A Deep Learning-Based Intelligent Drinking Water Behavior Management Method and System

Abstract

This invention discloses a deep learning-based intelligent drinking water behavior management method and system, comprising: S1, acquiring user drinking water behavior data; S2, constructing a behavior feature sequence and generating a behavior segment sequence; S3, inputting an improved SegRNN model, introducing a contact phase evolution mechanism in the segment recursive update layer, and generating a drinking water event sequence; S4, constructing an event feature set, and using a Mean Shift clustering algorithm with introduced order tide shifting to perform center convergence on the drinking water events, generating a drinking water event center set; S5, generating a drinking water behavior rhythm sequence; S6, constructing a drinking water behavior state and generating a drinking water management state; S7, outputting the drinking water behavior management result. This invention achieves accurate identification and dynamic scheduling of drinking water behavior, improving the accuracy, continuity, and personalization of drinking water management results.

Description

Technical Field

[0001] This invention relates to the field of intelligent behavior recognition and health management technology, and in particular to an intelligent drinking water behavior management method and system based on deep learning. Background Technology

[0002] With the continuous development of smart wearable devices, intelligent health management systems, and individual behavior sensing technologies, continuous monitoring and intelligent management of daily drinking behavior has gradually become an important research direction in the field of health management. Existing methods for managing drinking behavior mainly rely on manual recording, timed reminders, or rule-based judgments based on simple sensor data for drinking behavior identification and intervention. However, these methods generally suffer from the following problems in practical applications: The collected drinking behavior data is significantly affected by random human movement, posture changes, and environmental interference. There are obvious fluctuations and instabilities among limb movement data, hand displacement data, and container contact data, making it difficult to effectively distinguish drinking actions from other similar actions, thus affecting the accuracy of drinking behavior recognition. Multi-source behavioral data suffers from inconsistent sampling frequencies, time stamp offsets, and asynchrony across data channels during the collection process. Existing alignment methods cannot guarantee a strict correspondence of data in the time dimension, leading to information misalignment or continuity interruption during the construction of behavioral features. For drinking behaviors with obvious stage characteristics, traditional methods based on single time series modeling are unable to effectively characterize the evolutionary relationship between continuous behavioral stages such as approach, stay, and withdrawal, resulting in unclear boundaries of drinking events and incomplete expression of behavioral processes. In the process of drinking behavior event analysis, existing methods mostly use simple statistics or fixed window methods for clustering or grouping, lacking the ability to model the time sequence and dynamic change path of events, resulting in unstable event center localization and low accuracy of behavioral rhythm extraction.

[0003] Therefore, how to provide a method and system for intelligent drinking water behavior management based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose an intelligent drinking water behavior management method and system based on deep learning. This invention utilizes an improved SegRNN model and a Mean Shift clustering algorithm that introduces sequential tide shifting, combined with the construction process of drinking water behavior data streams, behavior feature sequences, behavior segment sequences, drinking water event sequences, and drinking water behavior rhythm sequences. It details the technical solutions for drinking water behavior recognition, drinking water event extraction, drinking water behavior rhythm analysis, and drinking water scheduling sequence generation, realizing the output of drinking water behavior management results driven by drinking water management status. It has the advantages of high accuracy in behavior recognition, strong continuity in event segmentation, strong rhythm modeling ability, and high degree of personalization in drinking water scheduling.

[0005] A method for intelligent drinking water behavior management based on deep learning according to an embodiment of the present invention includes the following steps: S1. Obtain user drinking behavior data, synchronize and arrange it over time, and generate a drinking behavior data stream; S2. Extract behavioral features from the drinking behavior data stream, construct a behavioral feature sequence, and segment the behavioral feature sequence using a sliding segmentation method to generate a behavioral segment sequence; S3. Input the sequence of behavioral segments into the improved SegRNN model, introduce a contact phase evolution mechanism in the segment recursive update layer, perform phase evolution transmission on adjacent behavioral segments, and generate a drinking event sequence based on the phase evolution results of consecutive behavioral segments. S4. Construct an event feature set based on the drinking water event sequence, and use the Mean Shift clustering algorithm with the introduction of the order tide migration center to form a migration path according to the time order of the events during the mean drift process, perform central convergence on the drinking water events, and generate a drinking water event center set; S5. Arrange the set of drinking water event centers, calculate the time interval relationship between adjacent drinking water event centers, and generate a drinking behavior rhythm sequence. S6. Construct a drinking behavior state based on the drinking behavior rhythm sequence, and generate a drinking management state; S7. Generate drinking water control quantities by mapping the state according to the drinking water management status, rearrange the drinking water event time intervals based on the drinking water control quantities to generate a drinking water scheduling sequence, and output the drinking water behavior management results according to the drinking water scheduling sequence.

[0006] Optionally, S1 specifically includes: Acquire user behavior drinking data during the drinking process and assign time stamps; the user behavior drinking data includes limb movement data, hand displacement data, head posture data, container contact data, and action duration data; The user behavior drinking data is sorted and aligned according to a unified time base to generate user aligned behavior drinking data. Based on the user's aligned drinking behavior data, a drinking behavior data stream is generated by time-series splicing.

[0007] Optionally, S2 specifically includes: Extract time series data corresponding to limb movement data, hand displacement data, head posture data, container contact data, and action duration data from the drinking behavior data stream; Calculate the action amplitude change sequence based on limb motion data, calculate the displacement change sequence based on hand displacement data, calculate the posture change sequence based on head posture data, calculate the contact change sequence based on container contact data, and calculate the duration change sequence based on action duration data. The action amplitude change sequence, displacement change sequence, posture change sequence, contact change sequence, and duration change sequence are spliced ​​together in chronological order to construct a behavioral feature sequence. The behavioral feature sequence is truncated according to a preset length, and adjacent truncated results are arranged in chronological order to generate a behavioral segment sequence.

[0008] Optionally, the improved SegRNN model specifically includes a behavior fragment encoding layer, a contact phase construction layer, a fragment recursive update layer, and an event generation layer; The behavior segment encoding layer maps each behavior segment in the behavior segment sequence to a corresponding segment representation vector, and arranges them in chronological order to form a segment representation sequence; The contact phase construction layer is based on the segment representation components corresponding to the contact change features, displacement change features, and posture change features in the behavioral feature sequence. It constructs a proximity phase representation according to the increasing trend of contact change and the approaching trend of displacement change, a dwelling phase representation according to the continuous trend of contact change and the stable trend of posture change, and a withdrawal phase representation according to the weakening trend of contact change and the moving away trend of displacement change. The proximity phase representation, dwelling phase representation, and withdrawal phase representation are written into the segment representation sequence in chronological order to form a contact phase sequence. The segment recursive update layer recursively updates the segment representation sequence and the contact phase sequence in chronological order. Between adjacent behavioral segments, the approach phase representation, dwell phase representation, and evacuation phase representation are updated together with the recursive hidden state. The sequential connection relationship of the approach phase representation, dwell phase representation, and evacuation phase representation is matched between consecutive behavioral segments. When consecutive behavioral segments sequentially satisfy the approach phase connection, dwell phase connection, and evacuation phase connection, the corresponding consecutive behavioral segments are merged into the same recursive segment chain. When the phase connection relationship is interrupted, the continuation of the recursive segment chain is terminated, and a phase recursive sequence is generated according to the chronological order of the recursive segment chain. The event generation layer merges continuous behavior segments into drinking events based on the start and end positions of the continuous behavior segments corresponding to the relay segment chain in the phase relay sequence, and arranges them in chronological order to form a drinking event sequence.

[0009] Optionally, S4 specifically includes: Extract the start time, end time, duration, and time interval between adjacent drinking water events from the drinking water event sequence to construct an event feature set; An event feature vector is generated based on the arrangement of the start time, end time, duration, and time interval between adjacent drinking events in the event feature set, and then the event feature vector is mapped to the event feature space. The event feature vectors in the event feature space are used as mean-shifted sample points, and the set of neighborhood sample points is determined based on the distance relationship between the sample points. The sample points in the neighborhood sample point set are sorted according to the time order in the drinking water event sequence, and adjacent sample points are connected in sequence to form a migration path. The local density center is calculated based on the set of neighboring sample points, and the moving direction and position of the sample points are adjusted in combination with the migration path to obtain the updated sample point position. Repeat the calculation of local density centers, adjustment of migration paths, and updating of sample point positions until the sample point positions converge; merge the converged sample point positions to generate a set of drinking water event centers.

[0010] Optionally, S5 specifically includes: The drinking water event centers are arranged according to the time order in the drinking water event center set to obtain an ordered drinking water event center sequence; The time interval between adjacent drinking water event centers is calculated based on the time positions of adjacent drinking water event centers in the ordered drinking water event center sequence, thus obtaining the time interval sequence. Based on the duration of adjacent drinking water event centers in the ordered drinking water event center sequence, the duration change relationship is calculated to obtain the duration change sequence. The relationship between event morphology changes is calculated based on the event feature vectors of adjacent drinking water event centers in the ordered drinking water event center sequence, and the event morphology change sequence is obtained. By splicing together the time interval sequence, the duration change sequence, and the event pattern change sequence in chronological order, a drinking behavior rhythm sequence is generated.

[0011] Optionally, S6 specifically includes: Arrange the time interval sequence, duration change sequence, and event pattern change sequence in chronological order to form a rhythm feature vector sequence; Based on the user's historical drinking behavior baseline, the offsets of time interval, duration and event pattern changes in the rhythm feature vector sequence are calculated respectively to obtain the time interval offset sequence, duration offset sequence and event pattern offset sequence. The time interval offset sequence, duration offset sequence, and event pattern offset sequence are spliced ​​together in chronological order to generate a drinking behavior state sequence. The drinking behavior state is determined based on the drinking behavior state sequence, and the drinking water management state is generated.

[0012] Optionally, the step of generating drinking water control quantities by mapping the state based on the drinking water management status specifically involves: Extract time interval offset sequence, duration offset sequence, and event pattern offset sequence from the drinking behavior state sequence; The time interval change sequence is obtained by calculating the difference between adjacent elements in the time interval offset sequence according to the time order; the duration change sequence is obtained by calculating the difference between adjacent elements in the duration offset sequence according to the time order; and the behavior change direction sequence is determined based on the change direction of adjacent elements in the event pattern offset sequence. By combining the time interval change sequence, the duration change sequence, and the behavior change direction sequence in chronological order, a control feature sequence is obtained. The time interval changes in the control feature sequence are accumulated to generate a time adjustment quantity; the duration changes in the control feature sequence are accumulated to generate a frequency adjustment quantity; the time adjustment quantity and the frequency adjustment quantity are arranged in chronological order to generate a drinking water control quantity.

[0013] Optionally, the step of rearranging the drinking event time intervals based on the drinking water control quantity to generate a drinking water scheduling sequence, and outputting the drinking water behavior management result according to the drinking water scheduling sequence, specifically includes: Based on the time adjustment factor in the drinking water control quantity, the time interval of drinking water events is increased or decreased to generate an adjusted time interval sequence; The time positions in the drinking water event center set are rearranged according to the adjusted time interval sequence to generate a drinking water scheduling sequence. The drinking water reminder time is generated according to the time position in the drinking water scheduling sequence, the drinking water frequency guidance information is generated according to the time interval in the drinking water scheduling sequence, and the behavior feedback information is generated according to the correspondence between the drinking water scheduling sequence and the drinking water management status. The drinking behavior management results are output based on drinking reminders, drinking frequency guidance information, and behavioral feedback information.

[0014] Optionally, a deep learning-based intelligent drinking water behavior management system includes the following modules: The drinking water data module is used to acquire user drinking water behavior data, synchronize and arrange it over time, and generate a drinking water behavior data stream. The feature construction module is used to extract behavioral features from drinking behavior data streams, construct behavioral feature sequences, and segment the behavioral feature sequences using a sliding segmentation method to generate behavioral segment sequences. The evolution modeling module is used to input the behavior segment sequence into the improved SegRNN model, introduce a contact phase evolution mechanism in the segment recursive update layer, perform phase evolution transmission on adjacent behavior segments, and generate a drinking event sequence based on the phase evolution results of consecutive behavior segments. The central convergence module is used to construct an event feature set based on the drinking water event sequence. It adopts the Mean Shift clustering algorithm that introduces the order tide migration center to form a migration path according to the time order of events during the mean drift process, and performs central convergence on the drinking water events to generate a central set of drinking water events. The rhythm generation module is used to arrange the set of drinking event centers, calculate the time interval relationship between adjacent drinking event centers, and generate a drinking behavior rhythm sequence. The state generation module is used to construct the drinking behavior state based on the drinking behavior rhythm sequence and generate the drinking management state. The scheduling output module is used to generate drinking water control quantities by mapping the state according to the drinking water management status, rearrange the drinking water event time intervals based on the drinking water control quantities to generate a drinking water scheduling sequence, and output the drinking water behavior management results according to the drinking water scheduling sequence.

[0015] The beneficial effects of this invention are: In the improved SegRNN model, a contact phase evolution mechanism is introduced. The evolutionary relationship between approach phase representation, dwell phase representation and withdrawal phase representation is constructed through the behavior segment encoding layer, contact phase construction layer, segment recursive update layer and event generation layer. This realizes the staged expression of drinking behavior segments and improves the completeness of drinking event sequence and the accuracy of boundary division. The Mean Shift clustering algorithm, which incorporates the tidal migration center, is adopted. During the mean drift process, the migration path is formed by combining the temporal order of drinking events. Stable convergence of the drinking event center set is achieved through local density center calculation and migration path adjustment, thereby improving the accuracy and continuity of the drinking behavior rhythm sequence construction. Based on the rhythm sequence of drinking behavior, a drinking behavior state is constructed and a drinking management state is generated. By rearranging the time interval of drinking events through drinking control quantities, a drinking scheduling sequence is generated, realizing the dynamic output of drinking reminder time, drinking frequency guidance information and behavior feedback information, thereby improving the personalization and adaptability of drinking behavior management results. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a deep learning-based intelligent drinking water behavior management method proposed in this invention; Figure 2 This is a schematic diagram of the improved SegRNN model proposed in this invention; Figure 3 This is a schematic diagram of an intelligent drinking water behavior management system based on deep learning proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-2 A deep learning-based intelligent drinking water behavior management method includes the following steps: S1. Obtain user drinking behavior data, synchronize and arrange it over time, and generate a drinking behavior data stream; S2. Extract behavioral features from the drinking behavior data stream, construct a behavioral feature sequence, and segment the behavioral feature sequence using a sliding segmentation method to generate a behavioral segment sequence; S3. Input the behavioral segment sequence into the improved SegRNN model, introduce a contact phase evolution mechanism in the segment recursive update layer, perform phase evolution transmission on adjacent behavioral segments, and generate a drinking event sequence based on the phase evolution results of continuous behavioral segments. S4. Construct an event feature set based on the drinking water event sequence, and use the Mean Shift clustering algorithm with the introduction of the order tide migration center to form a migration path according to the time order of the events during the mean drift process, perform central convergence on the drinking water events, and generate a drinking water event center set; S5. Arrange the set of drinking water event centers, calculate the time interval relationship between adjacent drinking water event centers, and generate a drinking behavior rhythm sequence. S6. Construct drinking behavior status based on drinking behavior rhythm sequence and generate drinking management status; S7. Generate drinking water control quantities by mapping the state based on the drinking water management status, rearrange the drinking water event time intervals based on the drinking water control quantities to generate a drinking water scheduling sequence, and output the drinking water behavior management results based on the drinking water scheduling sequence.

[0019] In this embodiment, S1 specifically refers to: The system collects data on user limb movement, hand displacement, head posture, container contact, and action duration during the drinking process. Limb movement data records the changes in the position of the upper limbs and trunk during the drinking action. Hand displacement data records the changes in the position of the hand during continuous sampling. Head posture data records the changes in the pitch and rotation of the head during continuous sampling. Container contact data records the changes in contact between the drinking container and the mouth as they approach and leave. Action duration data records the duration of the drinking action segment on the time axis. Time stamps are assigned to limb movement data, hand displacement data, head posture data, container contact data, and action duration data respectively. The time stamps are written to each data record in the order of sampling, so that limb movement data, hand displacement data, head posture data, container contact data, and action duration data at the same sampling time have the same time position. User drinking behavior data is sorted and aligned according to a unified time benchmark. During sorting, data records from different sources are arranged in ascending order of time identifier. During alignment, data records at the same time position are assigned to the same sampling time. If there are missing positions between adjacent sampling times, transitional data between adjacent times are written at the missing positions to ensure that limb movement data, hand displacement data, head posture data, container contact data, and action duration data maintain a corresponding relationship across all sampling times, thus obtaining aligned user drinking behavior data. Based on the user's aligned drinking behavior data, time-series stitching is performed. During stitching, the limb movement data, hand displacement data, head posture data, container contact data, and action duration data corresponding to the same sampling time are extracted sequentially from front to back according to the time identifier. The data at the same sampling time are arranged into a set of time data in a fixed order. Then, all the time data are connected end to end in chronological order to form a continuous time-series record, generating a drinking behavior data stream.

[0020] In this embodiment, S2 specifically refers to: From the drinking behavior data stream, limb movement data, hand displacement data, head posture data, container contact data, and action duration data are separated sequentially according to time identifiers. The limb movement data is arranged in chronological order to form a time series corresponding to the limb movement data, the hand displacement data is arranged in chronological order to form a time series corresponding to the hand displacement data, the head posture data is arranged in chronological order to form a time series corresponding to the head posture data, the container contact data is arranged in chronological order to form a time series corresponding to the container contact data, and the action duration data is arranged in chronological order to form a time series corresponding to the action duration data. Based on the difference in motion position between adjacent sampling times in the time series corresponding to limb motion data, the change in motion amplitude at each sampling time is obtained sequentially, and all the changes in motion amplitude are arranged in chronological order to form a sequence of motion amplitude changes. Based on the hand position difference between adjacent sampling times in the time series corresponding to hand displacement data, the displacement change corresponding to each sampling time is obtained in sequence, and all displacement changes are arranged in chronological order to form a displacement change sequence. Based on the pitch and rotation changes at adjacent sampling times in the time series corresponding to head pose data, the pose change at each sampling time is obtained sequentially, and all pose changes are arranged in chronological order to form a pose change sequence. Based on the contact state difference between adjacent sampling times in the time series corresponding to container contact data, the contact change amount corresponding to each sampling time is obtained in turn, and all contact changes are arranged in chronological order to form a contact change sequence. Based on the duration difference between adjacent sampling moments in the time series corresponding to the action duration data, the duration change corresponding to each sampling moment is obtained in turn, and all duration changes are arranged in chronological order to form a duration change sequence. The action amplitude change sequence, displacement change sequence, posture change sequence, contact change sequence, and duration change sequence are spliced ​​together in chronological order. During splicing, the action amplitude change, displacement change, posture change, contact change, and duration change are taken out sequentially at the same time position and arranged into a set of feature data in a fixed order. Then, the feature data corresponding to all time positions are connected end to end in chronological order to construct the behavior feature sequence. The behavioral feature sequence is truncated according to a preset length. During truncating, the first continuous feature data is truncated from the starting position of the behavioral feature sequence as the first behavioral segment. Then, the starting point is moved backward according to a fixed sliding step size. At the new starting position, continuous feature data of the same preset length is truncated as the next behavioral segment. The truncating is repeated until the end of the behavioral feature sequence is reached. All truncating results are arranged from front to back according to the starting time position to generate a behavioral segment sequence.

[0021] In this embodiment, the improved SegRNN model specifically includes a behavior fragment encoding layer, a contact phase construction layer, a fragment recursive update layer, and an event generation layer; Each behavior segment in the behavior segment sequence is input into the behavior segment encoding layer in chronological order. The change in motion amplitude, displacement, posture, contact, and duration are extracted from each behavior segment in sequence. The change in motion amplitude, displacement, posture, contact, and duration corresponding to the same time position are arranged in a fixed order as a set of segment feature data. The segment feature data corresponding to all time positions within the same behavior segment are then concatenated end to end to form a segment feature matrix. Each row of segment feature data in the segment feature matrix is ​​multiplied and added to the corresponding encoding parameters one by one to obtain the segment encoding result at the corresponding time position. Then, the segment encoding results of all time positions within the same row segment are arranged in chronological order to form the segment representation vector of the corresponding row segment. Finally, the segment representation vectors corresponding to all row segments are arranged in the order of the row segments in the row segment sequence to form the segment representation sequence. The fragment representation sequence is input into the contact phase construction layer, and the fragment representation components corresponding to contact change features, displacement change features, and attitude change features are separated from the fragment representation vector. Compare the segment representation components corresponding to the contact change characteristics at adjacent time positions. If the value at the later time position is greater than the value at the earlier time position, it is recorded as an enhanced contact change. If the value at the later time position is equal to the value at the earlier time position, it is recorded as a continued contact change. If the value at the later time position is less than the value at the earlier time position, it is recorded as a weakened contact change. Compare the segment representation components corresponding to the displacement change characteristics of adjacent time positions. If the distance from the hand position to the container contact area decreases at each time step, it is recorded as the displacement change is approaching. If the distance from the hand position to the container contact area increases at each time step, it is recorded as the displacement change is moving away. Compare the segment representation components corresponding to the attitude change characteristics of adjacent time positions. If the changes in head pitch and rotation remain in the same direction with small fluctuations within consecutive time positions, it is recorded as stable attitude change. The time positions corresponding to the increase in contact change and the approach of displacement change are combined to form the approach phase representation; the time positions corresponding to the continuation of contact change and the stabilization of attitude change are combined to form the dwell phase representation; and the time positions corresponding to the decrease in contact change and the distance of displacement change are combined to form the withdrawal phase representation. The approach phase representation, dwell phase representation, and evacuation phase representation are written into the corresponding time positions of the corresponding behavior segments in the segment representation sequence. When writing, the approach phase representation, dwell phase representation, and evacuation phase representation are arranged sequentially at the same time position and connected to the beginning and end of the original segment representation vector to form a contact phase sequence. Input the fragment representation sequence and the contact phase sequence into the fragment recursive update layer in the same behavior fragment order. Write the initial recursive hidden state at the first behavior fragment and write the updated recursive hidden state of the previous behavior fragment at the next behavior fragment. Between adjacent action segments, the approach phase representation, dwell phase representation, and evacuation phase representation corresponding to the previous action segment are arranged together with the recursive hidden state updated by the previous action segment and placed after the segment representation vector of the next action segment to form the recursive input sequence of the next action segment. Write the segment representation vector, approach phase representation, dwell phase representation, evacuation phase representation and recursive hidden state in the recursive input sequence into the recursive unit position by position. Update the recursive hidden state position by position in the same action segment according to the time order. Pass the recursive hidden state between adjacent action segments in the order of action segments. Compare the order of appearance of approach phase representation, dwell phase representation and evacuation phase representation in continuous action segments. If an approach phase representation appears at the end of the previous action segment and a dwell phase representation continues to appear at the beginning of the next action segment, it is recorded as an approach phase connection and a dwell phase connection being established. If an evacuation phase representation continues to appear after a dwell phase representation, it is recorded as an evacuation phase connection being established. When consecutive action segments sequentially satisfy the approach phase connection, stay phase connection, and withdrawal phase connection, the consecutive action segments that satisfy the connection relationship are merged into the same recurring segment chain in chronological order. The starting time position of the first action segment in the recurring segment chain is recorded as the starting position of the recurring segment chain, and the ending time position of the last action segment in the recurring segment chain is recorded as the ending position of the recurring segment chain. When the approach phase representation, dwell phase representation, and evacuation phase representation are missing, inverted, or interrupted between consecutive action segments, the action segments that have formed a connection before the interruption position are retained in the existing recurring segment chain, and the continuation of the recurring segment chain ends at the interruption position. Arrange all the recurring segments from front to back according to their starting time positions to form a phase recurrence sequence; The phase evolution sequence is input into the event generation layer, and the start and end positions are extracted from each evolution segment chain. All continuous behavior segments covered between the start and end positions are merged in chronological order to form a drinking event. All drinking events are arranged from front to back according to the start position to form a drinking event sequence.

[0022] In this embodiment, both the improved SegRNN model and the SegRNN model adopt a structure that segments the time series and recursively updates it at the segment level. Both use behavioral segment sequences as input and pass temporal information between adjacent segments through recursive hidden states. Both complete segment-level state updates in chronological order and output sequence-level results. The improved SegRNN model further expands the input data from single-segment features to a combination of multi-source behavioral features including contact change features, displacement change features, and posture change features. In the behavioral segment encoding layer, multi-dimensional features are aligned and rearranged according to their temporal positions to form segment representation vectors, enabling the segment representation to simultaneously encompass… It includes information on changes in motion amplitude, spatial displacement, and contact state. The improved SegRNN model introduces approach phase, dwell phase, and withdrawal phase representations corresponding to enhanced, sustained, and weakened contact changes in the contact phase construction layer. This establishes a correspondence between contact change trends and displacement direction and attitude stability, and incorporates contact phase information into the segment representation sequence for subsequent recursive updates, enabling segment representations to segmentate behavior stages. Furthermore, the improved SegRNN model incorporates contact phase representation and recursive hidden states together in the segment recursive update layer, passing the contact phase representation and hidden state between adjacent behavior segments. The model improves upon traditional SegRNN by using phase continuity relationships as the basis for segment merging during recursive updates. This is achieved by comparing the sequential relationships between close phase representations, stationary phase representations, and departing phase representations, thus avoiding the problem of traditional SegRNN relying solely on numerical changes and failing to distinguish behavioral stages. Furthermore, the model incorporates phase continuity relationships as the basis for segment merging during recursive updates. By grouping consecutive behavioral segments that satisfy close phase continuity, stationary phase continuity, and departing phase continuity into a recurring segment chain, the model output transforms from single-point prediction to structured event representation. Finally, the model further improves upon traditional SegRNN by using the starting position of the recurring segment chain and... The drinking event is generated at the end position, transforming the output from a fragment-level representation into a behavioral event with a clear time boundary, thereby improving the completeness and continuity of behavior recognition. By introducing a contact phase evolution structure and embedding the phase information transmission path in the recursive update process, the model can distinguish between the approach phase, the dwell phase, and the withdrawal phase when processing drinking behavior, reducing misjudgments between different phases and improving the ability to express complex behavioral processes. By embedding multi-source behavioral features and contact phase information together in the recursive update process, the model enhances its ability to characterize behavioral structures while maintaining its original time-series modeling capabilities, thereby improving the accuracy and stability of drinking event recognition.

[0023] In this embodiment, S4 specifically refers to: Extract the start and end positions of each drinking water event sequentially from the drinking water event sequence in chronological order, and take the time identifier corresponding to the start position as the start time of the drinking water event and the time identifier corresponding to the end position as the end time of the drinking water event. Subtract the start time of the drinking water event from the end time of the drinking water event to obtain the duration of the corresponding drinking water event, and arrange the drinking water events in the order of their occurrence in the drinking water event sequence to form a drinking water event duration sequence. Subtract the start time of the previous drinking event from the start time of the next drinking event to obtain the time interval between adjacent drinking events, and arrange the drinking events in the order of their chronological sequence to form a sequence of time intervals between adjacent drinking events. The start time, end time, duration, and time interval of adjacent drinking water events are arranged sequentially according to their correspondence in the drinking water event sequence. At the same drinking water event location, the start time, end time, duration, and time interval of adjacent drinking water events are combined in a fixed order to form a set of event feature data. All event feature data are arranged in chronological order to construct an event feature set. Each set of event feature data is extracted sequentially from the event feature set. The start time, end time, duration, and time interval between adjacent drinking events of the drinking water event are connected end to end in a fixed order to form the event feature vector of the corresponding drinking water event. All event feature vectors are then arranged in the order of the drinking water events in the drinking water event sequence. The event feature vector is mapped to the event feature space. During the mapping, the start time of the drinking water event, the end time of the drinking water event, the duration of the drinking water event, and the time interval between adjacent drinking water events are respectively used as different dimensions in the event feature vector, so that each drinking water event corresponds to a sample point position in the event feature space. The event feature vectors in the event feature space are used as mean-shifted sample points. The corresponding dimensional differences between any two sample points are squared one by one and then added together. The square root of the sum of squares is then taken to obtain the distance between the sample points. Taking each sample point as the center, sample points whose distance from the center sample point is less than the preset neighborhood range are grouped into the same neighborhood sample point set, and the drinking event time position corresponding to the center sample point is used as the reference time position of the neighborhood sample point set. The sample points in the neighborhood sample point set are sorted according to the time order in the drinking water event sequence. When sorting, the sample points with earlier time positions are arranged first, and the sample points with later time positions are arranged last. Then, the adjacent sample points after sorting are connected end to end to form a migration path. The local density center is calculated based on the set of neighborhood sample points. During the calculation, the values ​​of all sample points in the neighborhood sample point set in each dimension are added together, and the sum is divided by the number of neighborhood sample points to obtain the center coordinates of the corresponding dimension. The center coordinates of all dimensions are combined to form the local density center position. The migration path is combined to adjust the movement direction and position of the sample point. During the adjustment, the direction of the line connecting the current sample point position to the local density center position is used as the basic movement direction. Then, the path directions adjacent to the current sample point in the migration path are superimposed on the basic movement direction to obtain the updated movement direction. The current sample point is moved towards the local density center position along the updated movement direction. The position of the sample point after the movement is equal to the sum of the current sample point position and the updated displacement. The updated sample point positions are then used as new mean-shifted sample point positions. The process of determining the neighborhood sample point set, forming the migration path, calculating local density centers, and updating sample point positions is repeated until the difference between two consecutive updated sample point positions is less than a preset convergence range across all dimensions. The preset convergence range is defined as the difference between two consecutive sample point positions in each dimension of the event feature space being less than 3% of the difference between the maximum and minimum values ​​of the corresponding dimension. The locations of the converged sample points are merged. During merging, sample points with overlapping locations or location differences less than a preset merging range are merged into the same event center. All sample points corresponding to the same event center are then assigned to the same drinking water event center in chronological order to generate a drinking water event center set.

[0024] In this embodiment, S5 specifically refers to: The drinking water event centers are sorted from front to back according to their corresponding time positions in the drinking water event center set. The drinking water event centers with smaller time positions are arranged first, and the drinking water event centers with larger time positions are arranged last, forming an ordered drinking water event center sequence. Take out two adjacent drinking water event centers from the ordered drinking water event center sequence, subtract the time position of the previous drinking water event center from the time position of the latter drinking water event center to obtain the corresponding time interval value, and arrange all the time interval values ​​according to the order of the drinking water event centers in the ordered drinking water event center sequence to form a time interval sequence. Take out two adjacent drinking water event centers from the ordered drinking water event center sequence, subtract the duration of the previous drinking water event center from the duration of the subsequent drinking water event center to obtain the corresponding duration change value, and arrange all duration change values ​​according to the order of the drinking water event centers in the ordered drinking water event center sequence to form a duration change sequence. From the ordered drinking water event center sequence, the event feature vectors corresponding to two adjacent drinking water event centers are extracted sequentially. The differences in the corresponding dimensions of the two event feature vectors are calculated respectively. Then, the absolute values ​​of all dimension differences are taken and arranged in dimensional order to form the event morphology change vectors corresponding to adjacent drinking water event centers. All event morphology change vectors are arranged in the order of drinking water event centers in the ordered drinking water event center sequence to form the event morphology change sequence. The time interval sequence, duration change sequence, and event pattern change sequence are spliced ​​together in chronological order. During splicing, the corresponding time interval value, duration change value, and event pattern change vector are extracted sequentially from the center position of the same adjacent drinking event and arranged in a fixed order as a set of rhythmic feature data. All rhythmic feature data are then connected end to end in chronological order to generate a drinking behavior rhythmic sequence.

[0025] In this embodiment, S6 specifically refers to: The data at the same time position in the time interval sequence, duration change sequence, and event pattern change sequence are arranged in chronological order. At each time position, the time interval value, duration change value, and event pattern change vector are extracted in sequence and combined into a set of rhythm feature data in a fixed order. The rhythm feature data corresponding to all time positions are arranged from front to back to form a rhythm feature vector sequence. Extract the historical time interval baseline sequence, historical duration variation baseline sequence, and historical event pattern baseline sequence from the user's historical drinking behavior baseline, and establish a one-to-one correspondence between the time position of the rhythm feature vector sequence and the corresponding time position in the user's historical drinking behavior baseline. The time interval values ​​in the rhythm feature vector sequence are compared with the baseline values ​​at the corresponding time positions in the historical time interval baseline sequence. The time interval offset values ​​are obtained by subtracting the corresponding baseline values ​​from the time interval values ​​in the rhythm feature vector sequence. All time interval offset values ​​are then arranged in chronological order to form a time interval offset sequence. The duration variation values ​​in the rhythm feature vector sequence are compared with the baseline values ​​at the corresponding time positions in the historical duration variation baseline sequence. The duration offset values ​​are obtained by subtracting the corresponding baseline values ​​from the duration variation values ​​in the rhythm feature vector sequence. All duration offset values ​​are then arranged in chronological order to form a duration offset sequence. The event morphology change vector in the rhythm feature vector sequence is compared with the baseline vector at the corresponding time position in the historical event morphology baseline sequence. The difference between the corresponding dimensions of the two vectors is calculated to obtain the morphology offset value of the corresponding dimension. The morphology offset values ​​of all dimensions at the same time position are arranged in dimensional order to form the event morphology offset vector at the corresponding time position. Then, all event morphology offset vectors are arranged in chronological order to form the event morphology offset sequence. The time interval offset sequence, duration offset sequence, and event pattern offset sequence are spliced ​​together in chronological order. During splicing, the time interval offset value, duration offset value, and event pattern offset vector are taken out sequentially at the same time position and arranged into a set of state feature data in a fixed order. The state feature data corresponding to all time positions are connected from beginning to end to generate a drinking behavior state sequence. From the drinking behavior state sequence, extract the time interval offset value, duration offset value, and event form offset vector corresponding to each time position in sequence. Compare the positive and negative directions and absolute values ​​of the time interval offset value, the duration offset value, and the direction and magnitude of change of each dimension of the event form offset vector. Combine the comparison results of the same time position to form the drinking behavior state of the corresponding time position. Arrange the drinking behavior states corresponding to all time positions in chronological order to generate the drinking management state.

[0026] In this embodiment, the drinking water control quantity is generated by mapping the state based on the drinking water management status, specifically as follows: Separate the time interval offset sequence, duration offset sequence, and event pattern offset sequence from the drinking behavior state sequence in chronological order. Make each time position in the time interval offset sequence consistent with the corresponding time position in the drinking behavior state sequence, make each time position in the duration offset sequence consistent with the corresponding time position in the drinking behavior state sequence, and make each time position in the event pattern offset sequence consistent with the corresponding time position in the drinking behavior state sequence. Take out the offset values ​​of two adjacent time positions in the time interval offset sequence in turn, subtract the time interval offset value of the previous time position from the time interval offset value of the next time position to obtain the time interval change at the corresponding time position, and arrange all the time interval changes in chronological order to form a time interval change sequence. Take out the offset values ​​of two adjacent time positions in the duration offset sequence in turn, subtract the duration offset value of the previous time position from the duration offset value of the later time position to obtain the duration change of the corresponding time position, and arrange all duration changes in chronological order to form a duration change sequence. The event form offset vectors of two adjacent time positions in the event form offset sequence are extracted sequentially. The offset values ​​of the corresponding dimensions in the event form offset vectors of the two time positions are compared item by item. If the offset value of the later time position in the same dimension is greater than the offset value of the earlier time position, the corresponding dimension is recorded as a positive change. If the offset value of the later time position in the same dimension is less than the offset value of the earlier time position, the corresponding dimension is recorded as a negative change. If the offset value of the later time position in the same dimension is equal to the offset value of the earlier time position, the corresponding dimension is recorded as a constant change. The change directions of all dimensions at the same time position are arranged in dimensional order to form the behavior change direction. All behavior change directions are arranged in chronological order to form the behavior change direction sequence. The time interval change sequence, duration change sequence, and behavior change direction sequence are combined in chronological order. At the same time position, the time interval change, duration change, and behavior change direction are taken out in sequence and arranged in a fixed order to form a set of control feature data. The control feature data corresponding to all time positions are arranged from front to back to obtain the control feature sequence. The time interval changes in the control feature sequence are extracted sequentially. The time interval change at the first time position is used as the initial time adjustment amount. The time adjustment amount at the next time position is set as the sum of the time adjustment amount at the previous time position and the time interval change at the current time position. The time adjustment amounts corresponding to all time positions are obtained sequentially and arranged in chronological order to form a time adjustment amount sequence. The duration change in the control feature sequence is extracted sequentially. The duration change at the first time position is used as the initial frequency adjustment amount. The frequency adjustment amount at the next time position is set as the sum of the frequency adjustment amount at the previous time position and the duration change at the current time position. The frequency adjustment amounts corresponding to all time positions are obtained in sequence and arranged in chronological order to form a frequency adjustment amount sequence. The time regulation quantity sequence and the frequency regulation quantity sequence are spliced ​​together in chronological order. During splicing, the time regulation quantity and the frequency regulation quantity are taken out sequentially at the same time position and arranged into a set of control quantity data in a fixed order. The control quantity data corresponding to all time positions are connected from beginning to end to generate drinking water control quantity.

[0027] In this embodiment, the drinking water event time intervals are rearranged based on the drinking water control quantity to generate a drinking water scheduling sequence, and the drinking water behavior management results are output according to the drinking water scheduling sequence, specifically as follows: Separate the time regulation quantity sequence and the frequency regulation quantity sequence from the drinking water control quantity in chronological order, and keep each time position in the time regulation quantity sequence consistent with the corresponding time position in the time interval sequence, and keep each time position in the frequency regulation quantity sequence consistent with the corresponding drinking water event center in the drinking water event center set; The time interval values ​​in the time interval sequence and the corresponding time adjustment values ​​in the time adjustment value sequence are extracted sequentially. The time interval values ​​at the same time position are added to the time adjustment values ​​to obtain the adjusted time interval values. All the adjusted time interval values ​​are arranged in chronological order to form the adjusted time interval sequence. Take the drinking water event center with the earliest time position from the set of drinking water event centers, and take the time position corresponding to the drinking water event center with the earliest time position as the starting scheduling time position; Take out each time interval value in the adjusted time interval sequence in turn, add the previous time position to the current time interval value to get the scheduling time position corresponding to the next drinking water event center, and then get the scheduling time positions corresponding to all drinking water event centers in turn. Replace each drinking water event center in the drinking water event center set with its corresponding scheduling time position, and arrange the replaced drinking water event centers from front to back according to their scheduling time positions to generate a drinking water scheduling sequence. Extract the scheduling time position corresponding to each drinking water event center from the drinking water scheduling sequence, and arrange the scheduling time positions in chronological order to form a drinking water reminder time sequence; The scheduling time position difference between adjacent drinking water event centers is extracted sequentially from the drinking water scheduling sequence. All differences are arranged in chronological order to form a scheduling time interval sequence. The corresponding frequency adjustment amount in the frequency adjustment amount sequence is combined with the corresponding time interval value in the scheduling time interval sequence. The frequency adjustment amount and time interval value are arranged sequentially at the same time position to form drinking water frequency guidance information. Establish a correspondence between the time position in the drinking water scheduling sequence and the corresponding time position in the drinking water management status. Extract the drinking water management status at the same time position and arrange the drinking water management status and the corresponding scheduling time position in a fixed order to form behavioral feedback information. The system outputs drinking water reminder time sequence, drinking water frequency guidance information, and behavioral feedback information in chronological order to generate drinking water behavior management results.

[0028] refer to Figure 3 In this embodiment, a deep learning-based intelligent drinking water behavior management system includes the following modules: The drinking water data module is used to acquire user drinking water behavior data, synchronize and arrange it over time, and generate a drinking water behavior data stream. The feature construction module is used to extract behavioral features from drinking behavior data streams, construct behavioral feature sequences, and segment the behavioral feature sequences using a sliding segmentation method to generate behavioral segment sequences. The recursive modeling module is used to input the behavioral segment sequence into the improved SegRNN model, introduce a contact phase recursion mechanism in the segment recursive update layer, perform phase recursion propagation on adjacent behavioral segments, and generate a drinking event sequence based on the phase recursion results of consecutive behavioral segments. The central convergence module is used to construct an event feature set based on the drinking water event sequence. It adopts the Mean Shift clustering algorithm that introduces the order tide migration center to form a migration path according to the time order of events during the mean drift process, and performs central convergence on the drinking water events to generate a central set of drinking water events. The rhythm generation module is used to arrange the set of drinking water event centers, calculate the time interval relationship between adjacent drinking water event centers, and generate a drinking behavior rhythm sequence. The status generation module is used to construct drinking behavior status based on drinking behavior rhythm sequence and generate drinking management status. The scheduling output module is used to generate drinking water control quantities by mapping the state according to the drinking water management status, rearrange the drinking water event time intervals based on the drinking water control quantities to generate a drinking water scheduling sequence, and output the drinking water behavior management results according to the drinking water scheduling sequence.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to a health management scenario in an office park. The park has four office buildings with an average of 186 employees on duty daily, of whom 138 are sedentary. Their daily water-drinking behavior is significantly irregular. On-site investigation revealed that traditional timed reminders cannot distinguish whether users have already drunk water, nor can they identify differences in actual behavior such as holding a cup without drinking, short-term contact, or continuous small sips, leading to an imbalance in reminder frequency, inaccurate intervention timing, and decreased user acceptance. Therefore, behavior data collection terminals were deployed in the park's rest areas, workstations, and meeting areas to continuously collect data on limb movement, hand displacement, head posture, container contact, and action duration, forming a water-drinking behavior data stream based on a unified time benchmark. The data collection period was 30 consecutive days.

[0030] In practical applications, a behavioral feature sequence is first constructed from the drinking behavior data stream. Then, a behavioral segment sequence is formed through sliding segmentation and input into an improved SegRNN model. The model introduces a contact phase evolution mechanism in the segment recursive update layer, continuously transmitting the approach phase, dwell phase, and withdrawal phase to extract real drinking events from continuous actions. Subsequently, an event feature set is constructed based on the start time, end time, duration, and time interval between adjacent drinking events. A Mean Shift clustering algorithm incorporating sequential tide shifting is used to form a drinking event center set, generating a drinking behavior rhythm sequence, drinking behavior state, and drinking management state. Finally, drinking reminder times, drinking frequency guidance information, and behavioral feedback information are output. During system operation, manually labeled results are used as a comparison to statistically analyze the accuracy, false alarm rate, event boundary deviation, and effectiveness of drinking event identification and scheduling.

[0031] Table 1 Comparison of Drinking Behavior Recognition Results

[0032] As shown in Table 1, the method of this invention has significant advantages in the drinking behavior recognition stage. Traditional rule-based reminder methods mainly rely on fixed times or simple action triggers, making it difficult to distinguish between picking up a cup, pausing, and actual drinking, resulting in high false alarm rates and boundary deviations. Although ordinary temporal recognition methods improve overall recognition capabilities, they still suffer from confusion in recognizing continuous small sips of water and short pauses in meeting scenarios. Methods without introducing a contact phase progression mechanism can complete general segment modeling, but they lack sufficient expression of the continuous connection between approach phase, pausing phase, and withdrawal phase, resulting in a boundary deviation still close to 3 seconds. The method of this invention improves the accuracy of real drinking event recognition to 94.6% and reduces the false alarm rate to 4.1%, indicating that the expression of continuous behavior segments in the progression chain is more complete, and it can better solve the problems of overlapping drinking actions with other similar actions and unclear event boundaries in the background technology.

[0033] Table 2 Comparison of the Effectiveness of Drinking Water Behavior Management

[0034] As shown in Table 2, this invention not only improves recognition accuracy but also yields better practical benefits in management. While traditional rule-based reminder methods provide the most reminders, 36.5% of them are invalid, indicating a disconnect between reminder timing and actual drinking behavior, making it easy for users to ignore them. Ordinary time-series recognition methods and methods without incorporating tidal shift centers show some improvement in management effectiveness, but the insufficient stability of drinking event centers limits the accuracy of rhythmic sequence construction, resulting in discrepancies between the drinking schedule sequence and actual daily routines. This invention constructs a drinking behavior rhythmic sequence through a set of drinking event centers, and then generates drinking control quantities and drinking schedule sequences based on the drinking management status. This reduces the average number of reminders per person to 5.4 times per day, while increasing the user response rate to 81.5%, the drinking compliance rate to 86.9%, and the average drinking interval to 88 minutes. This demonstrates that this invention can improve management effectiveness and individual adaptability while reducing disturbances.

[0035] The results of 30 days of continuous application further demonstrate that this invention can effectively identify real drinking water events in office park scenarios, forming a stable set of drinking water event centers, and further generating a drinking behavior rhythm sequence that matches users' actual work and rest schedules. For users who experience significant delays in drinking water in the afternoon, prolonged periods without drinking water during meetings, or frequent small sips of water, the system can provide differentiated drinking water scheduling sequences based on drinking water management status, avoiding concentrated and ineffective reminders caused by traditional fixed reminders. Statistics show that in the first week of application, the number of people in the park meeting the drinking water standard was 108, increasing to 129 in the second week, 148 in the third week, and stabilizing at 162 in the fourth week, indicating that this invention has continuous improvement capabilities. This proves that this invention can solve the problems of difficulty in synchronizing multi-source behavioral data, difficulty in identifying drinking water event boundaries, instability of event centers, and lack of personalization in drinking water scheduling in existing technologies, demonstrating high feasibility and practical application value.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent drinking water behavior management method, characterized in that, Includes the following steps: S1. Obtain user drinking behavior data, synchronize and arrange it over time, and generate a drinking behavior data stream; S2. Extract behavioral features from the drinking behavior data stream, construct a behavioral feature sequence, and segment the behavioral feature sequence using a sliding segmentation method to generate a behavioral segment sequence; S3. Input the sequence of behavioral segments into the improved SegRNN model, introduce a contact phase evolution mechanism in the segment recursive update layer, perform phase evolution transmission on adjacent behavioral segments, and generate a drinking event sequence based on the phase evolution results of consecutive behavioral segments. S4. Construct an event feature set based on the drinking water event sequence, and use the Mean Shift clustering algorithm with the introduction of the order tide migration center to form a migration path according to the time order of the events during the mean drift process, perform central convergence on the drinking water events, and generate a drinking water event center set; S5. Arrange the set of drinking water event centers, calculate the time interval relationship between adjacent drinking water event centers, and generate a drinking behavior rhythm sequence. S6. Construct a drinking behavior state based on the drinking behavior rhythm sequence, and generate a drinking management state; S7. Generate drinking water control quantities by mapping the state according to the drinking water management status, rearrange the drinking water event time intervals based on the drinking water control quantities to generate a drinking water scheduling sequence, and output the drinking water behavior management results according to the drinking water scheduling sequence.

2. The intelligent drinking water behavior management method based on deep learning according to claim 1, characterized in that, Specifically, S1 is: Acquire user behavior drinking data during the drinking process and assign time stamps; the user behavior drinking data includes limb movement data, hand displacement data, head posture data, container contact data, and action duration data; The user behavior drinking data is sorted and aligned according to a unified time base to generate user aligned behavior drinking data. Based on the user's aligned drinking behavior data, a drinking behavior data stream is generated by time-series splicing.

3. The intelligent drinking water behavior management method based on deep learning according to claim 1, characterized in that, Specifically, S2 is: Extract time series data corresponding to limb movement data, hand displacement data, head posture data, container contact data, and action duration data from the drinking behavior data stream; Calculate the action amplitude change sequence based on limb motion data, calculate the displacement change sequence based on hand displacement data, calculate the posture change sequence based on head posture data, calculate the contact change sequence based on container contact data, and calculate the duration change sequence based on action duration data. The action amplitude change sequence, displacement change sequence, posture change sequence, contact change sequence, and duration change sequence are spliced ​​together in chronological order to construct a behavioral feature sequence. The behavioral feature sequence is truncated according to a preset length, and adjacent truncated results are arranged in chronological order to generate a behavioral segment sequence.

4. The intelligent drinking water behavior management method based on deep learning according to claim 1, characterized in that, The improved SegRNN model specifically includes a behavior segment encoding layer, a contact phase construction layer, a segment recursive update layer, and an event generation layer; The behavior segment encoding layer maps each behavior segment in the behavior segment sequence to a corresponding segment representation vector, and arranges them in chronological order to form a segment representation sequence; The contact phase construction layer is based on the segment representation components corresponding to the contact change features, displacement change features, and posture change features in the behavioral feature sequence. It constructs a proximity phase representation according to the increasing trend of contact change and the approaching trend of displacement change, a dwelling phase representation according to the continuous trend of contact change and the stable trend of posture change, and a withdrawal phase representation according to the weakening trend of contact change and the moving away trend of displacement change. The proximity phase representation, dwelling phase representation, and withdrawal phase representation are written into the segment representation sequence in chronological order to form a contact phase sequence. The segment recursive update layer recursively updates the segment representation sequence and the contact phase sequence in chronological order. Between adjacent behavioral segments, the approach phase representation, dwell phase representation, and evacuation phase representation are updated together with the recursive hidden state. The sequential connection relationship of the approach phase representation, dwell phase representation, and evacuation phase representation is matched between consecutive behavioral segments. When consecutive behavioral segments sequentially satisfy the approach phase connection, dwell phase connection, and evacuation phase connection, the corresponding consecutive behavioral segments are merged into the same recursive segment chain. When the phase connection relationship is interrupted, the continuation of the recursive segment chain is terminated, and a phase recursive sequence is generated according to the chronological order of the recursive segment chain. The event generation layer merges continuous behavior segments into drinking events based on the start and end positions of the continuous behavior segments corresponding to the relay segment chain in the phase relay sequence, and arranges them in chronological order to form a drinking event sequence.

5. The intelligent drinking water behavior management method based on deep learning according to claim 1, characterized in that, Specifically, S4 is: Extract the start time, end time, duration, and time interval between adjacent drinking water events from the drinking water event sequence to construct an event feature set; An event feature vector is generated based on the arrangement of the start time, end time, duration, and time interval between adjacent drinking events in the event feature set, and then the event feature vector is mapped to the event feature space. The event feature vectors in the event feature space are used as mean-shifted sample points, and the set of neighborhood sample points is determined based on the distance relationship between the sample points. The sample points in the neighborhood sample point set are sorted according to the time order in the drinking water event sequence, and adjacent sample points are connected in sequence to form a migration path. The local density center is calculated based on the set of neighboring sample points, and the moving direction and position of the sample points are adjusted in combination with the migration path to obtain the updated sample point position. Repeat the calculation of local density centers, adjustment of migration paths, and updating of sample point positions until the sample point positions converge; merge the converged sample point positions to generate a set of drinking water event centers.

6. The intelligent drinking water behavior management method based on deep learning according to claim 1, characterized in that, Specifically, S5 is: The drinking water event centers are arranged according to the time order in the drinking water event center set to obtain an ordered drinking water event center sequence; The time interval between adjacent drinking water event centers is calculated based on the time positions of adjacent drinking water event centers in the ordered drinking water event center sequence, thus obtaining the time interval sequence. Based on the duration of adjacent drinking water event centers in the ordered drinking water event center sequence, the duration change relationship is calculated to obtain the duration change sequence. The relationship between event morphology changes is calculated based on the event feature vectors of adjacent drinking water event centers in the ordered drinking water event center sequence, and the event morphology change sequence is obtained. By splicing together the time interval sequence, the duration change sequence, and the event pattern change sequence in chronological order, a drinking behavior rhythm sequence is generated.

7. The intelligent drinking water behavior management method based on deep learning according to claim 1, characterized in that, Specifically, S6 is: Arrange the time interval sequence, duration change sequence, and event pattern change sequence in chronological order to form a rhythm feature vector sequence; Based on the user's historical drinking behavior baseline, the offsets of time interval, duration and event pattern changes in the rhythm feature vector sequence are calculated respectively to obtain the time interval offset sequence, duration offset sequence and event pattern offset sequence. The time interval offset sequence, duration offset sequence, and event pattern offset sequence are spliced ​​together in chronological order to generate a drinking behavior state sequence. The drinking behavior state is determined based on the drinking behavior state sequence, and the drinking water management state is generated.

8. The intelligent drinking water behavior management method based on deep learning according to claim 1, characterized in that, The process of generating drinking water control quantities by mapping the state based on the drinking water management status is as follows: Extract time interval offset sequence, duration offset sequence, and event pattern offset sequence from the drinking behavior state sequence; The time interval change sequence is obtained by calculating the difference between adjacent elements in the time interval offset sequence according to the time order; the duration change sequence is obtained by calculating the difference between adjacent elements in the duration offset sequence according to the time order; and the behavior change direction sequence is determined based on the change direction of adjacent elements in the event pattern offset sequence. By combining the time interval change sequence, the duration change sequence, and the behavior change direction sequence in chronological order, a control feature sequence is obtained. The time interval changes in the control feature sequence are accumulated to generate a time adjustment quantity; the duration changes in the control feature sequence are accumulated to generate a frequency adjustment quantity; the time adjustment quantity and the frequency adjustment quantity are arranged in chronological order to generate a drinking water control quantity.

9. The intelligent drinking water behavior management method based on deep learning according to claim 1, characterized in that, The process involves rearranging the time intervals of drinking events based on drinking water control quantities to generate a drinking water scheduling sequence, and then outputting the drinking water behavior management results according to the drinking water scheduling sequence. Specifically: Based on the time adjustment factor in the drinking water control quantity, the time interval of drinking water events is increased or decreased to generate an adjusted time interval sequence; The time positions in the drinking water event center set are rearranged according to the adjusted time interval sequence to generate a drinking water scheduling sequence. The drinking water reminder time is generated according to the time position in the drinking water scheduling sequence, the drinking water frequency guidance information is generated according to the time interval in the drinking water scheduling sequence, and the behavior feedback information is generated according to the correspondence between the drinking water scheduling sequence and the drinking water management status. The drinking behavior management results are output based on drinking reminders, drinking frequency guidance information, and behavioral feedback information.

10. A deep learning-based intelligent drinking water behavior management system, executing the deep learning-based intelligent drinking water behavior management method according to any one of claims 1 to 9, characterized in that, Includes the following modules: The drinking water data module is used to acquire user drinking water behavior data, synchronize and arrange it over time, and generate a drinking water behavior data stream. The feature construction module is used to extract behavioral features from drinking behavior data streams, construct behavioral feature sequences, and segment the behavioral feature sequences using a sliding segmentation method to generate behavioral segment sequences. The evolution modeling module is used to input the behavior segment sequence into the improved SegRNN model, introduce a contact phase evolution mechanism in the segment recursive update layer, perform phase evolution transmission on adjacent behavior segments, and generate a drinking event sequence based on the phase evolution results of consecutive behavior segments. The central convergence module is used to construct an event feature set based on the drinking water event sequence. It adopts the MeanShift clustering algorithm that introduces the order tide migration center to form a migration path according to the time order of events during the mean drift process, and performs central convergence on the drinking water events to generate a central set of drinking water events. The rhythm generation module is used to arrange the set of drinking event centers, calculate the time interval relationship between adjacent drinking event centers, and generate a drinking behavior rhythm sequence. The state generation module is used to construct the drinking behavior state based on the drinking behavior rhythm sequence and generate the drinking management state. The scheduling output module is used to generate drinking water control quantities by mapping the state according to the drinking water management status, rearrange the drinking water event time intervals based on the drinking water control quantities to generate a drinking water scheduling sequence, and output the drinking water behavior management results according to the drinking water scheduling sequence.

Images