Artificial intelligence-based mental health monitoring method and system

Abstract

The application relates to the technical field of data processing, and discloses a mental health monitoring method and system based on artificial intelligence. The method comprises the following steps: collecting environment parameters through a multi-source sensing network, identifying scene features through an environment context reasoning engine; extracting physiological features by using an adaptive algorithm according to the environment features; analyzing the relationship between the environment and the behavior by using a Bayesian model; fusing multi-dimensional features through a graph attention network; constructing an environment adaptive baseline based on transfer learning; and combining multi-level anomaly detection to evaluate the psychological state and output an early warning. The application can realize accurate and real-time monitoring of the mental health state of a user under the condition of considering dynamic changes of the environment, and overcome the monitoring deviation caused by individual differences.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing, and particularly relates to a mental health monitoring method and system based on artificial intelligence. BACKGROUND

[0002] In the existing mental health monitoring technology, mainly adopts questionnaire evaluation, interview diagnosis and physiological index monitoring method. Questionnaire evaluation collects subjective feeling data through standardized scale; interview diagnosis relies on professional personnel to carry out state evaluation; physiological index monitoring adopts wearable device to collect heart rate, brain electricity, skin electricity and other physiological data. These methods analyze the data of single dimension or multiple dimensions to evaluate the mental health status of individuals. At the same time, part of the monitoring system begins to introduce artificial intelligence technology, uses machine learning algorithm to process and analyze the monitoring data, and establishes a mental state evaluation model.

[0003] However, the existing technology still has the following deficiencies: the traditional method adopts static evaluation mode, which is difficult to realize real-time dynamic monitoring of mental state; secondly, the existing monitoring system generally ignores the influence of environmental factors on mental state, and cannot accurately identify and compensate the interference brought by environmental changes; thirdly, due to significant individual differences, the general evaluation standard is difficult to adapt to the characteristics of different users, and evaluation deviation is easy to occur; fourthly, the existing technology lacks effective fusion mechanism of multi-source heterogeneous data, and it is difficult to comprehensively utilize multi-dimensional data such as environment, physiology and behavior for comprehensive analysis. SUMMARY

[0004] The present application provides a mental health monitoring method and system based on artificial intelligence, which is used to realize accurate real-time monitoring of user mental health status considering the dynamic changes of environment, and overcome the monitoring deviation caused by individual differences.

[0005] In a first aspect, the present application provides an artificial intelligence-based mental health monitoring method, comprising: collecting environmental parameters in real time through a multi-source sensing network, performing scene recognition processing through an environmental context reasoning engine to obtain environmental semantic feature data; performing feature extraction and individual difference modeling on multi-dimensional physiological signals through an environmental adaptive signal processing algorithm according to the environmental semantic feature data, to obtain an environmental-compensated personalized physiological feature vector; performing environmental-behavior interaction analysis through a Bayesian causal inference model using the environmental semantic feature data and the personalized physiological feature vector, to output environmental-behavior correlation features; inputting the environmental-behavior correlation features and the personalized physiological feature vector into a graph attention network, performing multi-dimensional feature fusion through an environmental condition gating mechanism, to generate dynamic feature representations; constructing sub-baselines for different environmental types through a hierarchical baseline transfer learning module according to the dynamic feature representations, to obtain an environmental adaptive baseline model; and evaluating a mental state through a multi-level anomaly detector based on the environmental adaptive baseline model, and outputting an early warning signal through an environmental constraint anomaly scoring mechanism.

[0006] In a second aspect, the present application provides an artificial intelligence-based mental health monitoring system, comprising:

[0007] A collection module for collecting environmental parameters in real time through a multi-source sensing network, performing scene recognition processing through an environmental context reasoning engine, and obtaining environmental semantic feature data;

[0008] A modeling module for performing feature extraction and individual difference modeling on multi-dimensional physiological signals through an environmental adaptive signal processing algorithm according to the environmental semantic feature data, to obtain an environmental-compensated personalized physiological feature vector;

[0009] An analysis module for performing environmental-behavior interaction analysis through a Bayesian causal inference model using the environmental semantic feature data and the personalized physiological feature vector, to output environmental-behavior correlation features;

[0010] An input module for inputting the environmental-behavior correlation features and the personalized physiological feature vector into a graph attention network, performing multi-dimensional feature fusion through an environmental condition gating mechanism, and generating dynamic feature representations;

[0011] A construction module for constructing sub-baselines for different environmental types through a hierarchical baseline transfer learning module according to the dynamic feature representations, to obtain an environmental adaptive baseline model;

[0012] An evaluation module for evaluating a mental state through a multi-level anomaly detector based on the environmental adaptive baseline model, and outputting an early warning signal through an environmental constraint anomaly scoring mechanism.

[0013] In the technical solutions provided in the present application, the environmental parameters are collected in real time through a multi-source sensing network and processed for scene recognition by an environmental context reasoning engine, which can comprehensively perceive and analyze the influence of environmental changes on the psychological state, effectively improving the accuracy and completeness of environmental factor analysis. The environmental adaptive signal processing algorithm extracts features from multi-dimensional physiological signals and models individual differences, solving the problem of insufficient environmental interference compensation in traditional methods and significantly improving the anti-interference ability of physiological feature extraction. The Bayesian causal inference model is used for environmental-behavior interaction analysis, establishing a causal correlation mapping between environmental changes and behavioral responses, and enhancing the explainability of behavioral feature analysis. The environmental-behavioral correlation features and personalized physiological feature vectors are input into a graph attention network, and multi-dimensional feature fusion is performed through an environmental condition gating mechanism, realizing deep fusion of multi-source heterogeneous data and overcoming the problem of information loss in traditional feature fusion methods. The hierarchical baseline transfer learning module constructs sub-baselines for different environmental types, solving the baseline shift problem caused by environmental changes and improving the adaptability of the baseline model to environmental changes. The multi-level anomaly detector combines with the environmental constraint anomaly scoring mechanism to realize multi-dimensional evaluation of the psychological state, reducing the false positive rate caused by environmental interference. The artificial intelligence algorithms in the overall scheme are optimized and designed for the specific needs of mental health monitoring, including environmental perception attention mechanisms, multi-modal data graph structure representations, and personalized modeling based on transfer learning. These algorithm features significantly improve the performance and reliability of the scheme in practical applications. Through the synergistic effect of algorithms and models, the problem of mental health monitoring under dynamic environmental changes is effectively solved. BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor based on these drawings.

[0015] Figure 1 An embodiment of the mental health monitoring method based on artificial intelligence in the present application;

[0016] Figure 2 A timing diagram for real-time collection of environmental parameters through a multi-source sensing network and scene recognition processing by an environmental context reasoning engine in the present application;

[0017] Figure 3 An initial feature map structure diagram in the present application;

[0018] Figure 4An embodiment of a mental health monitoring system based on artificial intelligence in the embodiments of the present application. DETAILED DESCRIPTION

[0019] The embodiments of the present application provide a mental health monitoring method and system based on artificial intelligence. The terms "first", "second", "third", "fourth" and the like (if any) in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" or "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0020] For ease of understanding, the specific processes of the embodiments of the present application are described below. Please refer to Figure 1 An embodiment of a mental health monitoring method based on artificial intelligence in the embodiments of the present application includes:

[0021] Step S101, real-time collection of environmental parameters through a multi-source sensing network, scene recognition processing by an environmental context reasoning engine, and obtaining of environmental semantic feature data;

[0022] Step S102, feature extraction and individual difference modeling of multi-dimensional physiological signals according to the environmental semantic feature data through an environmental adaptive signal processing algorithm, and obtaining of an environmental-compensated personalized physiological feature vector;

[0023] Step S103, environmental-behavior interaction analysis by a Bayesian causal inference model using the environmental semantic feature data and the personalized physiological feature vector, and output of environmental-behavior correlation features;

[0024] Step S104, input of the environmental-behavior correlation features and the personalized physiological feature vector into a graph attention network, multi-dimensional feature fusion by an environmental condition gating mechanism, and generation of dynamic feature representation;

[0025] Step S105, sub-baseline construction for different environmental types by a hierarchical baseline transfer learning module according to the dynamic feature representation, and obtaining of an environmental adaptive baseline model;

[0026] Step S106, evaluation of a mental state by a multi-level anomaly detector based on the environmental adaptive baseline model, and output of a warning signal by an environmental constraint anomaly scoring mechanism.

[0027] It can be understood that the execution subject of the present application can be an artificial intelligence-based mental health monitoring system, and can also be a terminal or a server, which is not limited here. The server is taken as an example for description in the embodiments of the present application.

[0028] Specifically, the environmental parameters are collected in real time through a multi-source sensing network. The multi-source sensing network is composed of temperature sensors, humidity sensors, light sensors, noise sensors, and air quality sensors, and the collection frequency is 1 Hz. After the environmental parameters are collected, scene recognition processing is performed through an environmental context reasoning engine. The environmental context reasoning engine standardizes the collected parameters, maps the temperature value to the 0-1 interval, converts the humidity value to a relative humidity percentage, converts the light intensity to a lumen value, converts the noise to a decibel value, and performs dimensionless processing on the air quality index. Through environmental semantic analysis, the characteristics of different scenes such as offices, homes, and outdoors are extracted to generate environmental semantic feature data.

[0029] For the collected environmental semantic feature data, an environmental adaptive signal processing algorithm is used to extract features from multi-dimensional physiological signals and model individual differences. The multi-dimensional physiological signals include electrocardiogram signals, electroencephalogram signals, skin electricity signals, and respiratory signals. The sampling frequency of the electrocardiogram signal is 250 Hz, the sampling frequency of the electroencephalogram signal is 1000 Hz, the sampling frequency of the skin electricity signal is 50 Hz, and the sampling frequency of the respiratory signal is 20 Hz. The signal is preprocessed by noise reduction to remove power frequency interference and baseline drift. Then, according to the environmental semantic feature data, adaptive filtering is performed according to the signal characteristics in different environments to extract the physiological characteristics of the individual in a specific environment. For the electrocardiogram signal, the RR interval and heart rate variability parameters are extracted; for the electroencephalogram signal, the energy features of the α, β, θ, and δ wave bands are extracted; for the skin electricity signal, the conductance level and response amplitude are extracted; and for the respiratory signal, the respiratory frequency and depth parameters are extracted. Through individual difference modeling, these features are compared with historical data to establish a personalized parameter mapping relationship, and a personalized physiological feature vector after environmental compensation is obtained. Based on the environmental semantic feature data and the personalized physiological feature vector, an environment-behavior interaction analysis is performed through a Bayesian causal inference model. The Bayesian causal inference model analyzes the influence path of environmental changes on individual physiological and behavioral states by constructing a conditional probability network. The environmental semantic feature data is divided according to the time window, and the environmental state transition probability matrix is calculated. At the same time, the personalized physiological feature vector is analyzed in time sequence, and the behavior response mode is extracted. Through conditional probability calculation, the causal association strength between environmental changes and behavior responses is obtained, and the environment-behavior association features are output.

[0030] The environment-behavior association features and personalized physiological feature vectors are input into a graph attention network, and feature fusion is performed through an environmental condition gating mechanism. The graph attention network constructs a feature graph structure, representing different features as nodes in the graph and the associations between features as edges. The importance weights of each node are calculated through an attention mechanism to achieve dynamic selection of features. The environmental condition gating mechanism adjusts the strength of feature connections based on the current environmental conditions, highlighting the role of key features in different environments. Through feature propagation and aggregation, multi-dimensional features are fused into a unified representation to generate dynamic feature representations. Based on the dynamic feature representations, a hierarchical baseline transfer learning module is used to construct sub-baselines for different environment types. The hierarchical baseline includes a long-term baseline, a medium-term baseline, and a short-term baseline. The long-term baseline reflects the individual's stable features, the medium-term baseline describes the environmental adaptation process, and the short-term baseline depicts the immediate state changes. For each environment type, a corresponding sub-baseline model is constructed. Through transfer learning, the existing baseline knowledge in the existing environment is transferred to the new environment to quickly establish an environment-adaptive baseline model.

[0031] Based on the environment-adaptive baseline model, the mental state is evaluated through a multi-level anomaly detector. Multi-level anomaly detection includes immediate anomaly detection, trend anomaly detection, and pattern anomaly detection. Immediate anomaly detection focuses on transient state deviations, trend anomaly detection analyzes change trend anomalies, and pattern anomaly detection identifies abnormal behavior patterns. Through an environmental constraint anomaly scoring mechanism, the influence of environmental conditions on anomaly state determination is considered to generate an anomaly score with environmental constraints. The anomaly score is subjected to multi-level threshold judgment to determine the warning level and output a warning signal.

[0032] Taking mental stress monitoring as an example, when the detection object is in an office environment, the environmental parameters show a room temperature of 23°C, a relative humidity of 45%, a light intensity of 500 lux, a noise level of 45 dB, and an air quality index of good. The environmental semantic features indicate that this is a typical office scenario. Physiological signal monitoring shows reduced heart rate variability, decreased alpha band energy, increased skin electricity level, and accelerated respiratory rate. Through environment-behavior interaction analysis, it is found that this state change is related to stress reactions caused by work pressure. After dynamic feature fusion, it is shown that the stress level continuously deviates from the baseline. The multi-level anomaly detection result shows that the trend anomaly detection finds that the stress level is rising, and the pattern anomaly detection identifies a persistent stress accumulation pattern. Considering the normal stress threshold in the office environment, the system generates a moderate warning signal, suggesting timely adjustment of work pace and stress management.

[0033] In the embodiments of the present application, the environmental parameters are collected in real time through a multi-source sensing network and processed for scene recognition by an environmental context reasoning engine, which can comprehensively perceive and analyze the influence of environmental changes on the psychological state, effectively improving the accuracy and completeness of environmental factor analysis. The environmental adaptive signal processing algorithm extracts features from multi-dimensional physiological signals and models individual differences, solving the problem of insufficient environmental interference compensation in traditional methods and significantly improving the anti-interference ability of physiological feature extraction. The Bayesian causal inference model is used for environmental-behavior interaction analysis, establishing a causal correlation mapping between environmental changes and behavioral responses and enhancing the explainability of behavioral feature analysis. The environmental-behavioral correlation features and personalized physiological feature vectors are input into a graph attention network, and multi-dimensional feature fusion is performed through an environmental condition gating mechanism, realizing deep fusion of multi-source heterogeneous data and overcoming the information loss problem in traditional feature fusion methods. The hierarchical baseline transfer learning module constructs sub-baselines for different environmental types, solving the baseline shift problem caused by environmental changes and improving the adaptability of the baseline model to environmental changes. The multi-level anomaly detector combines with the environmental constraint anomaly scoring mechanism to realize multi-dimensional evaluation of the psychological state, reducing the false positive rate caused by environmental interference. The artificial intelligence algorithms in the overall scheme are optimized and designed for the specific needs of mental health monitoring, including environmental perception attention mechanisms, multi-modal data graph structure representations, and personalized modeling based on transfer learning. These algorithm features significantly improve the performance and reliability of the scheme in practical applications. Through the synergistic effect of algorithms and models, the problem of mental health monitoring under dynamic environmental changes is effectively solved.

[0034] In a specific embodiment, the process of step S101 can specifically include the following steps:

[0035] (1) receiving temperature, humidity, light, noise, and air quality data through a multi-source sensing network, inputting the temperature, humidity, light, noise, and air quality data into an environmental parameter normalization unit for standardization processing to obtain standardized environmental parameters;

[0036] (2) obtaining barometric pressure, precipitation, and ultraviolet index data from a weather data interface, performing time series alignment processing on the barometric pressure, precipitation, and ultraviolet index data to obtain weather change characteristics;

[0037] (3) performing multi-source fusion positioning calculation on the GPS coordinates, WiFi signal strength, and Bluetooth beacon data collected by the positioning module to output scene location information;

[0038] (4) obtaining image streams and audio streams of the surrounding environment using an image acquisition unit and an acoustic acquisition unit, processing the image streams and audio streams through scene semantic analysis to obtain social environment features;

[0039] (5) The standardized environment parameters, weather change characteristics, scene location information, and social environment characteristics are input into an environment context reasoning engine for multi-modal feature fusion and scene semantic reasoning to obtain environment semantic feature data;

[0040] (6) The environment semantic feature data is stored and updated in real time to generate an environment semantic feature data stream;

[0041] (7) The environment semantic features are time-stamped and data-verified through the environment semantic feature data stream to output verified environment semantic feature data.

[0042] Specifically, as shown in Figure 2 the time sequence diagram of the embodiment of the present application shows the data transmission and processing relationship among the multi-source sensor network, the environment parameter normalization unit, the meteorological data interface, the positioning module, the image acquisition unit, the acoustic acquisition unit, the environment context reasoning engine, and the time sequence database. The time sequence diagram describes the complete process from environment parameter acquisition to environment semantic feature data generation, including key steps such as raw data acquisition, data standardization processing, time sequence alignment processing, multi-source fusion positioning, scene analysis, feature fusion, data storage, and verification.

[0043] The multi-source sensor network includes temperature sensors, humidity sensors, light sensors, noise sensors, and air quality sensors. The temperature sensors collect Celsius temperature values with a sampling frequency of 1 Hz; the humidity sensors collect relative humidity percentages with a sampling frequency of 1 Hz; the light sensors collect light intensity in lux with a sampling frequency of 1 Hz; the noise sensors collect sound pressure levels in decibels (dB) with a sampling frequency of 10 Hz; and the air quality sensors collect PM2.5, PM10, and CO2 concentrations with a sampling frequency of 0.1 Hz. The environment parameter normalization unit standardizes these raw data, maps the temperature values to the [0, 1] interval, keeps the humidity values in percentage form, logarithmically processes the light intensity, linearly normalizes the noise level, and maps the air quality indicators according to the national standard classification to obtain standardized environment parameters.

[0044] The weather data interface obtains the barometric pressure, precipitation, and ultraviolet index data of the weather station in real time through the API. The barometric pressure data is in units of hundred pascals, the precipitation data is in units of millimeters per hour, and the ultraviolet index is a dimensionless value. Due to the different collection frequencies of different data, the barometric pressure data is collected every 10 minutes, the precipitation data is collected every 5 minutes, and the ultraviolet index is collected every 30 minutes, which requires time series alignment processing. Time series alignment uses the interpolation method to unify all data to a 5-minute interval time series to form weather change characteristics. The data collected by the positioning module includes GPS coordinates, WiFi signal strength, and Bluetooth beacon data. GPS provides latitude and longitude information with a sampling frequency of 1 Hz; the WiFi module scans the MAC addresses and signal strengths (RSSI) of surrounding access points every 3 seconds; and the Bluetooth beacon broadcasts signal strength and a unique identifier every second. The multi-source fusion positioning calculation uses a weighted average method, mainly relying on GPS positioning outdoors and combining WiFi and Bluetooth beacon data for triangulation indoors. The weights are dynamically adjusted according to the reliability of the signal strength to output unified scene location information.

[0045] The image acquisition unit acquires an environmental image stream with a resolution of 1920x1080 and a frame rate of 30fps; the acoustic acquisition unit acquires an environmental audio stream with a sampling rate of 44.1kHz and a bit depth of 16bit. Scene semantic analysis performs scene segmentation and target detection on the image stream, identifying features such as crowd density, activity type, and scene category; and performs acoustic event detection on the audio stream, identifying features such as environmental noise type and voice density. These features are combined to form social environment features, describing the characteristics of the surrounding social scene. The environmental context reasoning engine receives standardized environmental parameters, weather change characteristics, scene location information, and social environment features, and performs multi-modal feature fusion. Feature fusion uses an attention mechanism to dynamically allocate weights based on the importance of different features in the current scene. Scene semantic reasoning is based on a rule base and a probabilistic graph model, mapping multi-modal features to predefined scene categories such as office environment, home environment, and outdoor environment, and extracting semantic attributes of the scene to generate environmental semantic feature data.

[0046] For the storage and update of environmental semantic feature data, a time series database is used for structured storage. The data is organized according to the timestamp, supporting fast time range query and aggregation analysis. The real-time update mechanism ensures that new environmental semantic features can be written into the database in a timely manner, while maintaining the time continuity of the data, generating an environmental semantic feature data stream. The environmental semantic feature data stream is marked with a timestamp to ensure the time sequence and traceability of the data. Data verification includes integrity verification, consistency verification, and outlier detection. Integrity verification ensures that there are no missing time points in the data stream; consistency verification ensures that the logical relationship between related features is reasonable; outlier detection identifies data points that are significantly outside the normal range. The verified environmental semantic feature data serves as an important input for subsequent mental health monitoring.

[0047] For example: mental health monitoring for a person working in an office. The multi-source sensor network collects indoor temperature 24℃, relative humidity 45%, light intensity 600 lux, average noise level 52dB, and air quality index 75. After standardization processing, these raw data are converted to a unified numerical range. At the same time, from the weather data interface, the atmospheric pressure of the day is obtained 1013.2 hundred Pa, no precipitation, ultraviolet index 3, after time alignment, form the weather feature sequence. The positioning module determines the location in the office area through WiFi signal strength and Bluetooth beacon positioning. Image stream analysis shows medium population density, and the activity type is mainly sitting at work; audio stream analysis shows low-intensity human voice conversation and keyboard tapping sound. The environmental context reasoning engine infers that the current is a standard office scene according to these features, and extracts related environmental semantic features. These features are structured and stored and updated in real time to form a continuous data stream. After timestamp marking and data verification, reliable environmental semantic feature data is output.

[0048] In a specific embodiment, the process of performing step S102 can specifically include the following steps:

[0049] (1) Collecting electrocardiogram, electroencephalogram, skin electricity, and respiration data through a physiological signal acquisition device, and performing time domain feature calculation on the electrocardiogram, electroencephalogram, skin electricity, and respiration data to obtain raw physiological feature data;

[0050] (2) Time aligning the environmental semantic feature data with the raw physiological feature data, compensating for environmental interference in the raw physiological feature data through weighted processing, and obtaining compensated physiological features;

[0051] (3) Frequency domain decomposition of the compensated physiological features, decomposition of the electrocardiogram signal into low, medium, and high frequency components, decomposition of the electroencephalogram signal into α, β, θ, and δ wave bands, and obtaining multi-band feature data;

[0052] (4) Calculate feature statistics including mean, variance, kurtosis, skewness from multi-band feature data, generate statistical feature vector;

[0053] (5) Calculate similarity between statistical feature vector and historical sample data, extract individual feature difference parameters, correct statistical feature vector for individualization, obtain corrected feature vector;

[0054] (6) Perform dimension conversion and normalization processing on the corrected feature vector through environment adaptive signal processing algorithm, obtain environment compensated individualized physiological feature vector.

[0055] Specifically, the electrocardiogram signal collected by the physiological signal collection device has a sampling frequency of 250 Hz, and the P wave, QRS complex and T wave features in the electrocardiogram waveform are collected; the electroencephalogram signal has a sampling frequency of 1000 Hz, and the potential activity of key brain regions such as frontal lobe and parietal lobe is collected; the skin electricity signal has a sampling frequency of 50 Hz, and the skin conductance change is recorded; the respiration signal has a sampling frequency of 20 Hz, and the respiration waveform is collected. The collected signals are calculated for time domain features, and the electrocardiogram signal is calculated for RR interval and heart rate variability; the electroencephalogram signal is calculated for amplitude and power; the skin electricity signal is calculated for baseline level and response amplitude; the respiration signal is calculated for respiration frequency and depth, to obtain original physiological feature data.

[0056] The time alignment of environmental semantic feature data and original physiological feature data uses interpolation method to unify all data to a time interval of 10 ms. The environmental interference is compensated through weighted processing, and the weight coefficient is dynamically adjusted according to the influence degree of environmental factors. For temperature influence, the heart rate variability parameter is adjusted according to the body temperature regulation response; for noise influence, the electroencephalogram amplitude is corrected; for light change, the skin electricity baseline level is corrected; for air quality influence, the respiration parameter is adjusted, to obtain compensated physiological features. The compensated physiological features are decomposed in frequency domain, and the signals are analyzed for frequency spectrum by using fast Fourier transform (FFT). The electrocardiogram signal is decomposed into low frequency (0.01-0.04 Hz) reflecting sympathetic nerve activity, medium frequency (0.04-0.15 Hz) reflecting blood pressure regulation, and high frequency (0.15-0.4 Hz) reflecting parasympathetic nerve activity. The electroencephalogram signal is decomposed into delta wave (0.5-4 Hz) reflecting deep sleep state, theta wave (4-8 Hz) reflecting shallow sleep and attention, alpha wave (8-13 Hz) reflecting relaxation state, and beta wave (13-30 Hz) reflecting wakefulness and attention concentration, to obtain multi-band feature data.

[0057] Statistical feature calculation performs statistical analysis on multi-band feature data. Mean reflects the overall level of the signal, such as the mean value of heart rate variability indicating the autonomic nervous regulation ability; variance describes the volatility of the signal, such as the variability of the amplitude of the brain waves indicating the stability of the brain activity; kurtosis measures the sharpness of the distribution, such as the burstiness of the skin electricity response; skewness represents the asymmetry of the distribution, such as the regularity of the breathing pattern. These statistics form a statistical feature vector.

[0058] Similarity calculation is performed between the statistical feature vector and the historical sample data, and the cosine similarity is used to measure the distance between the feature vectors. Individual feature difference parameters include baseline level difference, response sensitivity difference, regulation ability difference, etc. Personalized correction is performed on the statistical feature vector, and the dimension and range of the feature values are adjusted according to the individual feature difference parameters to obtain a corrected feature vector. The environment adaptive signal processing algorithm processes the corrected feature vector. Dimension conversion unifies different physiological indicators to the same feature space, and principal component analysis is used to reduce the feature dimension. Normalization processing maps the feature values to the [0, 1] interval, and considers the influence of environmental conditions on the normalization parameters to obtain the environment-compensated personalized physiological feature vector.

[0059] For example: when monitoring the subject in an office environment, physiological signals and environmental data are collected simultaneously. The electrocardiogram signal shows that the RR interval is shortened, indicating that the heart rate is accelerated; the beta wave energy in the electroencephalogram signal is enhanced, indicating that the attention is highly concentrated; the skin electricity level is elevated, reflecting emotional tension; the breathing becomes rapid. The environmental data shows that the room temperature is relatively high and the noise interference is large. Time alignment is performed to align the physiological data and the environmental data to a unified time axis. Then, compensation is performed according to the environmental influence, such as reducing the influence weight of temperature on heart rate and eliminating the interference of noise on electroencephalogram. Frequency domain decomposition is performed on the compensated signal, and analysis finds that the low-frequency component of the electrocardiogram signal is enhanced, indicating that the sympathetic nerve activity is increased; the proportion of beta waves in the electroencephalogram is increased, reflecting the increase of cognitive load. Statistical features are calculated, and it is found that the mean value of heart rate variability is reduced, the variance of electroencephalogram amplitude is increased, the kurtosis of skin electricity response is increased, and the regularity of breathing is decreased. Compared with the historical data of the user, it is found that the current state deviates significantly from the personal baseline level. After personalized correction and dimension conversion, a feature vector reflecting the current psychological stress state is generated, which provides a basis for subsequent stress level evaluation.

[0060] In a specific embodiment, the process of performing step S103 can specifically include the following steps:

[0061] (1) The environmental semantic feature data is divided into multiple time sequence segments according to a time window, and feature extraction is performed on each time sequence segment to obtain an environmental state sequence;

[0062] (2) A state transition matrix is constructed according to the environmental state sequence, the transition probability between environmental states is calculated through matrix decomposition, and an environmental change pattern is obtained.

[0063] (3) Segment the personalized physiological feature vector with a sliding window, calculate the statistics and trend in each window, and form a physiological state sequence;

[0064] (4) Use the environmental change pattern and the physiological state sequence to calculate the conditional probability distribution through the Bayesian causal inference model, and obtain the environmental-physiological correlation measure;

[0065] (5) Align the user activity data and social behavior data with the timestamp, and perform time series correlation analysis with the environmental-physiological correlation measure to generate behavior response features;

[0066] (6) Perform path analysis and weight calculation on the behavior response features through a multi-layer causal graph, and output the environmental-behavior correlation features.

[0067] Specifically, the environmental semantic feature data is divided according to a fixed window length of 10 minutes, and each window has an overlap rate of 50%, forming continuous time series segments. For each time series segment, extract features such as temperature change rate, humidity fluctuation amplitude, light intensity gradient, noise spectrum feature, and air quality index change, to form an environmental state vector. The environmental state vectors of multiple time series segments are arranged in chronological order to form an environmental state sequence. Based on the environmental state sequence, a state transition matrix is constructed to record the transition relationship between different environmental states. Through matrix decomposition method, the state transition matrix is decomposed into a basic state matrix and a transition probability matrix. The basic state matrix represents the typical state pattern of the environment, and the transition probability matrix describes the change rule between states. The matrix decomposition uses a non-negative matrix decomposition algorithm to ensure that the decomposition result has physical meaning. By analyzing the main paths and jump features in the transition probability matrix, the typical patterns of environmental change are identified.

[0068] When segmenting the personalized physiological feature vector with a sliding window, the window length is set to 5 minutes and the sliding step is 1 minute. In each window, calculate the heart rate variability mean, electroencephalogram energy distribution, skin electricity level change rate, and respiratory frequency fluctuation. At the same time, calculate the trend of these indicators, including first-order difference, change acceleration, etc. Organize these features in chronological order to form a physiological state sequence that describes the dynamic changes of physiological state. The Bayesian causal inference model calculates the conditional probability distribution to analyze the correlation between the environmental change pattern and the physiological state sequence. A conditional probability table is established from environmental state to physiological state, and then a Bayesian network structure is used to describe the causal dependence relationship between variables. When calculating the environmental-physiological correlation measure, both direct causal influence and indirect transmission effect are considered to obtain a quantitative correlation strength index.

[0069] User activity data includes motion state, posture change, operation behavior, etc. Social behavior data includes conversation frequency, social interaction intensity, etc. Align these data to the unified time axis according to the timestamp, and perform time correlation analysis on the environment-physiological correlation metrics. The time correlation analysis adopts the dynamic time warping algorithm to calculate the time-dependent relationship between behavior change and environment-physiological correlation, and generates a feature vector reflecting the individual behavior response characteristics.

[0070] The path analysis and weight calculation of the multi-layer causal diagram are expressed as:

[0071]

[0072] wherein, represents the correlation strength from the environmental factor i to the behavior response j, represents the direct influence coefficient of the environmental factor i on the intermediate state k, represents the action strength of the intermediate state k on the behavior response j, is the importance weight of the intermediate state k, and n is the total number of intermediate states.

[0073] In the calculation process, the normalized value of the environmental characteristics is calculated:

[0074]

[0075] wherein, is the original influence value of the environmental factor i on the state k, is the significance coefficient of the environmental factor i, and p is the total number of environmental factors.

[0076] The environment-behavior correlation characteristics are obtained by weighted combination:

[0077]

[0078] wherein, is the environment-behavior correlation characteristics, is the path importance weight, and q is the dimension number of the behavior response.

[0079] For example, when monitoring mental health in an office environment, the environmental semantic feature data records the changes in the environment from morning to afternoon. Through time window division, the morning environment state is captured, with low noise and moderate light. At noon, the temperature rises and the density of personnel increases. In the afternoon, the light weakens and the air quality decreases. These features constitute the environmental state sequence, and through state transition analysis, the typical office environment daily change pattern is found. The personalized physiological feature vector shows that the physiological indicators fluctuate with the changes in the environment, such as the increase in brain electrical beta wave activity when noise increases, and the adjustment of heart rate variability when temperature changes. Bayesian causal inference reveals the correlation between environmental changes and physiological states, such as the increased probability of respiratory rate changes due to poor air quality. Combined with recorded activity data and social behavior data, it is observed that environmental changes trigger behavior adjustments, such as a decrease in conversation frequency when noise increases and a decrease in activity intensity when temperature rises. Through multi-layer causal graph analysis, the direct and indirect paths of environmental factors affecting behavior are quantified, thereby constructing environmental-behavioral correlation features.

[0080] In a specific embodiment, the process of performing step S104 can specifically include the following steps:

[0081] (1) Align the environmental-behavioral correlation features and the personalized physiological feature vector by timestamp, and generate a feature sequence matrix through data completion processing;

[0082] (2) Perform feature node construction on the feature sequence matrix, calculate the connection weights between nodes, and obtain an initial feature graph structure;

[0083] (3) Use the attention calculation unit to score the importance of the nodes in the initial feature graph structure, generating an attention weight distribution;

[0084] (4) Perform gated operation on the attention weight distribution and environmental condition information to dynamically adjust the feature connection strength, obtaining a gated feature graph;

[0085] (5) Perform feature propagation and aggregation on the gated feature graph through a message passing mechanism to generate a node update vector;

[0086] (6) Perform time series combination and dimension transformation on the node update vector to output dynamic feature representation.

[0087] Specifically, the environmental-behavioral correlation features and the personalized physiological feature vector are time-synchronized. Interpolation is used to unify data with different sampling frequencies to the same time scale. For missing data, linear interpolation or nearest neighbor interpolation is used for completion to form a feature sequence matrix. Each row in the feature sequence matrix represents a time point, and each column corresponds to a feature dimension.

[0088] During the feature node construction process, each feature in the feature sequence matrix is represented as a node in the graph structure. The connection weights between nodes are calculated by the following formula:

[0089]

[0090] where, represents the connection weight from node i to node j, is the importance coefficient at time point t, is the association strength of nodes i and j at time t, is the time decay factor, T is the total number of time steps, and N is the total number of nodes.

[0091] The calculation of attention weight distribution uses the following formula:

[0092]

[0093] where, is the attention weight of node i, is the feature similarity between nodes i and j, is the feature importance measure of node j, is the position encoding factor.

[0094] When the attention weight distribution is gated with environmental condition information, a dynamic threshold mechanism is used. According to the current environmental conditions, set the gating threshold, selectively retain or suppress the feature connection strength. The gating operation considers the comprehensive influence of temperature, humidity, illumination, noise and other environmental factors, dynamically adjusts the connection relationship between features. The message passing mechanism updates the node features through iterative method. In each iteration, the node collects information from its neighbor nodes and aggregates them according to the connection weights. The information aggregation process considers the influence of direct connection and multi-hop path, and preserves the original feature information through residual connection. The node update process also considers the temporal dependence relationship, ensuring the temporal consistency of the features.

[0095] As shown in Figure 3 , it is the initial feature graph structure diagram in the embodiment of the present application, which contains three main parts: physiological feature nodes (including heart rate variability, brain electrical activity, skin electrical response and respiratory features), behavior feature nodes (including activity intensity, concentration degree and interaction frequency) and environmental feature nodes (including temperature state, noise level and illumination intensity). The arrows in the figure represent the connection relationship between different feature nodes, the bidirectional arrows represent the mutual influence between features, and the unidirectional arrows represent the influence of environmental features on physiological and behavioral features.

[0096] The node update vector is time-series combined to fuse features of different time scales. The dimension transformation process includes dimension reduction and feature mapping, which compresses high-dimensional features to a proper dimensional space while maintaining the discriminative information of the features, forming the final dynamic feature representation.

[0097] For example, in the mental health monitoring of an office environment, the environment-behavior association features include the influence pattern of environmental changes on behavior, and the personalized physiological feature vector includes heart rate, electroencephalogram, skin electricity, etc. physiological indicators. Align these features at one-minute intervals, and supplement the data for time points with missing samples by linear interpolation. In feature node construction, each physiological indicator and behavior feature is represented as a node, such as heart rate variability node, electroencephalogram alpha wave node, behavior activity node, etc. When calculating the connection weight between nodes, the correlation of different indicators in the office environment is considered, for example, the strong correlation between electroencephalogram beta wave and behavior concentration in a noisy environment. The importance of nodes is scored through the attention mechanism, and it is found that the importance of each indicator varies significantly under different environmental conditions. For example, during high-intensity work periods, the weight of electroencephalogram indicators is higher, while during rest periods, the weight of heart rate variability indicators increases. The gating mechanism dynamically adjusts the strength of feature connection according to environmental conditions such as temperature changes and noise levels. Finally, through message passing and feature aggregation, a dynamic feature representation reflecting the current psychological state is generated.

[0098] In a specific embodiment, the process of performing step S105 can specifically include the following steps:

[0099] (1) Time series decomposition is performed on the dynamic feature representation to separate long-term features, medium-term features, and short-term features, obtaining a hierarchical feature sequence;

[0100] (2) Different environmental scenarios are classified according to historical environmental data, an environmental type feature library is established, and an environmental classification standard is formed;

[0101] (3) The hierarchical feature sequence is grouped according to the environmental classification standard, and statistical features are calculated for each group of data to generate an environmental related feature set;

[0102] (4) The environmental related feature set is divided into subspaces, and a feature distribution matrix under different environmental types is constructed to obtain environmental subscale data;

[0103] (5) The environmental subscale data is mapped through a transfer transformation to establish a correlation mapping relationship between environmental types, and obtain baseline migration parameters;

[0104] (6) The baseline migration parameters and the environmental subscale data are integrated and optimized to output an environmental adaptive baseline model.

[0105] Specifically, the dynamic feature representation is decomposed by wavelet transform, separating the time series into different time scales. The long-term feature reflects the monthly or quarterly trend, with a time span of more than 30 days; the medium-term feature represents the weekly or ten-day fluctuation rule, with a time span of 7-30 days; the short-term feature depicts the instantaneous change within a day or between days, with a time span of 1-7 days. By wavelet reconstruction, the features of each scale are recombined to obtain a hierarchical feature sequence.

[0106] The type division of historical environmental data uses clustering analysis method. The environmental parameters such as temperature, humidity, illumination, and noise are standardized, and then the hierarchical clustering algorithm is used to group the environmental data. In the clustering process, the time sequence correlation and spatial distribution characteristics of environmental parameters are considered, and similar environmental states are classified into the same category. When establishing the environmental type feature library, the typical feature mode of each type of environment is extracted, including parameter mean, fluctuation range, change trend, etc. statistical characteristics, forming the environmental classification standard. When grouping data according to the environmental classification standard, the nearest neighbor matching principle is used. The environmental state at each time point is matched with the standard mode in the environmental type feature library to determine the current environmental type. The statistical characteristics of each group of data are calculated, including time domain features (mean, variance, kurtosis, skewness), frequency domain features (power spectral density, main frequency component), and nonlinear features (approximate entropy, sample entropy), to generate the environmental related feature set.

[0107] The subspace division of the environmental related feature set uses the local linear embedding algorithm. For different environmental types, the local neighborhood relationship of the features is constructed to maintain the topological structure of the data. The high-dimensional features are projected into low-dimensional subspaces through feature mapping, and each subspace corresponds to the feature distribution under a certain environmental type. The feature distribution matrix of all subspaces is combined to form the environmental sub-baseline data. In the migration transformation process, the domain adaptation method is used to map the environmental sub-baseline data. The feature distribution difference between the source and target environmental domains is identified, and then a feature transformation function is designed to map the source domain features to the feature space of the target domain. By minimizing the distribution difference and maintaining the discriminative nature of the features, the correlation mapping relationship between environmental types is established, and the baseline migration parameters are obtained.

[0108] The integration of baseline migration parameters and environmental sub-baseline data uses a weighted fusion strategy. According to the environmental similarity, the migration weight is set to adaptively adjust the baseline data. The optimization calculation process uses the gradient descent method to minimize the migration error and baseline deviation, and outputs the environmental adaptive baseline model.

[0109] For example, in the mental health monitoring scenario in an office environment, the collected dynamic feature representation is decomposed into a time series. Long-term features reflect the baseline of employees' mental state in different seasons, such as changes in circadian rhythms in air-conditioned environments in summer; medium-term features show the state differences between weekdays and weekends, including work pressure accumulation and recovery periods; short-term features capture state fluctuations in daily work, such as the immediate impact of events such as meetings and deadlines. By analyzing historical environmental data, the office environment is divided into multiple typical types: standard office environment (temperature 22-26℃, relative humidity 40-60%, illumination 500-700 lux, environmental noise 45-55 dB), meeting environment (increased personnel density, increased noise level), rest environment (reduced light intensity, reduced noise level), etc. A feature library is established for each environment type, recording the typical distribution characteristics of environmental parameters.

[0110] According to the environmental classification standard, the hierarchical feature sequence is grouped. In the standard office environment, heart rate variability remains within a stable range, and brain electrical beta wave activity is moderate; in the meeting environment, attention indicators increase, and skin electricity level fluctuations increase; in the rest environment, alpha wave activity is enhanced, and the autonomic nervous system tends to be in a balanced state. Calculate feature statistics for each case to construct a feature distribution matrix. When the environment changes, such as from a standard office environment to a meeting environment, adjust the baseline parameters through transfer learning methods. Considering the influence mode of environmental changes on physiological indicators, establish a feature mapping relationship to achieve smooth transition of the baseline model. The final environment-adaptive baseline model can dynamically adjust the evaluation criteria according to different environmental conditions and accurately reflect changes in mental health status.

[0111] In a specific embodiment, the process of performing step S106 can specifically include the following steps:

[0112] (1) Calculate the feature distance between real-time mental state data and the environment-adaptive baseline model, and normalize the calculation results to obtain a deviation degree index;

[0113] (2) Layered evaluation of mental state according to the deviation degree index, calculate the immediate abnormal score, trend abnormal score and mode abnormal score through multi-scale comparison, and obtain the abnormal feature vector;

[0114] (3) Calculate the spatio-temporal distribution characteristics of the abnormal state using the abnormal feature vector to generate an abnormal state distribution map;

[0115] (4) Correlation analysis of the abnormal state distribution map and the environmental conditions to calculate the environmental constraint weight coefficient to form a constraint score parameter;

[0116] (5) Weighted combination of the constraint score parameter and the abnormal feature vector to generate a warning level identifier through multi-level threshold judgment;

[0117] (6) According to the mapping relationship between the early warning level identifier and the environmental risk degree, an early warning signal is output.

[0118] Specifically, since the dimensions and distributions of different feature dimensions are different, the features are standardized to convert the feature dimensions to the same scale space. The deviation value of each feature dimension from the baseline model is calculated, and the covariance relationship between the features is considered to obtain a comprehensive deviation value. The deviation value is mapped to the [0, 1] interval through the max-min normalization to generate a standardized deviation degree index. The hierarchical evaluation process is based on a multi-scale analysis framework. The instantaneous abnormal score reflects the instantaneous deviation of the current state from the baseline, which is obtained by calculating the Euclidean distance of the feature vector; the trend abnormal score describes the dynamic characteristics of state changes, and the change rate and acceleration of the feature sequence are calculated through a sliding window; the pattern abnormal score represents the abnormal degree of the behavior pattern, and the difference degree from the normal pattern is calculated through sequence pattern matching. The three types of scores constitute a multi-dimensional abnormal feature vector.

[0119] The spatio-temporal distribution characteristics of the abnormal state include the time dimension and the state space dimension. In the time dimension, the duration, frequency and periodicity of the abnormal features are analyzed; in the state space dimension, the distribution range and aggregation degree of the abnormal state in different feature dimensions are calculated. These distribution characteristics are visualized to form an abnormal state distribution map.

[0120] The association analysis between the abnormal state distribution map and the environmental conditions uses a two-way association calculation. The modulating effect of environmental factors on the abnormal state distribution is identified, while considering the differences in the sensitivity of abnormal states to the environment. Based on correlation analysis, the weight coefficient of each environmental factor is calculated to form a comprehensive environmental constraint score parameter. The weighted combination of the constraint score parameter and the abnormal feature vector uses an adaptive weight strategy. According to the current environmental conditions, the weights of each feature are dynamically adjusted, focusing on the abnormal features with high environmental sensitivity. The multi-level threshold judgment divides the weighted abnormal degree into different levels to generate a graded early warning identifier.

[0121] The mapping of the early warning level identifier and the environmental risk degree uses a piecewise mapping function. According to the environmental risk level, the early warning threshold is dynamically adjusted, which reduces the early warning threshold in high-risk environments and improves the early warning sensitivity; in low-risk environments, the early warning threshold is increased to reduce false positives. The mapping result determines the level of the final output early warning signal.

[0122] For example, in an office environment, the mental health status of an employee is monitored. By comparing the real-time monitoring data with the environmental adaptive baseline model, it is found that the heart rate variability index is lower than the baseline mean, the brain beta wave energy is consistently high, and the skin electricity level fluctuates greatly. Based on these deviations, a comprehensive deviation index is calculated. Multi-scale analysis shows that the immediate abnormal score reflects the current state of high tension; the trend abnormal score indicates that the stress level is rising continuously; the pattern abnormal score indicates that the work-rest rhythm is broken. The abnormal state distribution chart shows that this state mainly occurs in the afternoon of weekdays, and has significant correlation with environmental noise level and work task density. Environmental constraint analysis finds that abnormal state is more likely to occur in high temperature and high noise environment. Considering the abnormal degree and environmental risk factors, it is determined that the current state reaches the medium warning level, and it is suggested to take timely intervention measures such as adjusting work rhythm and improving environmental conditions. This example shows the complete processing flow from data collection, feature extraction to anomaly detection and warning generation, and reflects the important role of environmental factors in mental health monitoring.

[0123] In a specific embodiment, the process of performing the step of hierarchical evaluation of mental state according to the deviation degree index can specifically include the following steps:

[0124] (1) Divide the deviation degree index into short-term sequence, medium-term sequence and long-term sequence according to the time window, and perform data smoothing processing on each sequence to obtain multi-scale time series data;

[0125] (2) Perform sliding window analysis on the short-term sequence in the multi-scale time series data, calculate the difference between the statistical quantity in the window and the baseline threshold, and generate the immediate abnormal score;

[0126] (3) Calculate the slope of the medium-term sequence in the multi-scale time series data by trend decomposition, and analyze the deviation of the slope change from the baseline trend to obtain the trend abnormal score;

[0127] (4) Extract pattern features from the long-term sequence in the multi-scale time series data, identify abnormal patterns by sequence similarity calculation, and output the pattern abnormal score;

[0128] (5) Combine and transform the features of the immediate abnormal score, the trend abnormal score and the pattern abnormal score to form an abnormal feature matrix;

[0129] (6) Perform dimension reduction and normalization processing on the abnormal feature matrix to generate an abnormal feature vector.

[0130] Specifically, short-term sequences adopt a 1-hour time window, reflecting immediate state changes; medium-term sequences adopt an 8-hour time window, corresponding to changes in a single workday; long-term sequences adopt a 24-hour time window, representing cross-day change patterns. Exponential weighted moving average smoothing processing is applied to these sequences respectively to eliminate the influence of random fluctuations, obtaining multi-scale time series data.

[0131] The sliding window analysis of short-term sequences adopts a 5-minute window length, with a window overlap rate of 50%. Within each window, statistical quantities such as heart rate variability bias, brain electrical energy bias, and skin electricity level bias are calculated and compared with the corresponding baseline threshold. The baseline threshold is determined according to the normal fluctuation range of historical data, and the degree of exceeding the threshold determines the size of the immediate abnormal score.

[0132] The trend analysis of medium-term sequences is calculated using the following formula:

[0133]

[0134] wherein, is the trend abnormal score, is the trend slope of the i-th time point, is the environmental modulation coefficient, is the weight factor, is the j-th component of the baseline trend vector, is the fluctuation compensation factor, n is the sequence length, and m is the feature dimension.

[0135] The pattern analysis of long-term sequences adopts the dynamic time warping algorithm to calculate the similarity between the current sequence and the historical normal pattern. By extracting the periodic features, mutation features, and cumulative features of the sequence, a pattern feature descriptor is constructed. Based on these features, the distance measure between sequences is calculated to identify abnormal behavior patterns and output the pattern abnormal score. The feature combination of the abnormal score adopts a multi-level structure. The first layer is the original abnormal score, including the immediate abnormal score, the trend abnormal score, and the pattern abnormal score; the second layer is the cross-feature, calculating the interaction between different types of abnormal scores; the third layer is the time sequence feature, describing the time evolution characteristics of the abnormal score. These feature combinations form a three-dimensional feature matrix.

[0136] The dimension reduction of the feature matrix adopts the principal component analysis method, retaining the principal components with a cumulative contribution rate of 95%. The features after dimension reduction are subjected to maximum and minimum normalization processing, mapping all features to the [0, 1] interval to generate a standardized abnormal feature vector.

[0137] For example, in the monitoring of mental health in an office environment, a day's monitoring data of an employee is continuously collected. The deviation degree index is divided according to different time scales: the short-term sequence records the state fluctuation in the working process, such as the attention change during the meeting; the medium-term sequence reflects the state change in different time periods in the morning and afternoon; and the long-term sequence contains the working-rest cycle mode of the whole day. Through sliding window analysis, it is found that during the important meeting, the real-time anomaly score increases, showing that the heart rate is accelerated and the skin electricity activity is enhanced. Trend analysis shows that there is a continuous stress accumulation trend in the afternoon period, and the slope is significantly higher than the baseline level. Pattern analysis finds that the original regular work-rest rhythm is broken, showing a continuous tension abnormal pattern. These anomaly scores are processed through feature combination and dimension reduction to finally form a feature vector reflecting the overall abnormal state. The vector contains not only the abnormal degree of the current state, but also the dynamic characteristics of the state change, providing a basis for subsequent early warning judgment.

[0138] In a specific embodiment, the process of performing the slope calculation step on the medium-term sequence in the multi-scale time series data through trend decomposition can specifically include the following steps:

[0139] (1) The medium-term sequence in the multi-scale time series data is segmented according to a fixed step length, and statistical feature extraction is performed on each segment of data to obtain sequence feature points;

[0140] (2) Seasonal decomposition is performed on the sequence feature points to separate the trend component, the periodic component and the random component, and trend sequence data is obtained;

[0141] (3) Linear fitting is performed on the trend sequence data using the least squares method to calculate the slope value in each time period, and a trend slope sequence is formed;

[0142] (4) The trend slope sequence is difference calculated with the baseline trend data to establish a slope deviation matrix, and a trend change feature is obtained;

[0143] (5) The trend change feature is aggregated in the time dimension by weighted average to generate a trend deviation vector;

[0144] (6) The trend deviation vector is amplitude normalized and threshold mapped to output a trend anomaly score.

[0145] Specifically, a fixed step segmentation strategy is adopted. The mid-term sequence is divided into multiple time segments with a fixed step of 30 minutes, and a 10% overlap interval is set between each time segment to ensure data continuity. Statistical feature extraction is performed on the data in each time segment, including mean, variance, kurtosis, skewness, and entropy. These statistics form a multi-dimensional feature vector, each corresponding to a feature point on the time sequence, forming a discrete sequence of feature point sets. The seasonal decomposition of the sequence feature points uses the X-12-ARIMA decomposition method. This method preprocesses the original sequence, including outlier detection and missing value filling, and then decomposes the sequence into trend component, periodic component and random component. The trend component reflects the long-term trend of the data, the periodic component captures the cyclic variation of the data, and the random component contains irregular fluctuations. The trend component obtained by decomposition forms the trend sequence data.

[0146] The linear fitting of the trend sequence data uses the piecewise least squares method. The trend sequence is divided into multiple subintervals according to the working period, and linear least squares fitting is applied in each subinterval. The fitting process takes into account the weight of the data points, with recent data points having a larger weight and distant data points gradually decaying. By calculating the slope of the fitted straight line in each subinterval, the trend slope sequence reflecting the rate of state change is obtained. The difference between the trend slope sequence and the baseline trend data is calculated using the dynamic time warping method. The baseline trend data comes from historical data statistics and represents the change trend under normal conditions. By calculating the difference between the current trend slope sequence and the baseline trend data at different time points, a slope deviation matrix reflecting the degree of trend deviation is constructed. Each row of the deviation matrix corresponds to a time point, and each column corresponds to a feature dimension.

[0147] The time dimension aggregation of the trend change feature uses the exponential weighted average method. For each feature dimension in the slope deviation matrix, a time decay factor is set so that recent deviation values have a larger weight. By weighted average calculation, the deviation values in the time dimension are aggregated into a vector, each component representing the comprehensive deviation degree in the corresponding feature dimension, forming a trend deviation vector. The amplitude normalization process uses an adaptive threshold method. The distribution characteristics of each component of the trend deviation vector are calculated to determine the normalization parameter. Then the deviation value is mapped to the standard interval, and the mapping function is dynamically adjusted according to the environmental conditions. Finally, through multiple threshold judgments, the normalized deviation value is converted into a trend anomaly score.

[0148] For example, in the mental health monitoring of an office environment, the monitoring data of an employee in a day is analyzed. The medium-term sequence is divided by 30-minute steps, and the features in each time period are extracted. For example, in the period from 9:00 to 9:30 in the morning, the statistical features of indicators such as heart rate variability, brain energy, and skin electricity level are calculated. Through seasonal decomposition, it is found that these indicators have obvious intra-day variation rules, such as the period of high energy in the morning and the period of fatigue in the afternoon. Through segmented fitting of the trend component, it is found that after continuous work for 4 hours, the stress indicators begin to rise rapidly, and the slope is obviously higher than the normal level. Compare this trend change with the historical baseline to construct a deviation matrix, and obtain a comprehensive deviation vector through time weighting. Finally, after normalization and threshold mapping, the trend anomaly score reflecting the degree of stress accumulation is output, which provides a basis for timely adjusting the work rhythm.

[0149] The mental health monitoring method based on artificial intelligence in the embodiments of the application is described above, and the mental health monitoring system based on artificial intelligence in the embodiments of the application is described below. Please refer to Figure 4 An embodiment of the mental health monitoring system based on artificial intelligence in the embodiments of the application includes:

[0150] The acquisition module 201 is configured to collect environmental parameters in real time through a multi-source sensor network, and perform scene recognition processing on the collected environmental parameters through an environmental context reasoning engine to obtain environmental semantic feature data.

[0151] The modeling module 202 is configured to perform feature extraction and individual difference modeling on multi-dimensional physiological signals according to the environmental semantic feature data through an environmental adaptive signal processing algorithm to obtain an individualized physiological feature vector after environmental compensation.

[0152] The analysis module 203 is configured to perform environmental-behavior interaction analysis on the environmental semantic feature data and the individualized physiological feature vector through a Bayesian causal inference model to output environmental-behavior correlation features.

[0153] The input module 204 is configured to input the environmental-behavior correlation features and the individualized physiological feature vector into a graph attention network, perform multi-dimensional feature fusion through an environmental condition gating mechanism, and generate a dynamic feature representation.

[0154] The construction module 205 is configured to construct sub-baselines for different environmental types from a hierarchical baseline transfer learning module according to the dynamic feature representation to obtain an environmental adaptive baseline model.

[0155] The evaluation module 206 is configured to evaluate a mental state through a multi-level anomaly detector based on the environmental adaptive baseline model, and output a warning signal through an environmental constraint anomaly scoring mechanism.

[0156] Through the synergistic cooperation of the above components, real-time collection of environmental parameters through a multi-source sensor network and scene recognition processing through an environmental context reasoning engine can comprehensively perceive and analyze the influence of environmental changes on the psychological state, effectively improving the accuracy and integrity of environmental factor analysis. The environmental adaptive signal processing algorithm extracts features from multi-dimensional physiological signals and models individual differences, solving the problem of insufficient environmental interference compensation in traditional methods and significantly improving the anti-interference ability of physiological feature extraction. The Bayesian causal inference model is used for environmental-behavior interaction analysis, establishing a causal correlation mapping between environmental changes and behavioral responses, and enhancing the explainability of behavioral feature analysis. The environmental-behavioral correlation features and personalized physiological feature vectors are input into the graph attention network, and through the environmental condition gating mechanism, multi-dimensional feature fusion is realized, overcoming the information loss problem in traditional feature fusion methods. The hierarchical baseline transfer learning module constructs sub-baselines for different environmental types, solving the baseline shift problem caused by environmental changes and improving the adaptability of the baseline model to environmental changes. The multi-level anomaly detector combines the environmental constraint anomaly scoring mechanism to realize multi-dimensional evaluation of the psychological state, reducing the false positive rate caused by environmental interference. The artificial intelligence algorithms in the overall scheme are optimized for the specific needs of mental health monitoring, including environmental perception attention mechanisms, multi-modal data graph structure representation, and personalized modeling based on transfer learning. These algorithm features significantly improve the performance and reliability of the scheme in practical applications. Through the synergistic effect of algorithms and models, the problem of mental health monitoring under dynamic environmental changes is effectively solved.

[0157] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, system and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be repeated here.

[0158] The above-described embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the foregoing embodiments of the present application have been described in detail, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An artificial intelligence-based mental health monitoring method, characterized by, The mental health monitoring method based on artificial intelligence comprises: Real-time collection of environmental parameters through a multi-source sensing network, scene recognition processing through an environmental context reasoning engine, and obtaining environmental semantic feature data; According to the environmental semantic feature data, feature extraction and individual difference modeling of multi-dimensional physiological signals are performed through an environmental adaptive signal processing algorithm, and an environment-compensated personalized physiological feature vector is obtained, comprising: Collecting electrocardiogram, electroencephalogram, skin electricity, and respiration data through a physiological signal collection device, calculating time domain features of the electrocardiogram, electroencephalogram, skin electricity, and respiration data, and obtaining original physiological feature data; time-aligning the environmental semantic feature data and the original physiological feature data, compensating for environmental interference in the original physiological feature data through weighted processing, and obtaining compensated physiological features; performing frequency domain decomposition on the compensated physiological features, decomposing electrocardiogram signals into low, medium, and high frequency components, decomposing electroencephalogram signals into alpha, beta, theta, and delta bands, and obtaining multi-band feature data; calculating feature statistics including mean, variance, kurtosis, and skewness from the multi-band feature data, and generating a statistical feature vector; calculating the similarity of the statistical feature vector and historical sample data, extracting individual feature difference parameters, and performing personalized correction on the statistical feature vector to obtain a corrected feature vector; performing dimension conversion and normalization processing on the corrected feature vector through an environmental adaptive signal processing algorithm, and obtaining an environment-compensated personalized physiological feature vector; Using the environmental semantic feature data and the personalized physiological feature vector, performing environment-behavior interaction analysis through a Bayesian causal inference model, and outputting environment-behavior correlation features; Inputting the environment-behavior correlation features and the personalized physiological feature vector into a graph attention network, performing multi-dimensional feature fusion through an environmental condition gating mechanism, and generating dynamic feature representations; According to the dynamic feature representations, constructing sub-baselines for different environment types through a hierarchical baseline transfer learning module, and obtaining an environment-adaptive baseline model; Based on the environment-adaptive baseline model, evaluating the mental state through a multi-level anomaly detector, and outputting a warning signal through an environmental constraint anomaly scoring mechanism; The use of the environmental semantic feature data and the personalized physiological feature vector, the environment-behavior interaction analysis through the Bayesian causal inference model, and the output of the environment-behavior correlation features, comprising: Dividing the environmental semantic feature data into multiple time window segments, extracting features from each time window segment, and obtaining an environment state sequence; According to the environment state sequence, constructing a state transition matrix, calculating the transition probability between environment states through matrix decomposition, and obtaining an environment change pattern; Sliding window segmentation of the personalized physiological feature vector, calculation of statistics and trends in each window, and formation of a physiological state sequence; Using the environment change pattern and the physiological state sequence, calculating the conditional probability distribution through the Bayesian causal inference model, and obtaining an environment-physiology correlation measure; aligning the user activity data and the social behavior data with timestamps with the environment-physiological correlation measure to perform time-series correlation analysis, and generating a behavior response feature; performing path analysis and weight calculation on the behavior response feature through a multi-layer causal diagram, and outputting an environment-behavior correlation feature.

2. The artificial intelligence-based mental health monitoring method of claim 1, wherein, The real-time collection of the environment parameters through the multi-source sensing network, the scene recognition processing through the environment context reasoning engine, and the obtaining of the environment semantic feature data include: receiving temperature, humidity, illumination, noise, and air quality data through the multi-source sensing network, inputting the temperature, humidity, illumination, noise, and air quality data into an environment parameter normalization unit for standardization processing to obtain standardized environment parameters; obtaining pressure, precipitation, and ultraviolet index data from a meteorological data interface, performing time-series alignment processing on the pressure, precipitation, and ultraviolet index data to obtain weather change features; performing multi-source fusion positioning calculation on GPS coordinates, WiFi signal strengths, and Bluetooth beacon data collected by a positioning module to output scene location information; obtaining image streams and audio streams of the surrounding environment by using an image acquisition unit and an acoustic acquisition unit, processing the image streams and the audio streams through scene semantic analysis to obtain social environment features; inputting the standardized environment parameters, the weather change features, the scene location information, and the social environment features into the environment context reasoning engine to perform multi-modal feature fusion and scene semantic reasoning, and obtaining environment semantic feature data; performing structured storage and real-time updating on the environment semantic feature data to generate an environment semantic feature data stream; performing timestamp marking and data verification on the environment semantic feature data through the environment semantic feature data stream, and outputting verified environment semantic feature data. 3.The artificial intelligence-based mental health monitoring method of claim 1, wherein, The environment-behavior correlation feature and the personalized physiological feature vector are input into a graph attention network, multi-dimensional feature fusion is performed through an environment condition gating mechanism, and a dynamic feature representation is generated, including: aligning the environment-behavior correlation feature and the personalized physiological feature vector with timestamps, generating a feature sequence matrix through data completion processing; performing feature node construction on the feature sequence matrix, calculating the connection weight between nodes, and obtaining an initial feature graph structure; scoring the importance of the nodes in the initial feature graph structure by using an attention calculation unit to generate an attention weight distribution; performing gating operation on the attention weight distribution and environment condition information to dynamically adjust the feature connection strength, and obtaining a gated feature graph; performing feature propagation and aggregation on the gated feature graph through a message passing mechanism to generate a node update vector; performing time-series combination and dimension transformation on the node update vector to output a dynamic feature representation. 4.The artificial intelligence-based mental health monitoring method of claim 1, wherein, According to the dynamic feature representation, a hierarchical baseline migration learning module is used to construct a sub-baseline for different environment types to obtain an environment adaptive baseline model, including: performing time series decomposition on the dynamic feature representation to separate long-term features, medium-term features, and short-term features to obtain hierarchical feature sequences; According to historical environmental data, different environmental scenes are classified, an environmental type feature library is established, and an environmental classification standard is formed; The hierarchical feature sequence is grouped according to the environmental classification standard, statistical features of each group of data are calculated, and an environmental related feature set is generated; The environmental related feature set is divided into subspaces, a feature distribution matrix under different environmental types is constructed, and environmental subspace baseline data is obtained; The environmental subspace baseline data is mapped by migration transformation, an associated mapping relationship between environmental types is established, and baseline migration parameters are obtained; The baseline migration parameters and the environmental subspace baseline data are integrated and optimized to output an environmental adaptive baseline model. 5.The artificial intelligence-based mental health monitoring method of claim 1, wherein, Based on the environmental adaptive baseline model, the mental state is evaluated by a multi-level anomaly detector, and a warning signal is output through an environmental constraint anomaly scoring mechanism, including: Calculate the feature distance between real-time mental state data and the environmental adaptive baseline model, and normalize the calculation result to obtain a deviation degree index; According to the deviation degree index, the mental state is evaluated in layers, and the abnormal feature vector is obtained by calculating the instant abnormal score, the trend abnormal score and the mode abnormal score through multi-scale comparison; Calculate the spatio-temporal distribution characteristics of the abnormal state using the abnormal feature vector, and generate an abnormal state distribution map; Correlate the abnormal state distribution map with the environmental conditions, calculate the environmental constraint weight coefficient, and form the constraint scoring parameter; Combine the constraint scoring parameter and the abnormal feature vector by weighting, and generate a warning level identifier through multi-level threshold judgment; According to the mapping relationship between the warning level identifier and the environmental risk degree, a warning signal is output.

6. The artificial intelligence-based mental health monitoring method of claim 5, wherein, According to the deviation degree index, the mental state is evaluated in layers, and the abnormal feature vector is obtained by calculating the instant abnormal score, the trend abnormal score and the mode abnormal score through multi-scale comparison, including: Divide the deviation degree index into short-term sequence, medium-term sequence and long-term sequence according to time window, and perform data smoothing processing on each sequence to obtain multi-scale time series data; Perform sliding window analysis on the short-term sequence in the multi-scale time series data, calculate the difference between the statistical quantity in the window and the baseline threshold, and generate the instant abnormal score; Perform slope calculation on the medium-term sequence in the multi-scale time series data through trend decomposition, perform deviation analysis on the slope change and the baseline trend, and obtain the trend abnormal score; Extract the mode feature from the long-term sequence in the multi-scale time series data, identify the abnormal mode through sequence similarity calculation, and output the mode abnormal score; Combine the instant abnormal score, the trend abnormal score and the mode abnormal score, and convert the dimensions to form an abnormal feature matrix; Perform dimension reduction and normalization processing on the abnormal feature matrix to generate an abnormal feature vector.

7. The artificial intelligence-based mental health monitoring method of claim 6, wherein, The slope calculation on the medium-term sequence in the multi-scale time series data through trend decomposition, the deviation analysis on the slope change and the baseline trend, and the trend abnormal score, include: The medium-term sequence in the multi-scale time series data is segmented according to a fixed step length, statistical feature extraction is performed on each segment of data, and sequence feature points are obtained; The sequence feature points are subjected to seasonal decomposition, and trend components, periodic components and random components are separated to obtain trend sequence data; The trend sequence data is linearly fitted by using the least square method, the slope value in each time period is calculated, and a trend slope sequence is formed; The trend slope sequence is subjected to difference calculation with baseline trend data, a slope deviation matrix is established, and trend change characteristics are obtained; The trend change characteristics are aggregated in the time dimension by weighted average to generate a trend deviation vector; The trend deviation vector is subjected to amplitude normalization and threshold mapping, and a trend anomaly score is output.

8. An artificial intelligence-based mental health monitoring system for implementing the artificial intelligence-based mental health monitoring method according to any one of claims 1-7, characterized by, The mental health monitoring system based on artificial intelligence comprises: A collection module is configured to collect environmental parameters in real time through a multi-source sensing network, perform scene recognition processing through an environmental context reasoning engine, and obtain environmental semantic feature data. A modeling module is configured to extract features of multi-dimensional physiological signals and model individual differences according to the environmental semantic feature data through an environmental adaptive signal processing algorithm, and obtain an environmental-compensated personalized physiological feature vector. An analysis module is configured to perform environmental-behavior interaction analysis through a Bayesian causal inference model using the environmental semantic feature data and the personalized physiological feature vector, and output environmental-behavior correlation features. An input module is configured to input the environmental-behavior correlation features and the personalized physiological feature vector into a graph attention network, perform multi-dimensional feature fusion through an environmental condition gating mechanism, and generate dynamic feature representations. A construction module is configured to construct sub-baselines for different environmental types through a hierarchical baseline transfer learning module according to the dynamic feature representations, and obtain an environmental adaptive baseline model. An evaluation module is configured to evaluate mental states through a multi-level anomaly detector based on the environmental adaptive baseline model, and output early warning signals through an environmental constraint anomaly scoring mechanism.

Images