An intent hotspot vector-based cross-scene AI conversation context management method

Abstract

The present application relates to the field of artificial intelligence and natural language processing, in particular to a cross-scene AI conversation context management method based on intention hot spot vectors, mapping multi-modal data to generate intention hot spot vector sequences, comparing time slices to determine turning points and cut conversation semantic continuous fragment lists, checking state acceptance directions to filter and generate cross-scene conversation switching time node sets, determining semantic acceptance relationships to splice and generate cross-scene continuous conversation context sequences, combining privacy budget consumption to form constraint relationships and generate conversation context management views. The present application establishes an evolution representation through intention progressive relationships, completes segmentation by combining change feature identification semantic intervals, determines cross-scene switching timing and reduces false association by correlating fracture positions and cache rounds, forms continuous conversation records based on semantic acceptance verification, synchronously manages state resources by combining device sources and privacy budget consumption, and enhances the order, consistency and traceability of cross-scene conversations.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing, and in particular to a cross-scenario AI conversation context management method based on intent hotspot vectors. Background Technology

[0002] The field of artificial intelligence and natural language processing includes technical methods for semantic understanding and context modeling of multi-source interactive data. The core content of this field includes multimodal data acquisition and fusion, semantic representation construction, user intent recognition, dialogue context management, and cross-scene information transfer processing. The overall technical field involves unified encoding and correlation analysis of text data, voice data, and behavioral feature data. By constructing user cognitive graphs, extracting semantic relevance features, calculating spatiotemporal behavioral patterns, and maintaining dynamic context states, it realizes a systematic processing process for semantic continuity and scene relevance in multi-turn conversations. This field also covers key technical aspects such as context fragment filtering, context window pruning, data encryption processing, and cross-scene data scheduling.

[0003] One method for cross-scenario AI conversation context management based on intent hotspot vectors refers to a method that generates intent hotspot vectors by constructing a user cognitive graph during multi-scenario interactions, and dynamically manages the conversation context by combining privacy budget consumption information. The main technical aspects covered include implicit intent mining, intent hotspot vector generation, cross-scenario context prediction, context fragment filtering and reorganization, sensitive data encryption, and a privacy budget-based feedback adjustment mechanism. Specifically, this involves collecting multimodal behavioral features and calculating the correlation degree of potential intents using temporal causal relationships; mapping high-probability intents to graph nodes and calculating nonlinear spatiotemporal weights between nodes to form hotspot subgraph vectors; combining location data, environmental fingerprint data, and historical behavioral patterns; calculating scene switching probabilities through tensor operations and filtering matching context indices; performing compression and sorting processing on context fragments based on terminal computing power; mapping sensitive data to large integer domain elements and constructing encrypted state data using weighted exponential coefficients and adaptive moduli; constructing a feedback scoring function based on differential privacy budget consumption values; and dynamically adjusting encryption parameters and context processing strategies.

[0004] Existing technologies focus on intent correlation calculation, hotspot subgraph construction, and scene switching probability assessment. In actual cross-scene interaction, the results of context fragment filtering rely more on scene matching relationships and historical behavioral characteristics. When user intent is in a continuously changing state or scene transitions occur frequently, it is easy to have unclear semantic boundaries and excessively large context fragment correlations, which leads to content from different stages being classified into the same context. For example, in scenarios such as continuous queries and cross-device connection continuation, there is a risk of context misconnection and content mixing. At the same time, privacy budget feedback is mainly for processing strategy adjustments and is difficult to directly reflect the correspondence between continuous session content and resource consumption, affecting the consistency of subsequent session state tracking and management processes. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides a cross-scenario AI session context management method based on intent hotspot vectors. The technical solution is as follows: A cross-scenario AI session context management method based on intent hotspot vectors includes the following steps: S1: Collect touch coordinates, speech recognition text, gaze coordinates and key timestamp sequences, normalize them to a unified timeline, divide them into time slices, and progressively update candidate intent states according to the boundaries of speech, touch, gaze and key, match preset intent labels and hotspot intensity, and generate intent hotspot vector time series; S2: Based on the progressive relationship of candidate intent states in the time series of intent hotspot vectors, compare the label direction, hotspot intensity and continuity of adjacent time slices, identify the reverse or interruption position as the turning point of change, merge the time slices without turning points, divide the unified time axis, and generate a list of continuous segments of conversation semantics. S3: Call the list of continuous semantic segments of the session, obtain the session context cache index table and the user session round time series, check the direction of the transition according to the turning point of the change, filter the break position of the direction change of the cache position change or the change of the session round as the switching node, and generate a cross-scene session switching time node set. S4: Based on the cross-scene session switching time node set, use the switching node marker as the anchor point to associate the semantic continuous segments of the previous and subsequent sessions, trace back the speech recognition text and candidate intent state, compare the semantic and state continuity of the beginning and end, retain the segments that can be continued, and splice them according to the device source and round order to generate a cross-scene continuous session context sequence. S5: Invoke the cross-scenario continuous session context sequence, match the session identifier, edge node identifier and device source according to the continuous session fragment marker, associate privacy budget consumption records, classify them according to time order, device source and session round, form privacy budget feedback constraint relationship, and generate a session context management view.

[0006] As a further aspect of the present invention, the intent hotspot vector time series includes a preset intent label, a hotspot intensity value, a time slice identifier, and a candidate intent state; the conversation semantic continuous segment list includes a segment number, a segment start time, a segment end time, and a semantic fluctuation identifier; the cross-scenario conversation switching time node set includes a switching node marker, a node time position, a cache position identifier, and a conversation round identifier; and the cross-scenario continuous conversation context sequence includes a continuous conversation segment marker, a device source identifier, a voice / text segment, and a candidate intent state record.

[0007] As a further aspect of the present invention, the step of obtaining S1 is as follows: S11: Collect touch coordinate sequences, speech recognition text sequences, gaze tracking coordinate sequences, and key timestamp sequences generated during terminal interaction. Organize touch coordinate times, speech / text times, gaze coordinate times, and key trigger times. Map each time item to a unified timeline according to the order of collection. Divide the timeline into continuous time slices and generate a time slice mapping list. Details are as follows: S12: Based on the time slice mapping list, read the voice keywords, keyword occurrence time, and touch displacement trajectory within the current time slice. Determine the candidate intent direction according to the order of keyword occurrence. Compare the touch displacement trajectory with the candidate intent direction, retaining candidate intents within the consistent range to obtain the candidate intent convergence interval; specifically as follows: S13: Call the candidate intent convergence interval, associate the gaze dwell area, duration and key trigger time, confirm the candidate intent attention object according to the gaze dwell area, mark the interaction completion boundary according to the key trigger time, inherit the candidate intent state of adjacent time slices in time sequence and match the preset intent label and hotspot intensity to generate an intent hotspot vector time series.

[0008] As a further aspect of the present invention, the step of obtaining S2 is as follows: S21: Call the candidate intent state progression relationship in the intent hotspot vector time series, read the previous time slice, the next time slice, the preset intent label and the hotspot intensity along the unified time axis, compare the preset intent labels corresponding to adjacent time slices, record the change direction according to the label succession relationship, and generate the label change trajectory according to the increase or decrease of the hotspot intensity status mark change amplitude. S22: Based on the label change trajectory, call the candidate intent state progression relationship, the candidate intent state of the previous time slice and the candidate intent state of the next time slice, compare the state continuation direction with the label change direction, and when the label direction is reversed or the continuation path is interrupted, record the corresponding time position and associate it with the adjacent time slices to obtain the semantic fluctuation interval. S23: Invoke the semantic fluctuation range and the label change trajectory, read the continuous time slices not marked as fluctuation boundaries, the succession order of candidate intention states and the change inflection points, determine the succession relationship of candidate intention states in adjacent time slices, merge the continuous time slices with consistent succession into continuous semantic segments of the conversation, divide the unified time axis according to the change inflection points, and generate a list of continuous semantic segments of the conversation.

[0009] As a further aspect of the present invention, the step of obtaining S3 is as follows: S31: Call the continuous segment of the session semantics in the list of continuous segment of the session semantics, obtain the session context cache index table consisting of session identifier, context segment identifier, cache location and cache time, and user session round time sequence consisting of session round number, round start and end time and device source, arrange adjacent segments along a unified time axis according to the turning point of change, and generate a session reference list; S32: Based on the session reference list, call the session context cache index table, user session round time sequence, semantic continuous segments of adjacent sessions and semantic fluctuation interval, check the deviation of the candidate intent state continuation direction in the semantic fluctuation interval, determine whether the deviation directions before and after the change inflection point are opposite, and obtain the direction change break position. S33: Invoke the direction change break position and the session reference list, read the cache position and cache time in the session context cache index table, read the session round number, round start and end time and device source in the user session round time series, filter the time position corresponding to the cache position change or session round change, set the switching node mark, and generate a cross-scene session switching time node set.

[0010] As a further aspect of the present invention, the process of checking the deviation of the candidate intent state continuation direction within the semantic fluctuation range is specifically as follows: Read the candidate intent state succession direction corresponding to the consecutive semantic segments of the preceding and following sessions, and align the candidate intent state succession direction with time according to the unified time axis. Compare the candidate intent state succession directions of adjacent time slices within the semantic fluctuation interval item by item. When the number of consecutive time slices with inconsistent directions reaches a preset deviation threshold, mark the corresponding time range as a deviation interval. The process of determining whether the deviation directions before and after the turning point of change are opposite is as follows: Read the deviation interval before the change inflection point and the deviation interval after the change inflection point respectively. Take the candidate intention state acceptance direction corresponding to the deviation interval before the change inflection point as the forward deviation direction and the candidate intention state acceptance direction corresponding to the deviation interval after the change inflection point as the backward deviation direction. When the forward deviation direction is opposite to the backward deviation direction, record the time position corresponding to the change inflection point as the direction change break position.

[0011] As a further aspect of the present invention, the process of filtering and matching the time position corresponding to changes in cache location or session rounds is specifically as follows: Read the time position corresponding to the direction change break position, the cache position, and the cache time, and read the session round number, the round start and end time, and the device source. When the time position corresponding to the direction change break position is consistent with the cache time when the cache position changes, or is consistent with the round boundary time in the round start and end time, retain the time position corresponding to the direction change break position and set the switching node marker.

[0012] As a further aspect of the present invention, the step of obtaining S4 is as follows: S41: Call the switching node marker in the cross-scene session switching time node set, use the position of the switching node marker in the unified time axis as the anchor point, associate the continuous semantic segments of the session before and after the switching node, trace back the speech recognition text and candidate intent state corresponding to each continuous semantic segment of the session, compare the order of the segments according to the start and end times of the segments, and generate the anchor point segment index. S42: Based on the anchor segment index, read the ending semantics of the semantic continuous segment of the previous session, the starting semantics of the semantic continuous segment of the next session, and the corresponding candidate intent states of the two. According to the time sequence, determine the continuation relationship of voice keywords, the corresponding situation of referential relationship, and the continuity of candidate intent states. Retain the consistent segment combination to obtain the semantic continuation segment combination. S43: Invoke the semantic continuation fragment combination and the anchor fragment index, read the switching node mark associated with the device source and the session round identifier, arrange the fragment combination according to the order of device source and session round identifier, filter adjacent fragments and splice them, set continuous session fragment mark on the splicing result, and generate a cross-scene continuous session context sequence.

[0013] As a further aspect of the present invention, the method further includes: S5: Invoke the cross-scenario continuous session context sequence, match the session identifier, edge node identifier and device source according to the continuous session fragment marker, associate privacy budget consumption records, classify them according to time order, device source and session round, form privacy budget feedback constraint relationship, and generate a session context management view; The session context management view includes continuous session segment records, device identification information, privacy budget consumption status, and session round records; As a further aspect of the present invention, the step of obtaining S5 is as follows: S51: Call the continuous session fragment marker in the cross-scenario continuous session context sequence, obtain the privacy budget consumption record sequence table consisting of continuous session fragment marker, device identifier, consumption amount and consumption time, and the edge node session identifier mapping table consisting of edge node identifier, session identifier and device source, and generate a budget mapping list by comparing the session identifier and device source according to the continuous session fragment marker. S52: Based on the budget mapping list, read the continuous session fragment marker, session identifier, edge node identifier, device source, consumption amount and consumption time, align the session identifier with the edge node identifier item by item, filter records with consistent device source, associate with the privacy budget consumption record sequence list, and mark the privacy budget consumption status in different device sources and session rounds to obtain the fragment consumption marker; S53: Invoke the fragment consumption flag and the budget mapping list, read the occurrence record of the same continuous session fragment, session round record, privacy budget consumption status, device source and session round identifier, arrange the occurrence record in the order of consumption time, classify the consumption status according to the device source, group and arrange the budget feedback constraint relationship according to the session round identifier, and generate a session context management view.

[0014] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this invention, an intent evolution representation is established by the progressive relationship of candidate intent states in continuous time slices. The semantic fluctuation range and stable segment are identified by combining the direction of change, the magnitude of change, and the path continuity. This enables the conversation content to be segmented and organized according to the actual cognitive continuation process. The timing of cross-scene switching is determined by combining the relationship between the break position of the direction change and the cache position and the conversation rounds. This reduces the phenomenon of context misconnection and misassociation during the scene transition. Furthermore, the continuous conversation segment record is formed by verifying the semantic continuity relationship. The continuous conversation segment is uniformly associated with the device source, the conversation rounds, and the privacy budget consumption status, realizing the synchronous management of the context continuity status and the resource consumption status. This enhances the organization, consistency, and traceability of cross-scene conversation content. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart illustrating the acquisition process of S1 in this invention; Figure 3 This is a flowchart illustrating the acquisition process of S2 in this invention; Figure 4 This is a flowchart illustrating the acquisition process of S3 in this invention; Figure 5 This is a flowchart illustrating the acquisition process of S4 in this invention; Figure 6 This is a flowchart of the acquisition process for S5 of the present invention. Detailed Implementation

[0016] The following detailed explanation illustrates the specific implementation methods: The basic implementation examples are as follows: Figure 1 As shown: A cross-scenario AI session context management method based on intent hotspot vectors includes the following steps: S1: Collect the touch coordinate sequence, speech recognition text sequence, gaze tracking coordinate sequence, and key timestamp sequence generated during terminal interaction. Normalize the timestamps of each sequence and map them to a unified time axis. Divide the time axis into continuous time slices. Within the current time slice, determine the direction of candidate intent based on speech keywords and their occurrence time. Combine the touch displacement trajectory to narrow down the range of candidate intents. Confirm the object of attention of candidate intents based on the gaze dwell area and duration. Determine the interaction completion boundary of the current candidate intent based on the key trigger time. Transform the candidate intent determined by the interaction completion boundary into the candidate intent state of the current time slice. Receive and update the candidate intent state of adjacent time slices in chronological order. Progressively pass along the continuous time slices. Match the preset intent label and hotspot intensity based on the progressive result to generate an intent hotspot vector time sequence. The intent hotspot vector time series includes preset intent labels, hotspot intensity values, time slice identifiers, and candidate intent states.

[0017] Please see Figure 2 Specifically, S1 includes: S11: Collect touch coordinate sequences, speech recognition text sequences, gaze tracking coordinate sequences, and key timestamp sequences generated during terminal interaction. Organize touch coordinate times, speech / text times, gaze coordinate times, and key trigger times. Map each time item to a unified timeline according to the order of collection. Divide the timeline into continuous time slices and generate a time slice mapping list. Details are as follows: The system collects touch coordinate sequences, speech recognition text sequences, gaze tracking coordinate sequences, and button timestamp sequences generated during terminal interactions. It utilizes hardware-level components such as a capacitive touchscreen panel (with a physical reporting rate of 200Hz), a MEMS multi-microphone acoustic baseband array (using the I2S audio bus transmission protocol), an infrared non-contact pupil-corneal reflection gaze tracking camera (with a sampling frame rate of 120Hz), and physical piezoelectric tactile buttons to capture the raw physical signals of each dimension. From these, the system extracts the raw millisecond-level timestamps corresponding to each point in the touch coordinate sequence. Extract the starting offset time of each character identified by the Hidden Markov Model decoder in the speech recognition text sequence. With end offset time Obtain the absolute time record of pupil sampling data for each frame in the gaze tracking coordinate sequence. Read the instantaneous value of key trigger from the key timestamp sequence. ; Call the preset time base offset This value is set in advance based on the user's actual situation. A subtraction operation is performed on all extracted timestamps to obtain the relative duration value, and a global time ratio adjustment coefficient is set. , here This is a dimensionless time calibration constant; Through compound multiplication Each raw second-level time parameter Uniform time scalar converted to milliseconds According to the converted The data points are arranged on a unified time axis from smallest to largest. If there are abnormal and scattered data points that exceed the preset time limit, the memory is directly discarded and cleared. Select the start time point of the current interaction flow ms and termination time point ms, sets the fixed length parameter of the continuous time slice. The total number of dimensionless time slices is calculated by subtracting the start time from the total duration and then dividing by a fixed length. Each time slice interval is divided into equal segments along a unified time axis. Traverse all mapped coordinates and text records on a unified timeline. In a multi-concurrency case, when a touch coordinate is touched in the first concurrent interaction stream... The mapping time falls within the interval When the time is ms, the coordinate point is assigned to the third time slice set. When the word "zoom in" in the voice text "View current location and zoom in on the map" is mapped to the time ms, the coordinate point is assigned to the third time slice set. At the millisecond interval, associate it with the corresponding third time slice index bit to extract the line-of-sight coordinates. exist Record the keystrokes at the specified milliseconds (ms) and mark them to that time slice. exist The trigger action at ms, and in order to carry over the subsequent zooming operation, in the second concurrent interaction stream, the voice keyword "zoom in" that falls into the subsequent 5th time slice (800 to 1000ms) and its corresponding mapped timestamp and touch trajectory are extracted together. The key-value pair combination of each time slice number and the address of the multimodal data parameter attached inside is used to generate a time slice mapping list.

[0018] S12: Based on the time-slice mapping list, read the voice keywords, keyword occurrence time, and touch displacement trajectory within the current time slice. Determine the candidate intent direction according to the order of keyword occurrence. Compare the touch displacement trajectory with the candidate intent direction, retaining candidate intents within the consistent range to obtain the candidate intent convergence interval; as detailed below. Based on the time-slice mapping list, the voice keywords, keyword occurrence times, and touch displacement trajectories within the current time slice are read. The voice keyword "amplify" and its corresponding mapping timestamp within the current time slice are retrieved through the high-speed static storage register. ms, extract the starting coordinates of the touch trajectory generated by the touch screen electrode array within the same time slice. and termination coordinates The coordinate units are all pixels, and the horizontal pixel displacement component is obtained by subtraction. Pixel and vertical pixel displacement components: Pixels, using square root operations. The magnitude of the touch displacement vector is calculated as follows: Pixels yield a dimensionless touch direction normalized vector. The direction angle of the displacement vector The target direction is compared with a preset set of candidate intent directions. A preset speech direction mapping table is invoked to retrieve the dimensionless intent direction normalized vector corresponding to the keyword "amplify". Calculate the cosine of the angle between the touch displacement direction and the intention direction vector: ,in and These represent the corresponding intent vectors. Norm modulus value; Substitute the data to obtain Set a dimensionless cosine similarity threshold. The size judgment is performed because If the touch displacement is established and falls within the consistent range, then the touch displacement is determined to have continuity with the voice intent direction. If the cosine value calculated in other interference cases is less than or equal to... If the touch is determined to be isolated false touch noise, the system will directly clear the probability value of the candidate direction in memory to zero and stop the current intent convergence calculation to prevent false triggering. Specifically, the stop control method involves clearing the processing context of the current time slice and resetting the data buffer within the preset intent space. In the middle, the center direction angle corresponding to the voice keyword is taken as Set dimensionless search bandwidth weight parameters ; Through calculation The initial angle search range is determined, and the endpoint coordinates of the touch displacement are verified through consistency checks. As the geometric center, set the radius For a circular region of a pixel, a two-dimensional spatial coordinate intersection operation is performed on the overlapping area between this region and the initial search range. All candidate intent codes falling within the intersection region are retained, while those with an angle cosine value lower than 1 are excluded. The invalid interference direction is used to obtain the candidate intention convergence interval.

[0019] S13: Invoke the candidate intent convergence interval, associate it with the gaze dwell area, duration, and button trigger time, confirm the candidate intent's focus object according to the gaze dwell area, mark the interaction completion boundary according to the button trigger time, and sequentially connect the candidate intent states of adjacent time slices and match them with preset intent tags and hotspot strengths to generate an intent hotspot vector time series; as detailed below. The candidate intent convergence interval is invoked, and the gaze dwell area, duration, and key trigger time are associated to extract the cluster of gaze dwell pixel coordinates mapped within the convergence interval. It calls the interface element layout coordinate table stored in read-only memory and maps the coordinate point clusters to the boundary rectangles of each functional component in the interface. Perform a Boolean intersection operation on the overlapping ranges. The line-of-sight sampling point falls within the rectangular area of ​​the "map zoom control". Within the specified area, the control is marked as the object of current interest. The number of consecutive frames in which the gaze point exists within the same area is counted and multiplied by the frame period to obtain the duration of continuous existence. ms, setting a baseline value for significant attention duration ms, determined by size comparison operation This confirms that the level of attention belongs to the high-level attention range, and the button trigger time is read. ms, calculates the key press time and the end time of the voice keyword through subtraction. The difference in milliseconds ms, compare the difference with the preset boundary threshold The ms size comparison will satisfy... The key press time point in milliseconds is confirmed as a valid interaction completion boundary. If it exceeds this range, it is judged as an isolated timeout event, the key trigger flag of the current time slice is cleared in memory and no further processing is performed. The dimensionless candidate intention state vector for each time slice With the The state vector of each time slice The system performs memory matrix concatenation, calls a preset intent tag library, inputs the state vector into the matching matrix, and calculates the dimensionless Euclidean distance between the state vector and the tag "Map_Zoom_In". ,because Less than the preset classification threshold The time slice is labeled "Map_Zoom_In", and the dimensionless initial coefficient of the hotspot intensity is set. Based on the duration of eye contact and the normalized button pressure level collected by the thin-film pressure sensor. The product terms are added together using the formula. Calculate the dimensionless hotspot intensity here. For hotspot intensity, These are the initial coefficients. The unit is ms, and the denominator is... Used to convert time to seconds to achieve dimensionless alignment. For dimensionless pressure levels, substituting specific data yields... The tag encoding and strength value are combined. Combined into a multidimensional dense vector, it generates an intentional hotspot vector time series.

[0020] S2: Invoke the candidate intent state progression relationship in the intent hotspot vector time series, compare adjacent time slices one by one along the unified time axis, compare the preset intent labels corresponding to adjacent time slices, determine the direction of change of the preset intent label from the previous time slice to the next time slice, determine the change amplitude based on the increase or decrease trend of hotspot intensity, verify the path continuity of the time slices before and after through the candidate intent state progression relationship, when the change direction of the preset intent label reverses or the path continuity is interrupted, determine the corresponding time position as the change inflection point, and mark the adjacent time slices before and after the inflection point as semantic fluctuation intervals, merge the continuous time slices without change inflection points according to the candidate intent state continuity order to obtain the continuous semantic segments of the conversation, divide the unified time axis according to the change inflection points, and generate a list of continuous semantic segments of the conversation. The list of continuous semantic segments in a conversation includes the segment number, segment start time, segment end time, and semantic fluctuation identifier.

[0021] Please see Figure 3 Specifically, S2 includes: S21: Invoke the candidate intent state progression relationship in the intent hotspot vector time series, read the previous time slice, the next time slice, the preset intent label, and the hotspot intensity along a unified time axis, compare the preset intent labels corresponding to adjacent time slices, record the direction of change according to the sequential relationship of the labels, and generate the label change trajectory based on the increase or decrease of the hotspot intensity status marker. Details are as follows: The candidate intent state progression relationship is retrieved from the intent hotspot vector time series, and the previous time slice is read along a unified time axis. (400-600ms), the next time slice Extract (600-800ms), preset intent tags and hotspot strength from the memory pipeline buffer. Corresponding Integer Index Value of Intent Label (Representing map zoom-in) and hotspot intensity values ,extract Corresponding Integer Index Value of Intent Label With hotspot intensity value Perform comparison operation to determine and If they are equal, record the label's direction change value as 0 if they are equal, otherwise... The intent label corresponding to (800-1000ms) becomes (This represents the intention to play music); Perform a subtraction operation to obtain the tag index difference This difference is recorded as the direction of change in the relationship between the preceding and following tags. The absolute deviation of the intensity is obtained by performing a subtraction operation on the intensity of the hotspot. It invokes the preset hotspot intensity amplitude determination rules and sets the dimensionless amplitude change weighting coefficient. Weighting coefficients for label changes ; Through arithmetic formulas Calculate the overall dimensionless change value, where Represents the current time slice number. Represents dimensionless intensity. Represents a dimensionless label index. This is a dimensionless comprehensive change index value, specifically expressed as follows: and These are the dimensionless hotspot intensity values ​​for the next time slice and the current time slice, respectively. and These are the dimensionless integer index values ​​of the intent labels for the next time slice and the current time slice, respectively. Substitution and Data derived Set a dimensionless threshold for significant changes ,because In Within the closed interval, the change in the marker is at the "stable" level, meaning that the interaction state remains highly consistent and stable. Substituting... and Data and settings ; get ,because In Within the open interval, the change amplitude is marked as "drastic", representing a sudden jump. The direction value, amplitude level and corresponding timestamp interval between each adjacent time slice are arranged in time sequence and stored in memory as a structure array composed of the change value, amplitude level code and timestamp interval of the time-series adjacent time slices, generating the label change trajectory.

[0022] S22: Based on the label reversal trajectory, the candidate intent state progression relationship, the candidate intent state of the previous time slice, and the candidate intent state of the next time slice are invoked. The state continuity direction is compared with the label change direction. When the label direction reverses or the continuity path is interrupted, the corresponding time position is recorded, and adjacent time slices are associated to obtain the semantic fluctuation interval; as detailed below: Based on the label-based trajectory, the algorithm invokes the progressive relationship of candidate intent states, the candidate intent states of the previous time slice, and the candidate intent states of the next time slice, and extracts them using an arithmetic logic unit. Candidate intention state probability vector of time slice ,extract Candidate intention state probability vector of time slice Here, the vector components correspond to the intent distribution of the map, entertainment, and global settings, respectively. The resulting value is obtained by summing the vector point integrals. Set the dimensionless state acceptance consistency benchmark value as The size comparison judgment is performed because If the user is within an inconsistency interval, the current state transition path is considered interrupted, indicating a sudden, discontinuous shift in user attention (meaning the user is in a transitional period between different scenarios). The specific memory reconstruction mechanism is as follows: direct inheritance of the preceding session context features is paused; an independent temporary queue is created in the memory pipeline to isolate and suspend this uncertain data, preventing its propagation into subsequent normal states until the intent probability distribution of subsequent time slices converges and stabilizes, at which point the tag-based trajectory is read. to If the label changes direction by 4, a sign bit judgment is performed. If the change value changes from positive to negative or the value jumps to a magnitude exceeding the preset dimensionless threshold 3, it is determined that the label direction has reversed. At this time, the current unified time axis position is recorded. ms is the directional correlation point, and one time slice period preceding this correlation point is selected. ms and the next time slice period ms, obtained by performing a union operation on the time intervals. Within a time period of milliseconds, extract the jitter frequency of all touch trajectories and the correction markers of the speech recognition text. If the speech text contains correction keywords such as "no, is..." or touch coordinates within that time period, further processing is performed. If the movement reciprocates within a pixel range, the merged interval is marked as a semantically uncertain region, and the corresponding time start point is determined. ms and endpoint The ms value is written to the fluctuation record table to obtain the semantic fluctuation range.

[0023] S23: Invoke the semantic fluctuation range and label change trajectory, read the continuous time slices not marked as fluctuation boundaries, the succession order of candidate intent states, and the change inflection points, determine the succession relationship of candidate intent states in adjacent time slices, merge continuous time slices with consistent succession into continuous semantic segments of the conversation, divide the unified time axis according to the change inflection points, and generate a list of continuous semantic segments of the conversation; as detailed below: By invoking semantic fluctuation intervals and label change trajectories, reading continuous time slices not marked as fluctuation boundaries, candidate intent state succession order, and change inflection points, and utilizing the timeline control handle of the central processing unit, extracting data from a unified timeline... to Continuous time slices not marked as fluctuations within ms The intent tag indexes corresponding to each slice are read sequentially. Perform an equality check on adjacent labels and confirm. and ,as well as and The dimensionless succession relationship scores are all Because this score is in Within the interval, if the sequence is deemed consistent, the set merging action will be executed. Encapsulate it as an independent, continuous semantic unit; if the acceptance score is lower than... Then isolate them as isolated conflict fragments and do not merge them, extracting the locations of the turning points of change. ms, at this position, perform timeline splitting operation, to ms is divided into the first semantic region, and After ms, the score was restored to The above sequence Divide into second semantic intervals, and call the arithmetic mean of the dimensionless hotspot intensity within each interval. If the average intensity of the first interval is... The average intensity of the second interval is The centroid time points for each interval are calculated separately. Specifically, the calculation method involves taking a weighted average of the center timestamps of all time slices within the interval based on their corresponding dimensionless hotspot intensities. The centroid time point is used as the geometric center reference point of each semantic block on the time axis for subsequent execution of temporal distance span verification. Each segmented semantic unit is structurally associated with the corresponding dimensionless average intensity, label type, and start and end time to form a semantic block object with a unique string identifier. The above semantic blocks are stored in a linear linked list in chronological order on the time axis to obtain a list of continuous semantic fragments of the conversation.

[0024] S3: Call the session semantic continuous fragments in the session semantic continuous fragment list, obtain the session context cache index table consisting of session identifier, context fragment identifier, cache position and cache time, and the user session round time sequence consisting of session round number, round start and end time and device source. Along a unified time axis with the change inflection point as a reference, check the direction of the candidate intent state of adjacent session semantic continuous fragments. When the direction of the candidate intent state continues to deviate within the semantic fluctuation range and the deviation direction before and after the change inflection point is inconsistent, record the corresponding time position as the direction change break position. Combine the session context cache index table and the user session round time sequence, filter the direction change break positions corresponding to the cache position change or session round change as switching nodes, set the switching node mark, and generate a cross-scene session switching time node set. The cross-scenario session switching time node set includes switching node marker, node time position, cache position identifier, and session round identifier.

[0025] Please see Figure 4 Specifically, S3 includes: S31: Retrieve the session semantic continuous segments from the session semantic continuous segment list, obtain the session context cache index table consisting of session identifier, context segment identifier, cache location, and cache time, and the user session round time sequence consisting of session round number, round start and end time, and device source, arrange adjacent segments along a unified time axis according to the turning point, and generate a session reference list; as follows: Each session semantic contiguous segment in the session semantic contiguous segment list is invoked, and the session identifier encoded by the session semantic contiguous segment is retrieved from the cache memory via the high-speed memory controller. Context fragment identifier Memory cache physical address and timestamp records The session context cache index table is composed of ms, and the session round number is extracted. Start and end time intervals ms, device source label "vehicle dashboard" and wheel number Start and end time intervals The user session round time series of the device source tag "central control touch screen" (ms) is used to extract the change inflection points determined in S23. ms, using this turning point as the numerical anchor, retrieves the semantic continuum of the previous session. The last offset (corresponding to map zoom intent) and the semantic continuum of the next session. The starting offset (corresponding to the intended music switch) will Intervals on a unified timeline and Intervals on a unified timeline ms are linearly arranged and read. The corresponding dimensionless intentional state vector terminal value ; Read The corresponding dimensionless intentional state vector initial value The time boundaries, device source attributes, and associated intent hotspot intensity values ​​of the two segments are sequentially filled into a preset reference matrix for multi-dimensional dense matrix block encapsulation. The globally unique sequence number of the current session is set to a dimensionless integer. The gaps between segments The ms is marked as a gap to be checked. The relative coordinate distribution of the two segments on the same time axis is obtained by coordinate subtraction. The structured data items containing all the above metadata information are written into the cache list in ascending order of time to generate a session reference list.

[0026] S32: Based on the session reference list, the session context cache index table, user session round time series, semantically continuous segments of adjacent sessions, and semantic fluctuation intervals are called to check the deviation of the candidate intent state's direction of transition within the semantic fluctuation interval, determine whether the deviation directions before and after the turning point are opposite, and obtain the position of the break in direction change; specifically as follows: The process of checking whether the direction of the candidate intent state deviates within the semantic fluctuation range is as follows: Read the candidate intent state succession direction corresponding to the semantic continuous segments of adjacent sessions, and align the candidate intent state succession direction in time according to a unified time axis. Compare the candidate intent state succession direction of adjacent time slices within the semantic fluctuation interval item by item. When the number of consecutive time slices with inconsistent directions reaches the preset deviation threshold, mark the corresponding time range as the deviation interval. The specific process for determining whether the deviation directions before and after a turning point are opposite is as follows: Read the deviation intervals before and after the change inflection point respectively. Take the direction of the candidate intention state corresponding to the deviation interval before the change inflection point as the forward deviation direction and take the direction of the candidate intention state corresponding to the deviation interval after the change inflection point as the backward deviation direction. When the forward deviation direction and the backward deviation direction are opposite, record the time position corresponding to the change inflection point as the direction change break position. Based on the session reference list, the session context cache index table, user session round time series, semantically continuous segments of adjacent sessions, and semantic fluctuation intervals are invoked to retrieve the fluctuation intervals. Multimodal data sampling points within milliseconds; Read the previous segment The dimensionless intention to carry the direction unit vector Read the next segment The dimensionless intention to carry the direction unit vector Within the fluctuation range, an interrupt is triggered using a hardware timer at fixed intervals. The instantaneous intentional state direction is extracted in milliseconds, resulting in a unit vector sequence of the observation direction. Through dot product algebra operations Calculate the cosine of the angle between the dimensionless instantaneous direction and the direction of the previous segment, where... For the first The dimensionless cosine of the included angle at each observation point For the first Unit vector in each observation direction, This serves as a unit vector to carry over the dimensionless intent from the previous segment. and They represent the corresponding vectors. Norm modulus, if the cosine value is in If the deviation occurs within a certain interval, it is determined to be due to inconsistent direction. Statistical analysis revealed that three consecutive time slices showed inconsistent direction. Therefore, a dimensionless integer threshold for the number of deviations was set. Perform a comparison operation to determine Established, within a time frame ms is marked as the forward deviation interval, and the same method is used to... Calculate the dimensionless cosine value after the inflection point based on the baseline, and then... ms is marked as the backward deviation interval, and the average direction unit vector of the forward deviation interval is extracted. Extract the average direction unit vector of the backward deviation interval. ; Performing the dot product operation yields The dimensionless threshold for determining opposite directions is set to a negative range. The comparison decision was performed because This is a serious deviation, indicating that the user is issuing a physical jump switching command (i.e., a discontinuous cross-device switching behavior of multimodal intent in the spatial dimension). The specific suspension protection process is as follows: the central processing unit sends a data suspension command to the currently active in-vehicle instrument panel hardware interface controller, locks the state stack register in the cache to prevent context loss, and sends a high-level synchronous interrupt signal to the target central control touchscreen to activate its data receiving mechanism, recording the turning point at this moment. The point at ms is the physical transition point of the semantic features. Its coordinate position is marked in the time axis mapping table to obtain the position of the direction change break.

[0027] S33: Invoke the break location of the call direction change and the session reference list, read the cache location and cache time from the session context cache index table, read the session round number, round start and end time and device source from the user session round time series, filter the time positions corresponding to the cache location change or session round change, set the switching node marker, and generate a cross-scenario session switching time node set; as detailed below: The process of filtering and caching changes or session round changes corresponding to time positions is as follows: Read the time position, cache position and cache time corresponding to the direction change break position, and read the session round number, round start and end time and device source. When the time position corresponding to the direction change break position is consistent with the cache time when the cache position changes, or is consistent with the round boundary time in the round start and end time, retain the time position corresponding to the direction change break position and set the switching node flag. The call direction changes, the break location is determined, and the session reference list is accessed. The cache location and cache time are read from the session context cache index table, and the current break location time point is extracted. ms, retrieve the most recent physical memory address switch record controlled by the memory management unit from the cache index table, and find that the memory address is in ms from dashboard mapping space Jump to central control mapping space Calculate the time difference between the two. ms, retrieve the user session round time series, and identify the first round. The end boundary time is ms, second round The initial boundary time is ms, defining the time boundary matching threshold at the ms level as ms; Perform subtraction operation ms, due to In Within the closed interval, and Also in Within the closed interval, the logical triggering conditions of cache position change or round boundary coincidence are met. If the condition is exceeded, it is judged as invalid touch jitter, the switching weight assignment at the break position is cleared in memory, the triggering is refused, the generation of cross-scene session switching time node set is terminated, and the time position is retained. ms, read the hardware conversion tag indicating the device source switched from "dashboard" to "central control screen", and read the session identifier. With context fragment identifier and Perform an attribute assignment operation at the fracture location, assigning a dimensionless switching weight value to that point. The switching type code is set to "cross-scene physical switching". A hardware synchronization interrupt signal flag is inserted into the timeline index structure. All time points with switching attributes and their associated round conversion metadata are integrated to generate a cross-scene session switching time node set.

[0028] S4: Call the switching node marker in the cross-scene session switching time node set, take the position of the switching node marker in the unified time axis as the anchor point, associate the continuous semantic segments of the session before and after the switching node, and backtrack the speech recognition text and candidate intent state corresponding to each continuous semantic segment of the session. According to the temporal comparison, the ending semantics of the previous continuous semantic segment, the starting semantics of the next continuous semantic segment, and the candidate intent state corresponding to the two are compared to determine the semantic continuity relationship before and after the switching node. When there is a corresponding speech keyword continuation or referential relationship, and the candidate intent state remains continuous, the combination of continuous semantic segments of the session before and after the switching node is retained. The retained combination of continuous semantic segments of the session is sequentially spliced ​​according to the device source and session round identifier associated with the switching node marker. The continuous session segment marker is set on the sequential splicing result to generate a cross-scene continuous session context sequence. The cross-scenario continuous session context sequence includes continuous session segment markers, device source identifiers, voice and text segments, and candidate intent state records.

[0029] Please see Figure 5 Specifically, S4 includes: S41: Call the switching node markers in the cross-scene session switching time node set, use the position of the switching node markers in a unified timeline as anchor points, associate the continuous semantic segments of the session before and after the switching node, trace back the speech recognition text and candidate intent state corresponding to each continuous semantic segment of the session, compare the preceding and following sequences according to the start and end times of the segments, and generate an anchor point segment index; as detailed below: Retrieve the switching node markers from the cross-scene session switching time node set and extract the physical location values ​​of the switching nodes on a unified timeline. ms, using this value as the central anchor point, perform a bidirectional depth-first search operation, backtracking to the left to read consecutive semantic segments of ended conversations. Extract the physical address of the text in the memory cache from the text buffer of the automatic speech recognition engine. The data is used to obtain the speech recognition text string "View current location and zoom in on the map", and then backtracking to the right to read the newly started continuous semantic segments of the conversation. Extract its physical address in memory cache The data is used to obtain the speech recognition text string "Switch to music playback interface", which is retrieved from the sequence generated in S13. End time slice The corresponding dimensionless candidate intention state probability vector The values ​​here correspond to the probability distribution of intents for map zooming, media control, and environment configuration, respectively. Start time slice The corresponding dimensionless candidate intention state probability vector Perform a comparison operation on the difference between the start and end times of the segment, and calculate... The end time ms and The start time The difference in milliseconds ms, setting the time proximity decay coefficient Through the exponential formula The calculation time carries weight, among which The time similarity weights are dimensionless. The decay constant is a constant with the reciprocal dimension of time. The time difference is in the milliseconds range; the result is obtained by substituting specific data. Set a dimensionless threshold range for determining correlation strength. In It is then determined to be in a "closely related" state. It is then determined to be in a "loosely correlated" state; if it is lower than... The system is then determined to be in a completely broken state and the context link is released. The current value... Determined to be in a "closely related" state, the metadata pointer encapsulation action is executed, and... and The identifiers, intent state components, and device source labels for "vehicle instrument panel" and "central control touchscreen" are written into the index entries in chronological order, establishing a system based on... Using milliseconds (ms) as the core semantic connection path, an anchor fragment index is generated.

[0030] S42: Based on the anchor fragment index, read the semantic keyword set K at the end of the semantic continuum fragment Seg_1 from the previous session. end={map, zoom in}, read the semantic keyword set K at the beginning of the semantic continuum Seg_2 in the next session. start ={Switch, Music}, calls the semantic domain association matrix preset in the domain knowledge graph hash table, and extracts the mutual information entropy value I between the "map" domain and the "music" domain. info =0.08 bits, the lexical parser performs a check on the correspondence of pronouns, scanning the Seg_2 text content for specific personal pronouns or demonstrative pronouns. No pronouns such as "it" or "this" were matched in the "Switch to music playback interface" section, so the dimensionless pronoun matching gain coefficient G was directly set. ref =0, retrieve the candidate intent state probability vector V e1 With V s2 Perform cosine similarity algebra operation, where V e1 and V s2 V represents the L2 norm magnitude of the intent probability vector, calculated as follows: e1 ≈0.8590 and V s2 ≈0.7714.

[0031] Cosθ = ( V e1 · V s2 ) / (| V e1 | × | V s2 |) Cosθ = 0.2205 / (0.8590 × 0.7714) ≈ 0.3328 To ensure that mutual information, referential matching gain, and cosine similarity have a unified comparison scale in the comprehensive scoring, the mutual information entropy value I is... info Perform dimensionless normalization to obtain normalized mutual information I. ′info The I ′info The value is located in the interval [0,1]; the I ′info Used to characterize the association strength of preceding and following semantic keyword sets within a predefined semantic domain association matrix, I ′info The larger the value, the stronger the semantic continuity between the preceding and following segments. The referential matching gain coefficient G... ref G is a dimensionless quantity, taking values ​​in the interval [0,1]. When a clear reference to the semantic object at the end of the preceding conversation's semantic segment is detected in the subsequent semantic segment, G... ref Take the higher value; when no corresponding referent is detected, G ref Take a lower value or zero. In this embodiment, no referential terms such as "it" or "this" were detected, so G is taken. ref=0. An interval mapping is performed on the cosine similarity Cosθ to obtain the state continuity score C. state So that its value is in the interval [0,1]; the C state C is used to characterize the degree of directional consistency between the candidate intent state vectors before and after in the intent space. state The larger the value, the stronger the continuity of the state.

[0032] C state = (Cos θ + 1) / 2 The formula for the comprehensive scoring of dimensionless semantic continuity is defined as follows, where S final This is a dimensionless comprehensive score used to characterize the overall semantic coherence of the segments before and after the switching node; I ′info The normalized mutual information score corresponds to the semantic domain association contribution item; G ref C represents the dimensionless referential matching gain coefficient, corresponding to the referential continuation contribution term; state The score represents the continuity of the state, corresponding to the contribution of the candidate intention state; W area W ref and W state All are dimensionless weighting coefficients, used to adjust the contribution ratios of semantic domain association, referential relationship continuation, and state vector continuity in the comprehensive score.

[0033] S final = I ′info × W area + G ref × W ref + C state × W state W area = 0.4, W ref = 0.3, W state = 0.3 W area + W ref + W state = 1 Among them, I ′info This can be achieved by using the mutual information entropy value I info Relative to the preset mutual information upper limit I maxObtained by performing normalization, when I info Greater than or equal to I max Time take I ′info =1, when I info When less than 0, take I. ′info =0, when 0≤I info max According to I ′info =I info / I max Calculation; in this embodiment, preset I max Take 1 bit, therefore from I info =0.08 bits to get I ′info =0.08. The G... ref The value can be directly assigned based on the matching result of the referential relationship or assigned according to the matching confidence level. This embodiment adopts a binarization processing method, that is, 1 is assigned when a valid referent is detected, and 0 is assigned when no valid referent is detected. The C state Obtained by cosine similarity mapping (Cosθ), where Cosθ theoretically takes values ​​in the range [-1, 1]. The mapped C state The value is located in [0,1], thus avoiding negative values ​​from directly participating in the comprehensive score and causing scale imbalance.

[0034] In this embodiment, C is obtained from Cosθ=0.3328. state =0.6664, and the overall score is calculated as follows.

[0035] I ′info = 0.08 C state = (0.3328 + 1) / 2 = 0.6664 S final = 0.08 × 0.4 + 0 × 0.3 + 0.6664 × 0.3 = 0.23192 0.23192 < 0.55 A dimensionless threshold for retaining consistency is set to 0.55. This threshold is used to distinguish between consistent segments that should be retained and weakly correlated segments that should be removed; when S final When S is greater than or equal to 0.55, the combination of preceding and following segments is considered to have continuity; when S final If the value is less than 0.55, the combination of preceding and following segments is determined not to meet the retention condition. A size comparison is performed. Since 0.23192 < 0.55, it falls within the weak correlation range, and the result is true. Therefore, the weakly correlated combination is removed from the current branch.

[0036] ​However, in another concurrent example, if Seg_2 receives the new text "zoom in on this map again", I is calculated using the same process. info =0.92 bit, G ref =1.0, Cosθ=0.95; following the same normalization and mapping rules, we obtain I ′info =0.92, C state =0.975; at this time I ′info A higher value indicates a high degree of correlation between the preceding and following semantic domains. ref =1.0 indicates the existence of a clear referential continuation relationship, C state =0.975 indicates that the candidate intent states before and after are highly consistent.

[0037] I ′info = 0.92 C state = (0.95 + 1) / 2 = 0.975 S final = 0.92 × 0.4 + 1.0 × 0.3 + 0.975 × 0.3 = 0.9605 0.9605 > 0.55 Since 0.9605 > 0.55, it is within the consistency preservation range. The fragment combination memory preservation operation is performed, and the verified fragment identifiers are stored in the semantic contiguous buffer to obtain the semantic continuation fragment combination.

[0038] S43: Invoke the semantic continuation fragment combination and anchor fragment index, read the switching node marker associated with the device source and session round identifier, arrange the fragment combination according to the order of device source and session round identifier, filter adjacent fragments and concatenate them, set continuous session fragment markers on the concatenation result, and generate a cross-scenario continuous session context sequence; as detailed below: The semantic continuation fragment combination and anchor fragment index are invoked, the device source tag associated with the switch node marker is read, and the hardware device tag of the first fragment is extracted as "vehicle dashboard" and its corresponding session round identifier. The hardware device tag extracted from the second segment is "central control touchscreen" and its corresponding session round identifier. Retrieve the dimensionless switching weight value set in S33 Perform an ascending sort operation based on the round identifiers, and confirm. and It has logical progression and extraction End position offset in milliseconds ms, extract Starting position offset in milliseconds ms, using linked list operators to perform fragment splicing operations on a unified timeline, will The data stream's starting address pointer is mounted to At the tail node of the linked list, for the concatenated continuous data blocks, the affine transformation matrix hardware acceleration card is called to perform heterogeneous device interaction normalization calculation. The underlying heterogeneous device interaction normalization calculation uses a preset spatial coordinate affine transformation matrix loaded by the hardware acceleration card to achieve coordinate equivalence mapping between displays with different physical resolutions, thus converting the original touch pixel coordinates of the dashboard... Mapped to the unified display coordinate system of the central control screen If the mapping exceeds the physical resolution boundary of the screen, an out-of-bounds exception interrupt is triggered and the screen is automatically reset to the center. A smooth transition of the intended hotspot intensity is then performed, and the dimensionless intensity value of the transition at the junction is calculated using a weighted average method. Assign and set globally unique consecutive session fragment string markers for the concatenated session stream. Write this flag into the identifier field of the cross-scene context data header and associate it with the corresponding start and end total duration. The system merges and stores the speech, touch, and gaze features of each segment according to a unified time slice index number, establishing a complete semantic link across scenarios and generating a continuous conversation context sequence across scenarios.

[0039] S5: Call the continuous session fragment marker in the cross-scenario continuous session context sequence, obtain the privacy budget consumption record sequence table consisting of continuous session fragment marker, device identifier, consumption amount and consumption time, and the edge node session identifier mapping table consisting of edge node identifier, session identifier and device source, match the session identifier, edge node identifier and device source according to the continuous session fragment marker, and associate the corresponding privacy budget consumption record, mark the privacy budget consumption status of continuous session fragments in different device sources and session rounds, classify the occurrence record, session round record and privacy budget consumption status of the same continuous session fragment with the device source in chronological order, and group and arrange them according to the session round identifier carried by the continuous session fragment marker to form a privacy budget feedback constraint relationship and generate a session context management view; The session context management view includes continuous session segment records, device identification information, privacy budget consumption status, and session round records.

[0040] Please see Figure 6 Specifically, in S5, this includes: S51: Invoke the continuous session fragment markers in the cross-scenario continuous session context sequence, obtain the privacy budget consumption record sequence table consisting of continuous session fragment markers, device identifiers, consumption limits, and consumption time, and the edge node session identifier mapping table consisting of edge node identifiers, session identifiers, and device sources. Compare the session identifiers and device sources according to the continuous session fragment markers to generate a budget mapping list; as detailed below: Retrieve continuous session fragment markers from the cross-scenario continuous session context sequence and extract the continuous session fragment markers generated in S43. The corresponding string index value is used to retrieve the privacy budget consumption record sequence table managed by the trusted execution environment's encryption chip from the underlying distributed security database, and the fragment of the first row of data in that table is marked as... The physical identifier of the equipment is a vehicle-mounted instrument panel. Privacy consumption limit is Unit privacy score, consumption time is ms, the segment of the second row of data is extracted and marked as The physical identifier of the device is a central control touchscreen. Privacy consumption limit is Unit privacy score, consumption time is ms, the segment of data extracted from the third row is marked as The physical identifier of the device is a central control touchscreen. Privacy consumption limit is Unit privacy score, consumption time is ms, synchronously obtain the edge node session identifier mapping table, and extract the hardware edge gateway node identifier from it. The associated session identifier is The corresponding original device source is the dashboard; the hardware edge gateway node identifier is extracted from another record. The associated session identifier is The original device source is the central control screen, which performs continuous session fragment marking. With session identifier The string hash value comparison operation, when two identifier strings are processed by the SHA-256 hash algorithm, the matching degree is 1. When the two belong to the same semantic lifecycle, if they do not match, isolation is triggered as an illegal access event. The specific isolation method is to mark the fragment as an abnormal pointer in memory and prohibit rendering to the front-end video memory, and call the device origin attribute field "dashboard" and the device identifier in the privacy record. Perform an equivalence check; if the names of the two names match by a certain degree... Then the current consumption amount will be... Mount to Under the corresponding logical node storage area, for the central control screen Perform the same name matching operation, and and Credit limit linked to In the logical node storage area, the consumption values, timestamps and node identifiers that have been verified and matched for identification are structurally encapsulated in a linear chronological order to construct an associated data array containing multiple device privacy overhead dimensions, and generate a budget mapping list.

[0041] S52: Based on the budget mapping list, read the continuous session fragment markers, session identifiers, edge node identifiers, device sources, consumption limits, and consumption times. Align the session identifiers with the edge node identifiers item by item, filter records with consistent device sources, associate them with the privacy budget consumption record sequence list, and mark the privacy budget consumption status in different device sources and session rounds to obtain the fragment consumption markers; specifically as follows: Based on the budget mapping list, read the consecutive session fragment tags. Session identifier Edge node identifier and The sources of equipment are the "dashboard" and "central control screen," as well as the consumption amount. , and Privacy score and time consumption record, execution session identifier The addressing alignment operation with the edge node identifier will exist Interaction logs at ms and Perform address pointer binding, exist ms and Interaction logs at ms and Perform address pointer binding, execute device source consistency filtering, extract all record rows belonging to "vehicle dashboard" from the budget mapping list, and retrieve the dimensionless historical average consumption baseline value from the privacy budget consumption record sequence table. Privacy score, which refers to the amount of money spent on the dashboard. Perform subtraction to obtain the dimensionless deviation value Privacy score refers to the amount of data consumed on the central control screen. Perform a subtraction operation to obtain the dimensionless deviation value: Privacy score: This deviation value is used as a reference for subsequent dynamic privacy compensation to adjust the noise injection gain of the hardware filter. It sets the opening and closing range for determining privacy consumption levels, based on the consumption limit itself. If the consumption limit is within... Within the closed interval, it is marked as a "low-consumption" state; if it is in... The range is then marked as "medium consumption". If it is greater than... It is then marked as a "high consumption" state and the differential privacy noise ambiguity mechanism is automatically activated, based on which the dashboard is determined in the first round. The consumption status is "low consumption", and the status flag is the code. The central control screen was judged in the second round. The consumption status is "medium consumption", and the status is marked by the code. Arrange these status bits according to the timeline The switching node of ms marks the forward and backward classification pointers, and sets the status mark through logical connection operations. Associated with The fragment will Associated with The fragments are processed, and corresponding consumption status attribute fields are set for the start and end offset positions of each consecutive fragment in the storage space to obtain the fragment consumption marker.

[0042] S53: Invoke the fragment consumption flag and budget mapping list, read the occurrence records of the same continuous session fragment, session round records, privacy budget consumption status, device source, and session round identifier, arrange the occurrence records in order of consumption time, classify the consumption status according to the device source, group and arrange the budget feedback constraint relationship according to the session round identifier, and generate a session context management view; as detailed below: Invoke fragment consumption flags and budget mapping list to read the same consecutive session fragments. Extract all internal records from the dashboard. MS, central control screen ms and Extracting privacy consumption events from session round records in MS. Corresponding dashboard Based on the hardware affiliation of the central control screen, read the privacy budget consumption status at each point in time. and Perform the timestamp ascending sort operation, sorting the records according to... The axial position of the milliseconds is physically rearranged, and the aforementioned records are divided into "dashboard data group" and "central control screen data group" based on the device source field, and the session round identifier is retrieved. and Enter the data from the "Dashboard Data Group" In the grouping structure, fill in the data for the "Central Control Screen Data Group". Within the grouping structure, the privacy budget feedback constraint formula rendered by the hardware graphics accelerator is invoked. Calculate the overall view metrics, where The privacy budget consumption density score representing the session context management view, with dimensions of: , The total number of interaction events within a consecutive session segment is a dimensionless integer value. and subscript Indicates the first time sequence after time reordering This privacy breach incident The dimensionless privacy consumption limit for each event (respectively) Privacy points Privacy points Privacy points). The dimensionless equipment source safety risk weighting coefficient (from the dashboard) Central control screen ), End timestamp ms, Start timestamp ms, by substituting the data and performing multiplication, addition, and division operations, we obtain... The calculated density score is compared with the preset dimensionless threshold for compliance upper limit. The system performs a size comparison. If the score is lower than the upper limit, the graphics rendering layer is displayed as a green safety constraint icon on the user's screen management interface. If the score is higher than the upper limit, a red interception alarm is popped up and the current communication link is interrupted. Finally, all round groupings, device classification information, fragment status markers, and calculated density scores are summarized into the front-end video memory visualization buffer to generate a session context management view.

[0043] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A cross-scenario AI session context management method based on intent hotspot vectors, characterized in that, Includes the following steps: S1: Collect touch coordinates, speech recognition text, gaze coordinates and key timestamp sequences, normalize them to a unified timeline, divide them into time slices, and progressively update candidate intent states according to the boundaries of speech, touch, gaze and key, match preset intent labels and hotspot intensity, and generate intent hotspot vector time series; S2: Based on the progressive relationship of candidate intent states in the time series of intent hotspot vectors, compare the label direction, hotspot intensity and continuity of adjacent time slices, identify the reverse or interruption position as the turning point of change, merge the time slices without turning points, divide the unified time axis, and generate a list of continuous segments of conversation semantics. S3: Call the list of continuous semantic segments of the session, obtain the session context cache index table and the user session round time series, check the direction of the transition according to the turning point of the change, filter the break position of the direction change of the cache position change or the change of the session round as the switching node, and generate a cross-scene session switching time node set. S4: Based on the cross-scene session switching time node set, use the switching node marker as the anchor point to associate the semantic segments of the preceding and following sessions, trace back the speech recognition text and candidate intent state, compare the semantic and state continuity of the beginning and end, retain the segments that can be continued, and splice them according to the device source and round order to generate a cross-scene continuous session context sequence.

2. The cross-scene AI session context management method based on intent hotspot vectors according to claim 1, characterized in that: The intent hotspot vector time series includes preset intent labels, hotspot intensity values, time slice identifiers, and candidate intent states. The conversation semantic continuous segment list includes segment number, segment start time, segment end time, and semantic fluctuation identifier. The cross-scenario conversation switching time node set includes switching node markers, node time positions, cache position identifiers, and conversation round identifiers. The cross-scenario continuous conversation context sequence includes continuous conversation segment markers, device source identifiers, voice and text segments, and candidate intent state records.

3. The cross-scene AI session context management method based on intent hotspot vectors according to claim 1, characterized in that: The steps for obtaining S1 are as follows: S11: Collect the touch coordinate sequence, speech recognition text sequence, gaze tracking coordinate sequence and key timestamp sequence generated during terminal interaction, organize the touch coordinate time, speech text time, gaze coordinate time and key trigger time, map each time to a unified time axis according to the collection order, divide the unified time axis into continuous time slices, and generate a time slice mapping list. S12: Based on the time slice mapping list, read the voice keywords, keyword occurrence time and touch displacement trajectory in the current time slice, determine the candidate intention direction according to the order of keyword occurrence, compare the touch displacement trajectory with the candidate intention direction, retain the candidate intention within the range of consistent connection, and obtain the candidate intention convergence interval. S13: Call the candidate intent convergence interval, associate the gaze dwell area, duration and key trigger time, confirm the candidate intent attention object according to the gaze dwell area, mark the interaction completion boundary according to the key trigger time, inherit the candidate intent state of adjacent time slices in time sequence and match the preset intent label and hotspot intensity to generate an intent hotspot vector time series.

4. The cross-scene AI session context management method based on intent hotspot vectors according to claim 1, characterized in that: The steps for obtaining S2 are as follows: S21: Call the candidate intent state progression relationship in the intent hotspot vector time series, read the previous time slice, the next time slice, the preset intent label and the hotspot intensity along the unified time axis, compare the preset intent labels corresponding to adjacent time slices, record the change direction according to the label succession relationship, and generate the label change trajectory according to the increase or decrease of the hotspot intensity status mark change amplitude. S22: Based on the label change trajectory, call the candidate intent state progression relationship, the candidate intent state of the previous time slice and the candidate intent state of the next time slice, compare the state continuation direction with the label change direction, and when the label direction is reversed or the continuation path is interrupted, record the corresponding time position and associate it with the adjacent time slices to obtain the semantic fluctuation interval. S23: Invoke the semantic fluctuation range and the label change trajectory, read the continuous time slices not marked as fluctuation boundaries, the succession order of candidate intention states and the change inflection points, determine the succession relationship of candidate intention states in adjacent time slices, merge the continuous time slices with consistent succession into continuous semantic segments of the conversation, divide the unified time axis according to the change inflection points, and generate a list of continuous semantic segments of the conversation.

5. The cross-scene AI session context management method based on intent hotspot vectors according to claim 1, characterized in that: The steps for obtaining S3 are as follows: S31: Call the continuous segment of the session semantics in the list of continuous segment of the session semantics, obtain the session context cache index table consisting of session identifier, context segment identifier, cache location and cache time, and user session round time sequence consisting of session round number, round start and end time and device source, arrange adjacent segments along a unified time axis according to the turning point of change, and generate a session reference list; S32: Based on the session reference list, call the session context cache index table, user session round time sequence, semantic continuous segments of adjacent sessions and semantic fluctuation interval, check the deviation of the candidate intent state continuation direction in the semantic fluctuation interval, determine whether the deviation directions before and after the change inflection point are opposite, and obtain the direction change break position. S33: Invoke the direction change break position and the session reference list, read the cache position and cache time in the session context cache index table, read the session round number, round start and end time and device source in the user session round time series, filter the time position corresponding to the cache position change or session round change, set the switching node mark, and generate a cross-scene session switching time node set.

6. The cross-scene AI session context management method based on intent hotspot vectors according to claim 5, characterized in that: The process of checking the deviation of the candidate intent state's direction within the semantic fluctuation range is as follows: Read the candidate intent state succession direction corresponding to the consecutive semantic segments of the preceding and following sessions, and align the candidate intent state succession direction with time according to the unified time axis. Compare the candidate intent state succession directions of adjacent time slices within the semantic fluctuation interval item by item. When the number of consecutive time slices with inconsistent directions reaches a preset deviation threshold, mark the corresponding time range as a deviation interval. The process of determining whether the deviation directions before and after the turning point of change are opposite is as follows: Read the deviation interval before the change inflection point and the deviation interval after the change inflection point respectively. Take the candidate intention state acceptance direction corresponding to the deviation interval before the change inflection point as the forward deviation direction and the candidate intention state acceptance direction corresponding to the deviation interval after the change inflection point as the backward deviation direction. When the forward deviation direction is opposite to the backward deviation direction, record the time position corresponding to the change inflection point as the direction change break position.

7. The cross-scene AI session context management method based on intent hotspot vectors according to claim 5, characterized in that: The process of filtering and caching position changes or session round changes corresponding to time positions is as follows: Read the time position corresponding to the direction change break position, the cache position, and the cache time, and read the session round number, the round start and end time, and the device source. When the time position corresponding to the direction change break position is consistent with the cache time when the cache position changes, or is consistent with the round boundary time in the round start and end time, retain the time position corresponding to the direction change break position and set the switching node marker.

8. The cross-scene AI session context management method based on intent hotspot vectors according to claim 1, characterized in that: The steps for obtaining S4 are as follows: S41: Call the switching node marker in the cross-scene session switching time node set, use the position of the switching node marker in the unified time axis as the anchor point, associate the continuous semantic segments of the session before and after the switching node, trace back the speech recognition text and candidate intent state corresponding to each continuous semantic segment of the session, compare the order of the segments according to the start and end times of the segments, and generate the anchor point segment index. S42: Based on the anchor segment index, read the ending semantics of the semantic continuous segment of the previous session, the starting semantics of the semantic continuous segment of the next session, and the corresponding candidate intent states of the two. According to the time sequence, determine the continuation relationship of voice keywords, the corresponding situation of referential relationship, and the continuity of candidate intent states. Retain the consistent segment combination to obtain the semantic continuation segment combination. S43: Invoke the semantic continuation fragment combination and the anchor fragment index, read the switching node mark associated with the device source and the session round identifier, arrange the fragment combination according to the order of device source and session round identifier, filter adjacent fragments and splice them, set continuous session fragment mark on the splicing result, and generate a cross-scene continuous session context sequence.

9. The cross-scene AI session context management method based on intent hotspot vectors according to claim 1, characterized in that: The method further includes: S5: Invoke the cross-scenario continuous session context sequence, match the session identifier, edge node identifier and device source according to the continuous session fragment marker, associate privacy budget consumption records, classify them according to time order, device source and session round, form privacy budget feedback constraint relationship, and generate a session context management view; The session context management view includes continuous session segment records, device identification information, privacy budget consumption status, and session round records.

10. The cross-scene AI session context management method based on intent hotspot vectors according to claim 9, characterized in that: The steps for obtaining S5 are as follows: S51: Call the continuous session fragment marker in the cross-scenario continuous session context sequence, obtain the privacy budget consumption record sequence table consisting of continuous session fragment marker, device identifier, consumption amount and consumption time, and the edge node session identifier mapping table consisting of edge node identifier, session identifier and device source, and generate a budget mapping list by comparing the session identifier and device source according to the continuous session fragment marker. S52: Based on the budget mapping list, read the continuous session fragment marker, session identifier, edge node identifier, device source, consumption amount and consumption time, align the session identifier with the edge node identifier item by item, filter records with consistent device source, associate with the privacy budget consumption record sequence list, and mark the privacy budget consumption status in different device sources and session rounds to obtain the fragment consumption marker; S53: Invoke the fragment consumption flag and the budget mapping list, read the occurrence record of the same continuous session fragment, session round record, privacy budget consumption status, device source and session round identifier, arrange the occurrence record in the order of consumption time, classify the consumption status according to the device source, group and arrange the budget feedback constraint relationship according to the session round identifier, and generate a session context management view.

Images