Method for constructing individual emotion baseline self-adaptively based on long-term physiological and activity data

Abstract

The present application relates to a method for constructing an individual emotion baseline based on long-term physiological and activity data, belonging to the field of artificial intelligence and digital health technology. The method comprises the following steps: S1, acquiring multi-modal physiological and activity data of the user, completing time alignment and cleaning, storing by time granularity layering, and generating an individual multi-modal data set; S2, traversing the data stream through a sliding window, detecting physiological feature mutations and behavior deviations, identifying abnormal disturbance scenarios and marking the data segment, and separating the steady-state reference data; S3, constructing a multi-time span emotion baseline based on the steady-state data, establishing a dynamic correlation model, dynamically allocating update weights, and iteratively outputting an optimized baseline; S4, constructing a physiological behavior feature coupling network, mining the mapping relationship between features and emotions, generating an emotion reference atlas, and outputting a multi-dimensional dynamic baseline result. The present application realizes the adaptive construction of an individual exclusive emotion baseline, avoids the pollution of abnormal disturbances to the baseline, and improves the individualization adaptability and accuracy of emotion evaluation.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and digital health technology, specifically involving an adaptive method for constructing an individual emotional baseline based on long-term physiological and activity data. Background Technology

[0002] With the rapid popularization of digital health technology and smart wearable devices, emotional state monitoring and mental health assessment technologies based on physiological and behavioral data have been widely used. These technologies collect various physiological indicators and daily behavioral data of users, and combine them with preset analysis models to judge users' emotional state, stress level and fatigue level, providing users with health guidance and risk warnings. This has become an important development direction in the field of digital health.

[0003] Currently, most mainstream emotion state analysis technologies use general thresholds or fixed rule models based on group statistics for data processing. These technologies do not fully consider the significant differences in physiological basis and behavioral patterns among different individuals. The same physiological indicator changes may correspond to completely different emotional states in different individuals, leading to frequent misjudgments in monitoring results and failing to meet the needs of personalized health monitoring. At the same time, users' physiological characteristics and behavioral habits will continuously change with their age, lifestyle adjustments, environmental changes, and health status evolution. The fixed baseline models used in existing technologies do not have the ability to be dynamically updated. As the usage time increases, the deviation between the model and the user's actual state will gradually increase, and the monitoring accuracy will continue to decline.

[0004] Furthermore, existing technologies cannot effectively distinguish between normal trend changes and temporary abnormal disturbances in users' physiological and behavioral data. Users' physiological and behavioral data are affected by various temporary factors, resulting in temporary fluctuations. Existing technologies do not effectively identify and distinguish these fluctuations, and directly include all data in the baseline calculation process, which can easily cause abnormal baseline drift, thereby affecting the accuracy and reliability of subsequent sentiment analysis results. More importantly, most existing technologies only focus on a single type of physiological indicator, ignoring the intrinsic correlation and synergistic effect between physiological state and daily behavior. They fail to fully explore the emotion-related information contained in multimodal data, making it difficult to construct a comprehensive and accurate reference model that reflects individual emotional characteristics. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides an adaptive method for constructing an individual emotional baseline based on long-term physiological and activity data. The objective of this invention can be achieved through the following technical solutions: include: S1: Acquire multimodal physiological and activity data, perform time series alignment and format verification, write the data to the storage module in layers according to preset time granularity, and generate individual multimodal datasets; S2: Based on the individual multimodal dataset, the sliding window algorithm is called to traverse the data stream; extract the physiological feature mutation threshold and behavioral pattern deviation, match abnormal disturbance situations and output situation state labels, mark abnormal data segments and separate steady-state baseline data; S3: Instantiate individual emotion baseline objects across multiple time spans based on steady-state baseline data and configure a dynamic correlation model; determine steady-state data segments and abnormal data segments based on preset thresholds; after identifying abnormal disturbances, perform graded decay processing on the baseline update weights and dynamically allocate update learning weights; perform adaptive iterative calculation of individual emotion baselines and output an optimized baseline model; S4: Construct a multidimensional dynamic correlation network of physiological and behavioral features by optimizing the baseline model, perform feature combination and emotional state mapping calculation, and output an individual emotional reference feature map; link the optimized baseline model to calculate the dynamic baseline multidimensional representation parameter set, perform parameter normalization and vectorization processing, and generate the final baseline result set.

[0006] Specifically, the process of generating the individual multimodal dataset is as follows: The system receives raw signals from multiple sensors, performs timestamp alignment, then performs missing value interpolation and noise filtering, and archives the cleaned data in layers according to daily and hourly granularity. Establish a unified data dictionary and index structure, perform field type conversion and memory mapping, and finally encapsulate it into a standardized format structured time-series dataset, merge it into the persistent storage module, generate standardized data archive files, and write them to the persistent storage medium.

[0007] Specifically, the process of calling the sliding window algorithm to traverse the data stream is as follows: Initialize a fixed-length time window and a preset step size parameter, load the structured time-series data stream into a memory buffer, extract continuous data segments sequentially according to the step size, and perform boundary smoothing processing on each segment; Record the start and end position indices of the window, and proceed cyclically until the entire data cycle is covered, forming a continuous overlapping local data sequence, which forms a segmented data stream and is passed to the feature extraction node.

[0008] Specifically, the process of extracting the physiological feature mutation threshold and the deviation from the behavioral pattern is as follows: Frequency domain transformation and statistical feature calculations were performed on local data sequences to obtain the standard deviation distribution of heart rate variability and skin conductance response. Calculate the dynamic confidence interval by combining historical normal distribution curves, and mark the points that exceed the extreme values ​​of the interval as the mutation threshold; Simultaneously, the offset between action frequency and duration is calculated by comparing the behavior log sequence, generating a quantization deviation index, which is then smoothed, encapsulated into a quantization feature vector, and sent to the feature processing bus.

[0009] Specifically, the process of marking abnormal data segments and separating steady-state reference data is as follows: Input the mutation threshold and deviation index into the classifier, compare it with the preset situation rule base, identify data segments containing drastic fluctuations or pattern breaks, and attach an abnormal state identifier. Extract consecutive segments that do not match the exception rules into a candidate set, perform time continuity verification and integrity filtering, remove fragmented records, filter out abnormal data blocks, and retain normal baseline data.

[0010] Specifically, the process of instantiating the multi-time-span individual emotion baseline object is as follows: Read the steady-state baseline data stream, parse the timestamp range and divide it into short, medium and long-term dimensions, allocate independent memory space for each dimension, and initialize the baseline parameter container and metadata descriptor; By loading historical statistical distributions as prior knowledge, a baseline template containing the mean trajectory and variance boundary is constructed; Perform object lifecycle registration and return instance handles for downstream modules to call, register baseline object instances and expose call interfaces.

[0011] Specifically, the process of configuring the dynamic association model is as follows: Parse the dimensional attributes of the baseline object instance, establish a parameter mapping table across time scales, initialize the state transition matrix and weight decay coefficient, and define the information transmission path and synchronization triggering conditions between baselines. Load the time-series dependency graph structure, perform topological sorting to eliminate circular references, perform model parameter binding and runtime environment configuration, and deploy the associated computing engine to the runtime environment.

[0012] Specifically, the process of determining steady-state data segments and abnormal data segments based on preset thresholds is as follows: The system acquires real-time incoming monitoring data streams, extracts the feature vector of the current time window, inputs it into the judgment engine, compares it dimension by dimension with the preset threshold boundary, and calculates the Euclidean distance and Mahalanobis distance scores. If the score is lower than the preset threshold, it is marked as a steady-state data segment; if it exceeds the preset threshold, it is marked as an abnormal data segment. A classification result label is generated and written to the status register. The status register is updated and the real-time classification result is recorded.

[0013] Specifically, the process of optimizing the baseline model through output is as follows: Collect weight update logs during iterative calculation, perform gradient accumulation and momentum correction, remove outlier parameters that deviate from the main distribution, and perform regularization constraint processing on the remaining parameters. Recalculate the baseline center trajectory and confidence interval boundaries, package the converged parameter set into a binary model file, attach a version identifier and checksum, persist it to the model repository and broadcast update notifications, and solidify the iteratively optimized baseline model version to the model repository.

[0014] Specifically, the process of constructing the multidimensional dynamic correlation network of physiological and behavioral characteristics is as follows: Extract the feature nodes from the optimized baseline model output, initialize the graph database instance, define the node type and edge weight attributes, and calculate the coupling strength between features based on the mutual information coefficient and Pearson correlation coefficient. Generate an adjacency matrix and construct an undirected graph topology. Execute a community detection algorithm to divide feature clusters. Perform network initialization and register a dynamic update callback interface. Construct a multi-dimensional feature dynamic association topology and load it into the graph database.

[0015] Specifically, the process of outputting the individual emotion reference feature map is as follows: Traverse the node and edge relationships in the dynamic association network, extract weighted connection paths and core feature clusters, execute the graph layout algorithm to optimize the spatial distribution of nodes, and render the feature hierarchy structure and interaction intensity heatmap. Generate a structured description file containing node attributes and edge weights, add timestamps and user identifiers, export it as visual graph data for the front-end rendering engine to parse, and encapsulate the output results in vector graphics format.

[0016] Specifically, the process of generating the final baseline result set is as follows: Receive the optimized baseline model and sentiment reference feature map, extract the multidimensional representation parameter set, perform normalization scaling and principal component projection transformation, and map the heterogeneous parameters to a unified vector space; Calculate the baseline stability score and deviation index, fuse the anomaly probability distribution and resilience coefficient, assemble them into a structured result array, perform integrity verification and write them into the output buffer to generate a standardized result array; The standardized result array is written to the output buffer, and redundant fields are compressed using a columnar storage format to generate a standardized data stream.

[0017] The beneficial effects of this invention are as follows: (1) By setting up a multimodal time series data hierarchical storage, abnormal disturbance situation identification and multi-time span baseline adaptive update mechanism, the collected physiological and activity data are standardized, cleaned and archived, abnormal disturbance data segments are accurately identified and separated, baseline update weights are dynamically allocated according to the stability of the data and iterative optimization is performed, the interference of abnormal fluctuations on the baseline model is effectively avoided, the individual emotional baseline is continuously self-adapted and adjusted, and the individual adaptability and long-term running accuracy of the baseline model are greatly improved. (2) By setting up a multidimensional association network of physiological and behavioral features and a dynamic baseline multidimensional representation output mechanism, the intrinsic coupling relationship between physiological and behavioral features is explored and an emotional reference feature map is constructed. Combined with the optimized baseline model, multidimensional baseline result parameters are generated to comprehensively reflect the comprehensive characteristics of individual emotional state, effectively improve the comprehensiveness and accuracy of emotional state assessment, and provide reliable underlying data support for subsequent emotional prediction and health intervention. Attached Figure Description

[0018] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0019] Figure 1 This is a flowchart illustrating the adaptive construction method for individual emotional baselines based on long-term physiological and activity data according to the present invention. Figure 2 This is a data flow diagram of the adaptive construction method for individual emotional baselines based on long-term physiological and activity data according to the present invention. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0021] Please see Figures 1-2 An adaptive method for constructing individual emotional baselines based on long-term physiological and activity data; include: S1: Acquire multimodal physiological and activity data, perform time series alignment and format verification, write the data to the storage module in layers according to preset time granularity, and generate individual multimodal datasets; S2: Based on the individual multimodal dataset, the sliding window algorithm is called to traverse the data stream; extract the physiological feature mutation threshold and behavioral pattern deviation, match abnormal disturbance situations and output situation state labels, mark abnormal data segments and separate steady-state baseline data; S3: Instantiate individual emotion baseline objects across multiple time spans based on steady-state baseline data and configure a dynamic correlation model; determine steady-state data segments and abnormal data segments based on preset thresholds; after identifying abnormal disturbances, perform graded decay processing on the baseline update weights and dynamically allocate update learning weights; perform adaptive iterative calculation of individual emotion baselines and output an optimized baseline model; S4: Construct a multidimensional dynamic correlation network of physiological and behavioral features by optimizing the baseline model, perform feature combination and emotional state mapping calculation, and output an individual emotional reference feature map; link the optimized baseline model to calculate the dynamic baseline multidimensional representation parameter set, perform parameter normalization and vectorization processing, and generate the final baseline result set.

[0022] In this embodiment, multimodal physiological and activity data refers to a collection of various time-series data collected from an individual that can reflect their physiological state and daily activity state, specifically including: heart rate data, heart rate variability data, respiratory rate data, skin temperature data, sleep structure data, resting heart rate data, nighttime wakefulness data, circadian rhythm data, daily activity data, sedentary duration data, exercise intensity data, device usage behavior data, social interaction frequency data, geographic activity range data, and schedule change data; Physiological characteristic mutation thresholds refer to the critical values ​​that can determine whether a physiological indicator has undergone significant abnormal fluctuations relative to an individual's normal distribution. Specifically, they include: heart rate mutation threshold, heart rate variability mutation threshold, respiratory rate mutation threshold, skin temperature mutation threshold, sleep structure mutation threshold, and resting heart rate fluctuation threshold. Behavioral pattern deviation refers to an indicator that quantifies the degree of difference between current behavioral performance and an individual's historical normal behavioral patterns. Specifically, it includes: daily activity level deviation, sedentary time deviation, exercise intensity deviation, device usage behavior deviation, social interaction frequency deviation, geographical activity range deviation, and schedule change deviation. Abnormal disturbances refer to specific scenarios that can cause temporary fluctuations in an individual's physiological and behavioral data and disrupt the baseline stability. These include: sleep deprivation, high-intensity exercise, disease recovery, physiological cycle changes, irritant effects, time zone switching, and high-stress work. Physiological and behavioral characteristics refer to a set of quantifiable features extracted from physiological and behavioral data that can reflect an individual's emotional state. Specifically, these include: heart rate fluctuation characteristics, heart rate variability distribution characteristics, respiratory rhythm characteristics, sleep structure characteristics, activity intensity characteristics, sedentary duration characteristics, social frequency characteristics, device usage duration characteristics, and geographical activity range characteristics. Multi-source sensors refer to acquisition units deployed at different acquisition terminals that can acquire different types of physiological and behavioral signals. Specifically, they include: heart rate acquisition sensors, skin conductance acquisition sensors, body temperature acquisition sensors, acceleration acquisition sensors, position acquisition sensors, and device usage status acquisition units. The dimensional attributes of a baseline object instance refer to descriptive information that can identify the classification features and operational status of baseline objects across different time spans. Specifically, these include: baseline time span attributes, baseline parameter boundary attributes, baseline update status attributes, baseline convergence attributes, and baseline lifecycle attributes. Dynamic association model parameters refer to adjustable computational parameters that can support information transmission and state linkage between baselines with different time spans. Specifically, they include: state transition matrix parameters, weight decay coefficient parameters, information transmission path parameters, synchronization trigger threshold parameters, and time-dependent weight parameters. Core features refer to the set of key features that have the strongest association with emotional state in the physiological behavior association network and play a dominant role in baseline determination. Specifically, they include: resting heart rate features, heart rate variability features, sleep quality features, daily activity level features, social interaction features, and stress recovery features.

[0023] Specifically, the process of generating the individual multimodal dataset is as follows: The system receives raw signals from multiple sensors, performs timestamp alignment, then performs missing value interpolation and noise filtering, and archives the cleaned data in layers according to daily and hourly granularity. Establish a unified data dictionary and index structure, perform field type conversion and memory mapping, and finally encapsulate it into a standardized format structured time-series dataset, merge it into the persistent storage module, generate standardized data archive files, and write them to the persistent storage medium.

[0024] Specifically, the process of calling the sliding window algorithm to traverse the data stream is as follows: Initialize a fixed-length time window and a preset step size parameter, load the structured time-series data stream into a memory buffer, extract continuous data segments sequentially according to the step size, and perform boundary smoothing processing on each segment; Record the start and end position indices of the window, and proceed cyclically until the entire data cycle is covered, forming a continuous overlapping local data sequence, which forms a segmented data stream and is passed to the feature extraction node.

[0025] In this embodiment, a typical configuration example for supplementing key operating parameters based on actual application scenarios is as follows: Key parameters of the sliding window: In a typical emotion monitoring scenario, the typical length of the time window is 30 minutes and the typical sliding step size is 5 minutes; for a long-term baseline steady-state verification scenario, the long analysis window length can be configured to 24 hours and the sliding step size to 1 hour; the window boundary smoothing adopts the Hanning window function and the detrending processing adopts the first-order difference method.

[0026] Statistical threshold parameters: The typical confidence level of the dynamic confidence interval for physiological feature mutation detection is 95%, corresponding to 2 times the standard deviation boundary; the threshold for Mahalanobis distance anomaly detection is the 95th percentile value under the corresponding feature dimension; the typical subsampling size for isolated forest anomaly detection is 256, and the typical number of base decision trees is 100.

[0027] Weight update parameters: The typical value of the exponential forgetting factor is 0.995; the typical values ​​of the weight decay coefficients corresponding to the three levels of anomalies are: 0.8 for mild anomalies, 0.3 for moderate anomalies, and 0 for severe anomalies; the typical value of the weight normalization period is 24 hours.

[0028] Baseline parameters for multiple time spans: intraday baselines cover a 24-hour period with a time granularity of 1 hour; weekly baselines cover a 7-30 day period; quarterly baselines cover a 90-365 day period; baseline confidence boundaries are uniformly set at ±2 standard deviations.

[0029] Network construction parameters: In the calculation of feature coupling strength, the mutual information coefficient accounts for 0.6% of the weight, and the Pearson correlation coefficient accounts for 0.4% of the weight; the typical value of the weak connection elimination threshold is 0.1; the typical value of the modularity threshold found by the Louvain community is 0.3.

[0030] Specifically, the process of extracting the physiological feature mutation threshold and the deviation from the behavioral pattern is as follows: Frequency domain transformation and statistical feature calculations were performed on local data sequences to obtain the standard deviation distribution of heart rate variability and skin conductance response. Calculate the dynamic confidence interval by combining historical normal distribution curves, and mark the points that exceed the extreme values ​​of the interval as the mutation threshold; Simultaneously, the offset between action frequency and duration is calculated by comparing the behavior log sequence, generating a quantization deviation index, which is then smoothed, encapsulated into a quantization feature vector, and sent to the feature processing bus.

[0031] Specifically, the process of marking abnormal data segments and separating steady-state reference data is as follows: Input the mutation threshold and deviation index into the classifier, compare it with the preset situation rule base, identify data segments containing drastic fluctuations or pattern breaks, and attach an abnormal state identifier. Extract consecutive segments that do not match the exception rules into a candidate set, perform time continuity verification and integrity filtering, remove fragmented records, filter out abnormal data blocks, and retain normal baseline data.

[0032] Specifically, the process of instantiating the multi-time-span individual emotion baseline object is as follows: Read the steady-state baseline data stream, parse the timestamp range and divide it into short, medium and long-term dimensions, allocate independent memory space for each dimension, and initialize the baseline parameter container and metadata descriptor; By loading historical statistical distributions as prior knowledge, a baseline template containing the mean trajectory and variance boundary is constructed; Perform object lifecycle registration and return instance handles for downstream modules to call, register baseline object instances and expose call interfaces.

[0033] Specifically, the process of configuring the dynamic association model is as follows: Parse the dimensional attributes of the baseline object instance, establish a parameter mapping table across time scales, initialize the state transition matrix and weight decay coefficient, and define the information transmission path and synchronization triggering conditions between baselines. Load the time-series dependency graph structure, perform topological sorting to eliminate circular references, perform model parameter binding and runtime environment configuration, and deploy the associated computing engine to the runtime environment.

[0034] Specifically, the process of determining steady-state data segments and abnormal data segments based on preset thresholds is as follows: The system acquires real-time incoming monitoring data streams, extracts the feature vector of the current time window, inputs it into the judgment engine, compares it dimension by dimension with the preset threshold boundary, and calculates the Euclidean distance and Mahalanobis distance scores. If the score is lower than the preset threshold, it is marked as a steady-state data segment; if it exceeds the preset threshold, it is marked as an abnormal data segment. A classification result label is generated and written to the status register. The status register is updated and the real-time classification result is recorded.

[0035] Specifically, the process of optimizing the baseline model through output is as follows: Collect weight update logs during iterative calculation, perform gradient accumulation and momentum correction, remove outlier parameters that deviate from the main distribution, and perform regularization constraint processing on the remaining parameters. Recalculate the baseline center trajectory and confidence interval boundaries, package the converged parameter set into a binary model file, attach a version identifier and checksum, persist it to the model repository and broadcast update notifications, and solidify the iteratively optimized baseline model version to the model repository.

[0036] Specifically, the process of constructing the multidimensional dynamic correlation network of physiological and behavioral characteristics is as follows: Extract the feature nodes from the optimized baseline model output, initialize the graph database instance, define the node type and edge weight attributes, and calculate the coupling strength between features based on the mutual information coefficient and Pearson correlation coefficient. Generate an adjacency matrix and construct an undirected graph topology. Execute a community detection algorithm to divide feature clusters. Perform network initialization and register a dynamic update callback interface. Construct a multi-dimensional feature dynamic association topology and load it into the graph database.

[0037] Specifically, the process of outputting the individual emotion reference feature map is as follows: Traverse the node and edge relationships in the dynamic association network, extract weighted connection paths and core feature clusters, execute the graph layout algorithm to optimize the spatial distribution of nodes, and render the feature hierarchy structure and interaction intensity heatmap. Generate a structured description file containing node attributes and edge weights, add timestamps and user identifiers, export it as visual graph data for the front-end rendering engine to parse, and encapsulate the output results in vector graphics format.

[0038] Specifically, the process of generating the final baseline result set is as follows: Receive the optimized baseline model and sentiment reference feature map, extract the multidimensional representation parameter set, perform normalization scaling and principal component projection transformation, and map the heterogeneous parameters to a unified vector space; Calculate the baseline stability score and deviation index, fuse the anomaly probability distribution and resilience coefficient, assemble them into a structured result array, perform integrity verification and write them into the output buffer to generate a standardized result array; The standardized result array is written to the output buffer, and redundant fields are compressed using a columnar storage format to generate a standardized data stream.

[0039] In this embodiment, the specific process of generating an individual multimodal dataset is as follows: receiving raw physiological and activity signals transmitted from multiple acquisition terminals, performing a unified timestamp alignment operation with an accuracy of 1ms, resampling all signals to a sampling rate of 1Hz, using linear interpolation to imputate segments with a missing rate of less than 20%, directly removing segments with a missing rate exceeding the standard, removing high-frequency noise from the signal through a third-order Butterworth low-pass filter, archiving the cleaned data in layers according to three time granularities of hourly, daily, and weekly levels, building a unified data dictionary and field-level clustered index structure, performing field type conversion and memory mapping processing, and finally encapsulating it into a structured time-series data set in a standardized columnar storage format, and writing it to a persistent storage medium to form a standardized data archive file; The specific process of calling the sliding window algorithm to traverse the data stream is as follows: initialize a time window with a length of 30 minutes and a sliding step size of 5 minutes, load the structured time series data stream into the memory buffer on a daily basis, extract continuous data segments sequentially according to the set step size, use the Hanning window function to perform smoothing processing on the beginning and end boundaries of each data segment to suppress spectral leakage, simultaneously perform detrending processing on the data within the window to eliminate baseline drift, record the start and end position indices of each window segment, and cyclically advance the window until it covers the entire data period, forming a continuous overlapping local data sequence, generating a segmented data stream and passing it to the feature extraction node; The specific process for extracting the physiological characteristic mutation threshold and behavioral pattern deviation is as follows: Fast Fourier frequency domain transform and time domain statistical feature calculation are performed on the segmented local data sequence to obtain the standard deviation of all sinus intervals of heart rate variability, the root mean square of the difference between adjacent intervals, and the power distribution in the low-frequency and high-frequency bands of the frequency domain. Simultaneously, the standard deviation distribution of skin conductance is calculated. A sliding kernel density estimation method is used to fit the user's historical normal distribution curve window by window, and a 95% confidence interval is calculated. Combined with cumulative sum and mutation detection algorithms to assist in the judgment, points exceeding the upper and lower extreme values ​​of the interval are marked as physiological characteristic mutation thresholds. At the same time, the offset of action frequency and duration is calculated by comparing with the behavioral log sequence, and the change in behavioral temporal entropy is supplemented. A quantified deviation index is generated using the standard score method, and after smoothing by moving average, it is encapsulated into a quantified feature vector and transmitted to the feature processing bus. In this embodiment, the specific process of marking abnormal data segments and separating steady-state baseline data is as follows: The physiological mutation threshold and behavioral deviation index are input into the classifier. The isolated forest algorithm is combined with the rule engine matching method. The isolated forest is set with 100 base decision trees and a sub-sampling scale of 256. The rule base is configured with feature matching thresholds for 7 types of disturbance scenarios. The two types of detection results are fused using a majority voting mechanism. Data segments containing violent fluctuations or pattern breaks are identified and abnormal state identifiers are attached. A buffer zone of 10 minutes is reserved before and after the abnormal segment and removed to avoid residual abnormal data contamination at the boundary. The segments that do not hit the abnormal rules are extracted as a candidate set. Time continuity verification is performed. Complete segments with a continuous duration of not less than 4 hours and no abnormal markings are retained. After removing fragmented abnormal records, abnormal data blocks are filtered to obtain normal steady-state baseline data. The specific process of instantiating multi-time-span individual sentiment baseline objects is as follows: read the steady-state baseline data stream, parse the timestamp range of the data and divide it into three time dimensions: daily, weekly, and quarterly. Allocate independent parameter runtime memory space for each dimension and initialize the baseline parameter container and metadata descriptor. The daily baseline adopts an hourly segmented mean model, the weekly baseline adopts an exponentially weighted moving average model, and the quarterly baseline adopts a Gaussian process regression model to fit the baseline trajectory. Load the user's previously accumulated historical statistical distribution as prior knowledge. In the cold start phase, a general group distribution is used as the initial value, which is gradually replaced with the individual-specific distribution as data accumulates. Construct a baseline template that includes the mean trajectory and upper and lower 2 standard deviation boundaries. Perform object lifecycle registration and bind an independent version number and update timestamp. Open the calling interface for downstream modules to call. The specific process of state stability assessment and dynamic weight allocation update is as follows: Acquire the real-time incoming monitoring data stream, extract the multi-dimensional feature vector X of the current time window, input it into the judgment engine, and compare it dimension-by-dimensionally with the preset threshold boundary. Calculate the Euclidean distance and Mahalanobis distance scores respectively. The Mahalanobis distance anomaly judgment formula is as follows: , in, This is the mean vector of historical steady-state characteristics. The corresponding covariance matrix is ​​used; the Mahalanobis distance corresponding to twice the standard deviation is used as the steady-state determination threshold. If the data is below the threshold, it is marked as a steady-state data segment; if it exceeds the threshold, it is marked as an abnormal data segment. A classification result label is generated and written to the status register. An exponential forgetting factor λ with a value of 0.995 is introduced to gradually reduce the weight of long-term steady-state data. A momentum term is added to the weight update to smooth weight fluctuations and avoid jumps. The recursive formula for weight decay in the t-th window is as follows: , In the formula As the weight of the previous window, This is the weight decay coefficient corresponding to the anomaly level. The current window's state matching degree is determined; the anomaly level is divided into three levels: mild anomaly corresponds to a weight decay coefficient of 0.8, moderate anomaly corresponds to a decay coefficient of 0.3, and severe anomaly has its weight reset to zero and the baseline update of that data segment is paused; weight normalization is performed once every 24-hour data cycle to ensure the stability of the weight range; The specific process of performing graded decay processing on baseline update weights is as follows: Based on the steady-state and anomaly judgment results obtained from the state stability assessment, and combined with the degree of influence of the anomaly disturbance scenario, the abnormal data segments are divided into three levels; mild anomalies correspond to a weight decay coefficient of 0.8, and the update weight of the corresponding data segment is reduced proportionally, while retaining part of its baseline iteration contribution; moderate anomalies correspond to a weight decay coefficient of 0.3, which significantly reduces the weight ratio of the data segment, allowing only minor corrections to the baseline parameters; severe anomalies directly reset the update weight of the corresponding data segment to zero, completely shielding its impact on baseline iteration; the weight decay results at each level are synchronously input into the dynamic correlation model, and linked updates are performed between short-term, medium-term, and long-term baselines to achieve synchronous correction of baseline parameters at different time scales; an exponential forgetting factor with a value of 0.995 is introduced in the weight calculation process to perform smooth decay on long-term steady-state data, while a momentum term is added to smooth weight fluctuations and avoid jumps; full weight normalization is performed once every 24-hour data cycle to maintain the stability of the weight value range; The specific process of constructing a multidimensional dynamic association network of physiological and behavioral features is as follows: 12 core physiological and behavioral features output from the optimized baseline model are extracted as network nodes. The nodes are divided into two categories: physiological features and behavioral features. Graph storage instances are initialized and node types and edge weight attributes are defined. Mutual information coefficients and Pearson correlation coefficients are calculated separately. The features are weighted and fused with weights of 0.6 and 0.4 to obtain the coupling strength between features as edge weight values. An adjacency matrix is ​​generated and an undirected graph topology is constructed. The Louvain community detection algorithm is used to divide feature clusters. The modularity threshold is set to 0.3. Graph sparsity processing is performed to remove weak connections with weights below 0.1 to reduce computational redundancy. Network initialization is performed and a dynamic update callback interface is registered so that the association network recalculates edge weights synchronously with each update of the baseline model. The specific process for generating the final baseline result set is as follows: Two types of inputs are received: an optimized baseline model and an emotion reference feature map. Multiple multidimensional representation parameters are extracted, including baseline stability, physiological deviation, behavioral deviation, probability of emotional abnormality, recovery ability index, and baseline reliability. An extreme value normalization method is used to map all parameters to the 0-1 interval. Principal component analysis is used to perform a projection transformation, retaining principal components with a cumulative variance contribution rate of no less than 85%, thus mapping heterogeneous parameters to a unified vector space. The baseline stability score is calculated based on the baseline fluctuation variance over seven consecutive days, using the following formula: , In the formula, T represents the number of days in the statistical period. Let be the baseline center value on day i. The score represents the mean of the baseline center over the period; a higher score indicates better long-term baseline stability. The probability of emotional abnormality is calculated using a combination of log-likelihood ratio and Mahalanobis distance. Based on the assumption of a multivariate normal distribution, the probability of abnormality corresponding to the current feature vector is: , In the formula, The Mahalanobis distance between the current feature and the steady-state baseline is given by d, where d is the feature dimension. A probability value closer to 1 indicates a higher risk of emotional abnormality. The resilience index is obtained by fitting the rate of regression to the baseline after deviation, and the calculation formula is as follows: , In the formula, This represents the maximum deviation of the feature from the baseline during the abnormal period. The index represents the time it takes for the emotion to fall back from the peak to the normal baseline range. A higher index indicates that the emotion recovers from the abnormal state to a steady state faster. The data is assembled into a structured result array, written to the output buffer after integrity verification, and redundant fields are compressed using a columnar storage format to generate a standardized output data stream.

[0040] Example 1: Taking a smart wearable user's daily emotional health monitoring scenario as an example, the specific process is as follows: Physiological data such as heart rate, heart rate variability, skin temperature, and sleep structure continuously output by wearable devices, as well as behavioral data such as daily activity level, sedentary time, and exercise intensity, are collected. A unified timestamp alignment operation is performed, and the data is cleaned by imputation of missing values ​​and low-pass filtering. The data is then archived and stored in a hierarchical manner according to multi-level time granularity to generate a standardized individual multimodal time series dataset. Initialize a sliding window to traverse the dataset, perform frequency domain transformation and time domain feature calculation on the data within the window, fit the historical normal distribution and calculate the dynamic confidence interval, extract the physiological feature mutation threshold and behavioral pattern deviation, match disturbance scenarios such as sleep deprivation, high-intensity exercise, and disease recovery and generate labels, mark abnormal data segments and separate the steady-state baseline data. Based on steady-state baseline data, emotion baselines for multiple time spans are constructed. Cross-scale parameter mapping and state transition association models are configured. An exponential forgetting factor is introduced. Weights are dynamically allocated and updated according to the state stability assessment results. The weights of steady-state data segments are increased, while the weights of abnormal data segments are reduced according to the corresponding decay coefficients. The optimized individual emotion baseline model is output after iterative calculation. A multidimensional dynamic correlation network is constructed using the core features of the optimized baseline model output as nodes. The coupling strength between features is calculated and feature clusters are divided to generate an individual emotion reference feature map. Further calculations are made of multidimensional representation parameters such as baseline stability, physiological deviation, probability of emotional abnormality, and recovery ability index. After normalization, a standardized baseline result set is output for daily emotional state assessment and abnormal early warning.

[0041] Example 2: Taking a company's long-term employee mental health management scenario as an example, the specific process is as follows: Collect physiological data recorded by employees' wearable devices, combine it with behavioral data such as working hours, schedule density, and device usage frequency recorded by the office system, complete unified sampling and timestamp alignment, and after noise filtering and outlier removal, store it in a multi-level time granularity layer to generate the corresponding individual multimodal dataset for each employee. The sliding window mechanism is used to analyze the time series data stream segment by segment. The isolated forest combined with the rule engine is used to detect abrupt changes in physiological indicators and deviations in behavioral patterns. Abnormal disturbances such as high-pressure work, working overtime, and cross-time zone business trips are identified. The buffer zones before and after the abnormal segments are removed. After time continuity verification, the steady-state reference data that meets the requirements is extracted. Based on steady-state baseline data, instantiate multi-time-span sentiment baseline objects, load historical statistical distribution as prior knowledge to construct an initial baseline template, establish a dynamic correlation model between baselines at different time scales, distinguish steady-state and abnormal states with a set judgment threshold, dynamically adjust the baseline update weights according to the real-time status of employees, perform weight normalization periodically, and complete the adaptive iterative optimization of the baseline. We construct a coupled correlation network between physiological characteristics and work behavior characteristics, explore the mapping relationship between feature combinations and emotional stress state, generate an individual emotional reference feature map, calculate multi-dimensional parameters such as baseline credibility, stress deviation degree, and risk of emotional abnormality, and retain principal components with high variance contribution rate to generate the final baseline results, providing data support for employee mental health assessment and graded intervention.

[0042] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An adaptive method for constructing an individual emotional baseline based on long-term physiological and activity data, characterized in that, include: S1: Acquire multimodal physiological and activity data, perform time series alignment and format verification, write the data to the storage module in layers according to preset time granularity, and generate individual multimodal datasets; S2: Based on the individual multimodal dataset, the sliding window algorithm is used to traverse the data stream; Extract physiological feature mutation thresholds and behavioral pattern deviations, match abnormal perturbation scenarios and output scenario state labels, mark abnormal data segments and separate steady-state baseline data; S3: Instantiate individual emotion baseline objects across multiple time spans based on steady-state baseline data and configure a dynamic correlation model; determine steady-state data segments and abnormal data segments based on preset thresholds; perform graded decay processing on baseline update weights after identifying abnormal disturbances, and dynamically allocate update learning weights. Perform adaptive iterative calculation of individual emotion baselines and output an optimized baseline model; S4: By optimizing the baseline model, a multidimensional dynamic correlation network of physiological and behavioral features is constructed, feature combination and emotional state mapping calculations are performed, and an individual emotional reference feature map is output. The optimized baseline model is used to calculate the dynamic baseline multidimensional representation parameter set, and parameter normalization and vectorization are performed to generate the final baseline result set.

2. The method according to claim 1, characterized in that, The specific process for generating the individual multimodal dataset is as follows: The system receives raw signals from multiple sensors, performs timestamp alignment, then performs missing value interpolation and noise filtering, and archives the cleaned data in layers according to daily and hourly granularity. Establish a unified data dictionary and index structure, perform field type conversion and memory mapping, and finally encapsulate it into a standardized format structured time-series dataset, merge it into the persistent storage module, generate standardized data archive files, and write them to the persistent storage medium.

3. The method according to claim 1, characterized in that, The specific process of invoking the sliding window algorithm to traverse the data stream is as follows: Initialize a fixed-length time window and a preset step size parameter, load the structured time-series data stream into a memory buffer, extract continuous data segments sequentially according to the step size, and perform boundary smoothing processing on each segment; Record the start and end position indices of the window, and proceed cyclically until the entire data cycle is covered, forming a continuous overlapping local data sequence, which forms a segmented data stream and is then passed to the feature extraction node.

4. The method according to claim 1, characterized in that, The specific process for extracting the physiological characteristic mutation threshold and the deviation from the behavioral pattern is as follows: Frequency domain transformation and statistical feature calculations were performed on local data sequences to obtain the standard deviation distribution of heart rate variability and skin conductance response. Calculate the dynamic confidence interval by combining historical normal distribution curves, and mark the points that exceed the extreme values ​​of the interval as the mutation threshold; Simultaneously, the offset between action frequency and duration is calculated by comparing the behavior log sequence, generating a quantization deviation index, which is then smoothed, encapsulated into a quantization feature vector, and sent to the feature processing bus.

5. The method according to claim 1, characterized in that, The specific process of marking abnormal data segments and separating steady-state reference data is as follows: Input the mutation threshold and deviation index into the classifier, compare it with the preset situation rule base, identify data segments containing drastic fluctuations or pattern breaks, and attach an abnormal state identifier. Extract consecutive segments that do not match the exception rules into a candidate set, perform time continuity verification and integrity filtering, remove fragmented records, filter out abnormal data blocks, and retain normal baseline data.

6. The method according to claim 1, characterized in that, The specific process for instantiating the multi-time-span individual emotion baseline object is as follows: Read the steady-state baseline data stream, parse the timestamp range and divide it into short, medium and long-term dimensions, allocate independent memory space for each dimension, and initialize the baseline parameter container and metadata descriptor; By loading historical statistical distributions as prior knowledge, a baseline template containing the mean trajectory and variance boundary is constructed; Perform object lifecycle registration and return instance handles for downstream modules to call, register baseline object instances and expose call interfaces.

7. The method according to claim 1, characterized in that, The specific process for configuring the dynamic association model is as follows: Parse the dimensional attributes of the baseline object instance, establish a parameter mapping table across time scales, initialize the state transition matrix and weight decay coefficient, and define the information transmission path and synchronization triggering conditions between baselines. Load the time-series dependency graph structure, perform topological sorting to eliminate circular references, perform dynamic association model parameter binding and runtime environment configuration, and deploy the association computing engine to the runtime environment.

8. The method according to claim 1, characterized in that, The specific process of determining steady-state data segments and abnormal data segments based on preset thresholds is as follows: The system acquires real-time incoming monitoring data streams, extracts feature vectors for the current time window, compares the input judgment engine with preset threshold boundaries dimension by dimension, and calculates Euclidean distance and Mahalanobis distance scores. If the score is lower than the preset threshold, it is marked as a steady-state data segment; if it exceeds the preset threshold, it is marked as an abnormal data segment. A classification result label is generated and written to the status register. The status register is updated and the real-time classification result is recorded.

9. The method according to claim 1, characterized in that, The specific process of optimizing the baseline model through output is as follows: Collect weight update logs during iterative calculation, perform gradient accumulation and momentum correction, remove outlier parameters that deviate from the main distribution, and perform regularization constraint processing on the remaining parameters. Recalculate the baseline center trajectory and confidence interval boundaries, package the converged parameter set into a binary model file, attach a version identifier and checksum, persist it to the model repository and broadcast update notifications, and solidify the iteratively optimized baseline model version to the model repository.

10. The method according to claim 1, characterized in that, The specific process of constructing the multidimensional dynamic correlation network of physiological and behavioral characteristics is as follows: Extract the feature nodes from the optimized baseline model output, initialize the graph database instance, define the node type and edge weight attributes, and calculate the coupling strength between features based on the mutual information coefficient and Pearson correlation coefficient. Generate an adjacency matrix and construct an undirected graph topology. Execute a community detection algorithm to divide feature clusters. Perform network initialization and register a dynamic update callback interface. Construct a multi-dimensional feature dynamic association topology and load it into the graph database.

11. The method according to claim 1, characterized in that, The specific process of outputting the individual emotion reference feature map is as follows: Traverse the node and edge relationships in the dynamic association network, extract weighted connection paths and core feature clusters, execute the graph layout algorithm to optimize the spatial distribution of nodes, and render the feature hierarchy structure and interaction intensity heatmap. Generate a structured description file containing node attributes and edge weights, add timestamps and user identifiers, export it as visual graph data for the front-end rendering engine to parse, and encapsulate the output results in vector graphics format.

12. The method according to claim 1, characterized in that, The specific process for generating the final baseline result set is as follows: Receive the optimized baseline model and sentiment reference feature map, extract the multidimensional representation parameter set, perform normalization scaling and principal component projection transformation, and map the heterogeneous parameters to a unified vector space; Calculate the baseline stability score and deviation index, fuse the anomaly probability distribution and resilience coefficient, assemble them into a structured result array, perform integrity verification and write them into the output buffer to generate a standardized result array; The standardized result array is written to the output buffer, and redundant fields are compressed using a columnar storage format to generate a standardized data stream.

Images