of fixed length for subsequent feature extraction. S3. Multi-dimensional feature extraction is performed on the collected adolescent speech information. Three modules are used: the first module extracts the emotional features in the speech through the emotion tendency extraction module, and the emotional features include positive, negative and neutral. The second module extracts the fundamental frequency and its dynamic change features through the intonation extraction module. The third module calculates the number of syllables or the effective speech ratio per unit time through the speech rate extraction module to form a multi-dimensional speech feature set. S4. Input the multi-dimensional speech features of emotion, tone, and speech rate extracted in step S3 into the self-attention mechanism neural network model. By focusing on the correlation between features through the self-attention mechanism, adaptive weight allocation and deep feature fusion are performed on the input features. Then, through inference calculation, the probability of non-suicidal self-harm behavior in adolescents is output, and finally, accurate early warning of non-suicidal self-harm risk in adolescents is achieved. In step S4, the self-attention mechanism neural network model automatically assigns differentiated weights to the three-dimensional features by learning the correlation between features; After deep feature fusion and inference calculation, the system outputs the risk probability of non-suicidal self-harm behavior in adolescents and the source of negative emotions, thus achieving early warning driven by multi-dimensional features. The specific structure of the self-attention mechanism neural network model includes an input layer, an embedding layer, four self-attention encoding layers, and an output layer: The input layer receives multi-dimensional feature vectors; the embedding layer converts the dialogue object classification features into 256-dimensional embedding vectors, which are then concatenated with continuous features of emotion, tone, and speech rate to form a 512-dimensional feature matrix. The self-attention encoding layer calculates the correlation weights between features through a multi-head attention mechanism, adaptively strengthens the influence of key features, and stabilizes training through residual connections and LayerNormalization. The output layer uses the Sigmoid activation function to output a risk probability value between 0 and 1, and divides it into 4 risk levels according to the probability value; S5. Provide corresponding intervention plans based on the warning risk level; In step S5, the probability of non-suicidal self-harm among adolescents is divided into four levels: A probability of less than 40% indicates almost no risk and no risk of self-harm. A probability of 40-60% indicates a risk, with the possibility of self-harm. A probability of 60-80% indicates a relatively serious condition, with self-harming behavior. A probability of 80-100% indicates a very serious condition; the self-harm behavior is severe and requires immediate medical attention. Based on the analysis results of risk level and sources of negative emotions, online or offline interventions are provided, with online interventions involving the setting of relevant psychological intervention strategies within the system.

2. The method for early warning of non-suicidal self-harm risks in adolescents according to claim 1, characterized in that, In step S1, a portable miniature sensor is used to collect voice information. The portable miniature sensor is suitable for daily wear scenarios of teenagers to achieve continuous and stable collection of voice information.

3. The method for early warning of non-suicidal self-harm risks in adolescents according to claim 1, characterized in that, In step S2, when preprocessing the extracted speech signal, firstly, blank segments in the signal are removed by endpoint detection, then a noise reduction algorithm is used to suppress environmental noise interference, and finally, speech separation technology is used to separate the speech components of other people. Finally, the processed clean speech continuous signal is divided into short frames of fixed length for subsequent feature extraction.

4. The method for early warning of non-suicidal self-harm risks in adolescents according to claim 1, characterized in that, In step S3, the sentiment tendency extraction module of the first module uses a convolutional neural network or a long short-term memory network as the core algorithm architecture to perform feature learning and classification on the preprocessed speech data, and finally outputs the sentiment tendency results in the speech of teenagers, providing core input features of the sentiment dimension for the subsequent risk warning model.

5. The method for early warning of non-suicidal self-harm risks in adolescents according to claim 1, characterized in that, In step S3, the intonation extraction module of the second module uses preset gender classification logic to perform targeted processing on the speech signals of different genders by adopting extraction parameters adapted to their fundamental frequency range, so as to improve the accuracy and stability of intonation feature extraction and provide more reliable intonation dimension features for subsequent risk warning models.

6. The method for early warning of non-suicidal self-harm risks in adolescents according to claim 1, characterized in that, In step S3, the speech rate extraction module of the third module distinguishes between male and female speech, uses the ratio of the number of pronunciation units in the effective speech segment to the effective duration as the core logic for speech rate quantification, and outputs a value that reflects the speech rate characteristics of teenagers, providing input data for the speech rate dimension of the subsequent risk warning model.

7. A risk warning system for non-suicidal self-harm among adolescents, characterized in that, include: The information acquisition module (1) acquires the speech information of teenagers within a set period through the sound-to-electric conversion sensor and converts the speech information into text; The dialogue object classification module (2) classifies the collected voice information according to the dialogue object. The dialogue objects include six categories: talking to oneself, talking to parents, talking to teachers, talking to friends, talking to relatives and talking to strangers. The signal preprocessing module (3) removes blank segments from the signal, reduces signal noise, separates other people's speech, and finally divides the processed pure speech continuous signal into short frames of fixed length. The multimodal speech signal analysis module (4), connected to the signal preprocessing module (3), is configured to perform multi-dimensional feature extraction on the collected adolescent speech information, specifically including: The sentiment extraction submodule (4a) is used to extract sentiment features from speech information, including positive sentiment, negative sentiment and neutral sentiment; The intonation extraction submodule (4b) is used to extract the fundamental frequency of speech and its dynamic variation features; The speech rate extraction submodule (4c) is used to calculate the number of syllables or the percentage of effective speech per unit time to characterize speech rate features; The multimodal speech signal analysis module (4) integrates the above-mentioned emotional tendency, intonation, and speech rate features to form a multi-dimensional speech feature set; The risk prediction module (5), connected to the multimodal speech signal analysis module (4), is configured as follows: The system receives the multi-dimensional voice features as input; it uses a self-attention mechanism neural network model to process the input features, focusing on the correlation between features through the self-attention mechanism to achieve adaptive weight allocation and deep feature fusion of the input features; through model inference calculation, it outputs the risk probability of adolescents exhibiting non-suicidal self-harm behavior, thus completing accurate risk warning; The self-attention mechanism neural network model learns the correlation between features and automatically assigns differentiated weights to the three-dimensional features. After deep feature fusion and inference calculation, it outputs the risk probability of non-suicidal self-harm behavior in adolescents and outputs the source of negative emotions, realizing early warning driven by multi-dimensional features. The specific structure of the self-attention mechanism neural network model includes an input layer, an embedding layer, four self-attention encoding layers, and an output layer: The input layer receives multi-dimensional feature vectors; the embedding layer converts the dialogue object classification features into 256-dimensional embedding vectors, which are then concatenated with continuous features of emotion, tone, and speech rate to form a 512-dimensional feature matrix. The self-attention encoding layer calculates the correlation weights between features through a multi-head attention mechanism, adaptively strengthens the influence of key features, and stabilizes training through residual connections and LayerNormalization. The output layer uses the Sigmoid activation function to output a risk probability value between 0 and 1, and divides it into 4 risk levels according to the probability value; The intervention plan generation module (6) is connected to the risk prediction module (5) and is configured to automatically match and generate the corresponding intervention plan based on the early warning risk level output by the risk prediction module (5). The probability of non-suicidal self-harm among adolescents is divided into four levels: A probability of less than 40% indicates almost no risk and no risk of self-harm. A probability of 40-60% indicates a risk, with the possibility of self-harm. A probability of 60-80% indicates a relatively serious condition, with self-harming behavior. A probability of 80-100% indicates a very serious condition; the self-harm behavior is severe and requires immediate medical attention. Based on the analysis results of risk level and sources of negative emotions, online or offline interventions are provided, with online interventions involving the setting of relevant psychological intervention strategies within the system.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the method for early warning of non-suicidal self-harm risks in adolescents as described in any one of claims 1-6.

Images