Voice interaction system for elevator emergency evacuation robot

Abstract

This invention provides a voice interaction system for an elevator emergency evacuation robot, relating to the field of voice interaction technology. The system collects segmented continuous voice signals from trapped individuals during an elevator emergency, generating corresponding text information and recognition confidence levels. Based on the text information, it extracts semantic elements such as urgency, physical condition, and intent to seek help, constructing semantic state data that evolves over time. Within a preset time window, it performs cumulative analysis of the semantic state data and controls its influence based on the recognition confidence level, generating semantic evolution results that reflect the development trend of emergency needs. Based on the semantic evolution results, it determines the type of emergency need and controls the elevator emergency evacuation robot to perform corresponding voice interaction and emergency response operations.

Description

Technical Field

[0001] This invention relates to the field of voice interaction technology, and in particular to the voice interaction system of an elevator emergency evacuation robot. Background Technology

[0002] Existing elevator emergency evacuation robots mostly employ interactive systems based on speech recognition and semantic understanding to reassure, guide, and collect information from trapped individuals. These speech interaction systems typically include a speech acquisition module, a speech recognition module, and a pre-set command matching module. The speech recognition module generally uses acoustic and language models to identify the acquired speech signals and matches the recognition results with a pre-set emergency command library, thereby triggering corresponding broadcast or control logic. For example, after an elevator malfunctions, the robot continuously broadcasts prompts to guide trapped individuals to utter fixed phrases such as "need help" or "someone is injured." After recognizing the corresponding keywords, the system uploads status information to the backend monitoring system or executes the corresponding evacuation assistance procedures.

[0003] However, in real-world elevator emergency evacuation applications, the aforementioned existing technologies are susceptible to interaction failures due to environmental and human conditions. Specifically, when an elevator stops due to a malfunction, the interior of the car is often accompanied by complex background noise, including alarms, motor noises, and multiple people talking simultaneously. Trapped individuals, in a state of tension, panic, or physical discomfort, may speak at excessively low volumes, with abnormal speaking speeds, or incomplete expressions. In such situations, existing voice interaction systems, which primarily rely on matching preset keywords or standard phrases, are prone to misidentifying "incomplete instructions" or "non-standard expressions" as invalid input, resulting in the repeated playback of default prompts and an inability to promptly understand the trapped individuals' true needs. For example, when a trapped individual repeatedly shouts "I...can't breathe..." or "My child is dying," the system fails to recognize the complete preset instruction phrase and thus does not trigger the medical assistance process, delaying the emergency response and reducing the practicality and reliability of elevator emergency evacuation robots in real-world accident scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a voice interaction system for an elevator emergency evacuation robot, which aims to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: The voice interaction system for the elevator emergency evacuation robot includes: The voice signal segmentation acquisition module is used to acquire continuous voice signals of people trapped in the elevator car in an emergency situation, and to segment the continuous voice signals according to a preset time interval to generate multiple voice segments. The speech-text and confidence generation module is used to perform speech recognition processing on each speech segment, convert the speech segment into corresponding text information, and generate a corresponding recognition confidence score for each piece of text information during the conversion process. The emergency semantic state construction module is used to extract semantic elements representing the degree of urgency, physical state, and intention to seek help from text information, and combine the semantic elements with the corresponding recognition confidence to generate semantic state data that represents the emergency state of trapped personnel and evolves over time. The time window semantic evolution analysis module is used to accumulate multiple semantic state data in chronological order within a preset time window, and to differentiate the influence of different semantic state data on the accumulation process based on the recognition confidence. Based on the changes in semantic state data during the accumulation process, it generates semantic evolution results that reflect the development trend of emergency needs. An emergency demand determination triggering module is used to compare the semantic evolution result with preset emergency demand determination conditions. The emergency demand determination conditions include triggering conditions corresponding to at least one determination stage. When the emergency demand determination conditions are met, the corresponding emergency demand type is determined. The emergency interaction execution module is used to control the elevator emergency evacuation robot to perform corresponding voice feedback, emergency information reporting, or evacuation guidance operations according to the type of emergency need.

[0006] Preferably, the emergency semantic state construction module includes: The semantic fragment localization and generation unit is used to locate and process the content related to urgency, physical condition and intention to seek help in the text information, and generate corresponding semantic fragments and time markers that correspond one-to-one with the semantic fragments. The semantic category mapping generation unit is used to perform category mapping processing on the semantic fragments based on their content, and generate corresponding semantic category identifiers; The confidence level admission determination unit is used to compare the recognition confidence level corresponding to the semantic segment with the preset confidence level admission conditions to generate an admission determination result that represents whether the semantic segment is allowed to participate in subsequent processing. The semantic state initial generation unit is used to determine the corresponding initial state type based on the semantic category identifier when the admission judgment result characterization allows participation in subsequent processing, and to assign an initial influence degree to the initial state type in combination with the recognition confidence of the semantic segment, thereby generating the initial semantic state. The semantic state temporal update unit is used to update the initial semantic state over time according to the time identifier, and generate an updated semantic state that reflects the current influence of the semantic state. The semantic state solidification output unit is used to set the updated semantic state to an irreversible state and use it as semantic state data when the updated semantic state meets the preset solidification conditions. The semantic state data includes the corresponding recognition confidence level.

[0007] Preferably, the time window semantic evolution analysis module includes: The analysis time window generation unit is used to generate at least one time window for semantic evolution analysis based on the current time and a preset time span. The semantic state window mapping unit is used to map semantic state data to a time window according to its corresponding time identifier, and generate a set of semantic states within the window. The semantic state order determination unit is used to sort the semantic state set within the window in a temporal order according to the time identifier corresponding to the semantic state data, and generate an ordered semantic state sequence. The semantic state influence adjustment unit is used to adjust the influence of each semantic state in the sequence accumulation process according to the recognition confidence of each semantic state in the ordered semantic state sequence, so as to generate a controlled accumulation sequence. The semantic state change trajectory generation unit is used to record the state trajectory of the semantic state corresponding to the same semantic category changing over time in a controlled cumulative sequence. The semantic evolution result locking unit is used to generate and lock the semantic evolution result for emergency demand determination based on the state trajectory.

[0008] Preferably, the emergency demand determination triggering module includes: The demand determination candidate generation unit is used to parse and process the semantic evolution results, extract the determination elements that reflect changes in urgency, physical state, and intention to seek help, and match the determination elements with the preset emergency demand determination dimensions. Based on the corresponding results, demand candidate information for emergency demand determination is generated. The demand candidate threshold comparison unit is used to compare the demand candidate information with the preset multi-level emergency demand judgment thresholds and generate the corresponding threshold matching results. The pre-triggering state generation unit is used to generate a pre-triggering state identifier corresponding to the emergency demand type when the threshold matching result meets the pre-triggering condition. The confirmation condition continuous verification unit is used to continuously compare and verify the semantic evolution results obtained after the pre-trigger state identifier is generated, and generate confirmation verification results. The emergency demand type determination unit is used to determine the corresponding emergency demand type based on the pre-trigger status identifier when the verification result meets the confirmation conditions, and set the emergency demand type as a non-reversible determination result.

[0009] Preferably, the semantic state time-series update unit includes: The update time window construction unit is used to construct a discrete update time window for semantic state update based on the time identifier corresponding to the semantic fragment, and generate the corresponding update time window identifier. The update trigger determination unit is used to determine whether there is a new semantic fragment entering the update time window based on the update time window identifier, and to generate an update trigger determination result; The semantic category conflict determination unit is used to compare the semantic category corresponding to the newly added semantic fragment with the semantic category corresponding to the existing semantic state when the update triggers the determination result representation to perform update processing, and generate a semantic category conflict determination result. The update method selection unit is used to compare the recognition confidence of the new semantic segment with the recognition confidence of the existing semantic state when the semantic category conflict determination result indicates that the new semantic segment and the existing semantic state belong to the same semantic category. If the comparison result meets the preset coverage condition, it determines to perform a coverage update on the existing semantic state. If the comparison result does not meet the preset coverage condition, it determines to perform a suppression update on the existing semantic state. The update state generation unit is used to update the influence of the existing semantic state according to the method of overwrite update or suppress update, and generate an updated semantic state as input for subsequent processing.

[0010] Preferably, the semantic state influence adjustment unit includes: The confidence segmentation unit is used to divide the ordered semantic state sequence into multiple confidence influence segments based on the recognition confidence contained in each semantic state in the ordered semantic state sequence, and generate segment identifiers. The segment participation eligibility determination unit is used to determine whether the corresponding semantic state is eligible to participate in the cumulative processing based on the preset participation rules according to each confidence level of the affected segment, and to generate the participation eligibility determination result. The cross-segment impact blocking determination unit is used to generate a blocking determination result to block the impact of a low-confidence impact segment on the semantic state of a high-confidence impact segment when the semantic state of the low-confidence impact segment does not meet the participation qualification determination result. The accumulation method determination unit is used to determine the participation method of each semantic state in the accumulation process based on the participation qualification determination result and the blocking determination result; The controlled accumulation result generation unit is used to perform accumulation processing on semantic states that have passed the participation qualification determination and have not been blocked, according to the participation method, to generate a controlled accumulation sequence as input for subsequent trajectory analysis.

[0011] Preferably, the semantic state change trajectory generation unit includes: The category state sequence extraction unit is used to extract semantic states belonging to the same semantic category based on the semantic categories corresponding to the semantic states in the controlled cumulative sequence, and generate a semantic category state sequence. The state change inflection point recognition unit is used to compare the degree of influence of adjacent semantic states in the semantic category state sequence, identify the position where the change direction changes or the change amplitude reaches the preset amplitude change threshold, and generate trajectory inflection point identifiers. The trajectory segmentation generation unit is used to divide the semantic category state sequence into multiple state change segments with the trajectory inflection point marker as the boundary, and generate a segmented trajectory sequence. The segmented change feature generation unit is used to extract segmented change features that characterize the semantic state change trend within each state change segment based on each state change segment. The trajectory result combination unit is used to combine segmented change features in chronological order to generate a semantic state change trajectory that serves as the basis for generating semantic evolution results.

[0012] Preferably, the cross-segment impact closure determination unit includes: The segment state acquisition unit is used to acquire the participation qualification determination results corresponding to each semantic state in the low-confidence influence segment and the high-confidence influence segment before performing semantic state accumulation processing; The cross-segment influence relationship determination unit is used to determine whether the semantic state in the low-confidence influence segment constitutes a cross-segment influence on the semantic state in the high-confidence influence segment when the semantic state in the low-confidence influence segment does not meet the participation qualification determination result. The blocking determination generation unit is used to generate a blocking determination result when the cross-segment influence relationship determination is established, which is used to block the influence of the semantic state of the low-confidence influence segment on the semantic state of the high-confidence influence segment.

[0013] Preferably, the accumulation mode determination unit includes: The cumulative judgment result collection unit is used to collect the participation qualification judgment results and cross-segment impact blocking judgment results corresponding to each semantic state before performing semantic state accumulation processing, and generate the cumulative judgment input results for accumulation method judgment. The cumulative participation mode determination unit is used to perform the following determination processes on each semantic state in sequence based on the cumulative determination input result: When the corresponding semantic state has cross-segment impact on the blocking determination result, the cumulative participation method of the semantic state is determined to be prohibited from participating in the cumulative processing; When the corresponding semantic state does not have a cross-segment impact on the blocking determination result and the participation qualification determination result indicates that participation is allowed, the cumulative participation method of the semantic state is determined to be the cumulative processing of allowed participation. When the corresponding semantic state does not affect the blocking determination result across segments and the participation qualification determination result indicates that participation is not allowed, the cumulative participation method of the semantic state is determined to be restricted participation cumulative processing. The restricted participation cumulative processing includes restricting the semantic state to participate in the cumulative processing without affecting other semantic states.

[0014] Preferably, the state change inflection point identification unit includes: The change comparison object construction unit is used to construct adjacent state comparison objects for change analysis in a semantic category state sequence for adjacent semantic states. The adjacent state comparison objects contain the degree of influence of adjacent semantic states and the corresponding time identifier. The change direction determination unit is used to determine the change direction between adjacent semantic states based on the influence degree change relationship between adjacent semantic states in the adjacent state comparison object, and generate the change direction determination result. The change magnitude determination unit is used to determine whether the change amount meets the preset change magnitude threshold based on the change in the influence degree of adjacent semantic states in the adjacent state comparison object, and to generate the change magnitude determination result. The inflection point determination and generation unit is used to determine the corresponding position as the state change inflection point when the change direction determination result represents a change in the direction of change of the adjacent semantic state, or when the change magnitude determination result represents a change in the amount of change of the adjacent semantic state that meets the preset magnitude change threshold, and to generate a trajectory inflection point identifier corresponding to the state change inflection point.

[0015] The above-described solution of the present invention has at least the following beneficial effects: First, by collecting segmented continuous voice signals from trapped individuals during elevator emergencies and simultaneously generating recognition confidence during the voice recognition process, the system no longer relies on the recognition results of single, complete sentences. Instead, it can acquire continuous voice information with time attributes and confidence indicators, providing a more stable input basis for subsequent emergency judgments. This reduces the impact of complex noise environments and non-standard expressions on the effectiveness of voice interaction.

[0016] Based on this, by extracting semantic elements such as urgency, physical condition, and intention to seek help from text information, and combining them with recognition confidence to construct semantic state data that evolves over time, the scattered, repetitive, or incomplete expressions of trapped personnel can be continuously accumulated and tracked in the form of semantic states. This avoids the problem in existing technologies where input is directly judged as invalid due to the failure to match preset keywords, thereby improving the system's ability to acquire real emergency semantics.

[0017] Furthermore, by accumulating semantic state data in chronological order within a preset time window, and by differentially controlling the influence of different semantic state data on the accumulating process based on the recognition confidence level, high-confidence semantics dominate the comprehensive judgment, while low-confidence semantics do not unduly interfere with the overall judgment. This allows the generated semantic evolution results to more accurately reflect the development trend of the emergency needs of trapped personnel.

[0018] Finally, by comparing the semantic evolution results with the emergency demand determination conditions that include multiple determination stage trigger conditions, and determining the corresponding emergency demand type when the conditions are met, the triggering of emergency demands no longer depends on a single voice or fixed command, but is based on the semantic evolution process for staged confirmation. This enables the timely triggering of the corresponding emergency response process when trapped personnel exhibit persistent physical abnormalities or a clear intention to seek help, thereby improving the response reliability and practicality of the elevator emergency evacuation robot in real accident scenarios. Attached Figure Description

[0019] Figure 1 This is an architecture diagram of the voice interaction system of the elevator emergency evacuation robot provided in an embodiment of the present invention. Detailed Implementation

[0020] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0021] like Figure 1 As shown, an embodiment of the present invention proposes a voice interaction system for an elevator emergency evacuation robot, the system comprising: The voice signal segmentation acquisition module is used to acquire continuous voice signals of people trapped in the elevator car in an emergency situation, and to segment the continuous voice signals according to a preset time interval to generate multiple voice segments. The speech-text and confidence generation module is used to perform speech recognition processing on each speech segment, convert the speech segment into corresponding text information, and generate a corresponding recognition confidence score for each piece of text information during the conversion process. The emergency semantic state construction module is used to extract semantic elements representing the degree of urgency, physical state, and intention to seek help from text information, and combine the semantic elements with the corresponding recognition confidence to generate semantic state data that represents the emergency state of trapped personnel and evolves over time. The time window semantic evolution analysis module is used to accumulate multiple semantic state data in chronological order within a preset time window, and to differentiate the influence of different semantic state data on the accumulation process based on the recognition confidence. Based on the changes in semantic state data during the accumulation process, it generates semantic evolution results that reflect the development trend of emergency needs. An emergency demand determination triggering module is used to compare the semantic evolution result with preset emergency demand determination conditions. The emergency demand determination conditions include triggering conditions corresponding to at least one determination stage. When the emergency demand determination conditions are met, the corresponding emergency demand type is determined. The emergency interaction execution module is used to control the elevator emergency evacuation robot to perform corresponding voice feedback, emergency information reporting, or evacuation guidance operations according to the type of emergency need.

[0022] In this embodiment of the invention, by segmenting the continuous voice signals of trapped personnel in an elevator emergency, and converting each voice segment into text information while generating corresponding recognition confidence, subsequent processing can obtain continuous voice expression content with confidence indicators in chronological order, thereby providing traceable input basis for emergency semantic judgment.

[0023] By extracting semantic elements representing urgency, physical condition, and intention to seek help from text information, and combining these semantic elements with corresponding recognition confidence levels to generate semantic state data that evolves over time, the requests for help from trapped individuals at different points in time can be continuously recorded in a unified data format. The continuity of semantic state data in the time dimension supports a comprehensive analysis of the emergency status of the same trapped individual.

[0024] By accumulating multiple semantic state data in chronological order within a preset time window, and using recognition confidence to differentiate the degree of influence of different semantic state data in the accumulation process, the influence paths of information with different sources of reliability on the overall judgment can be distinguished during the accumulation process. This avoids undue interference from low-confidence expressions on the comprehensive judgment within the time window, and ultimately outputs semantic evolution results that reflect the development trend of emergency needs.

[0025] By comparing the semantic evolution results with emergency demand determination conditions that include at least one determination stage trigger condition, the corresponding emergency demand type is determined when the trigger condition is met. This makes the determination process of emergency demand type have stage constraints, and the output emergency demand type can be consistent with the semantic evolution trend of trapped personnel within the time window. It can also serve as the control basis for subsequent voice feedback, emergency information reporting, or evacuation guidance operations.

[0026] For example, in a scenario where the elevator stops and people inside the car are continuously calling for help, the system segments the trapped person's continuous voice into text at time intervals and generates a recognition confidence score for each text segment. During the semantic state construction process, the system extracts semantic elements such as "difficulty breathing," "dizziness," "cannot open the door," and "needs rescue" from the text and forms semantic state data that evolves over time. Within a preset time window, the system accumulates the semantic state data in chronological order and adjusts the impact of each semantic state segment on the accumulated result based on the recognition confidence score, generating a semantic evolution result that reflects the trend of demand development. When the semantic evolution result meets the triggering conditions of the corresponding judgment stage, the system determines that the emergency demand type is medical assistance or emergency rescue, and accordingly controls the elevator emergency evacuation robot to perform voice reassurance and information confirmation for the trapped person, report emergency information to the external monitoring terminal, and provide evacuation guidance prompts.

[0027] In a preferred embodiment of the present invention, the speech text and confidence level generation module includes: First, the speech segments output by the speech signal segment acquisition module are subjected to speech validity detection to confirm that the speech segments contain identifiable speech content. After confirming the validity of the speech segments, feature extraction processing is performed on the speech segments to convert them into feature representations for speech recognition. Subsequently, the feature representations are input to a pre-configured speech recognition processing unit to recognize the speech segments segment by segment and output the corresponding text information. While generating the text information, the credibility of the matching results of each speech unit during the speech recognition process is evaluated, and a recognition confidence score representing the reliability of the text information is generated based on the consistency, stability, and continuity of the speech units in the entire speech segment. Finally, the text information and its corresponding recognition confidence score are correlated and output as input data for subsequent semantic state construction.

[0028] In a preferred embodiment of the present invention, the emergency semantic state construction module includes: The semantic fragment localization and generation unit is used to locate and process the content related to urgency, physical condition and intention to seek help in the text information, and generate corresponding semantic fragments and time markers that correspond one-to-one with the semantic fragments. The semantic category mapping generation unit is used to perform category mapping processing on the semantic fragments based on their content, and generate corresponding semantic category identifiers; The confidence level admission determination unit is used to compare the recognition confidence level corresponding to the semantic segment with the preset confidence level admission conditions to generate an admission determination result that represents whether the semantic segment is allowed to participate in subsequent processing. The semantic state initial generation unit is used to determine the corresponding initial state type based on the semantic category identifier when the admission judgment result characterization allows participation in subsequent processing, and to assign an initial influence degree to the initial state type in combination with the recognition confidence of the semantic segment, thereby generating the initial semantic state. The semantic state temporal update unit is used to update the initial semantic state over time according to the time identifier, and generate an updated semantic state that reflects the current influence of the semantic state. The semantic state solidification output unit is used to set the updated semantic state to an irreversible state and use it as semantic state data when the updated semantic state meets the preset solidification conditions. The semantic state data includes the corresponding recognition confidence level.

[0029] In this embodiment of the invention, by locating content related to urgency, physical condition, and intent to seek help in the text information and generating semantic segments with time stamps, the speech recognition results are no longer processed as part of the overall text. Instead, they enter the subsequent analysis process as semantic units with clear time attributes, thereby improving the traceability of semantic information in the time dimension. Furthermore, by classifying the semantic segments and combining them with recognition confidence for admission determination, only semantic segments that meet the confidence conditions participate in the construction of semantic states, avoiding the direct impact of unstable or ambiguous speech content on emergency state judgment. An initial semantic state is generated based on semantic category and recognition confidence, and the influence of the semantic state is updated over time in subsequent processing. This allows the semantic expression of the same trapped person at different points in time to be continuously reflected in the form of state evolution. When the semantic state meets the solidification condition, it is set as an irreversible state, which is conducive to the stable preservation of confirmed important semantic information during the emergency process and avoids repeated corrections to the emergency judgment formed due to subsequent changes in expression.

[0030] In a preferred embodiment of the present invention, the semantic fragment localization and generation unit includes: First, the text information output by the speech and confidence generation module is scanned sentence by sentence to identify keywords or phrases in the text that indicate an emergency, physical sensations, or a need for help. After identifying the relevant content, the text segment containing the content is identified as a candidate semantic segment. Then, based on the time position of the speech segment in the original speech signal, the occurrence time corresponding to the candidate semantic segment is determined, and a time identifier corresponding to the semantic segment is generated. Finally, the semantic segment is bound to its time identifier and output, so that each semantic segment has a clear time attribute for subsequent timing processing.

[0031] In a preferred embodiment of the present invention, the semantic category mapping generation unit includes: First, the semantic fragments output by the semantic fragment localization and generation unit are received, and the core expressive content in the semantic fragments is parsed. Then, the parsed expressive content is compared one by one with a preset set of semantic categories, which includes at least urgency category, physical state category, and help-seeking intention category. During the comparison process, the target semantic category to which the semantic fragment belongs is determined based on the main semantic features reflected in the semantic fragment. When the semantic fragment contains multiple semantic features, the main semantic category is determined according to a preset priority rule. Finally, a semantic category identifier corresponding to the determined semantic category is generated for the semantic fragment and output for subsequent semantic state construction processing.

[0032] In a preferred embodiment of the present invention, the confidence level admission determination unit includes: First, the recognition confidence level corresponding to the semantic segment is obtained. Then, the recognition confidence level is compared with a pre-set confidence level admission condition, which is used to limit the minimum level of confidence required for the semantic segment to enter the semantic state construction process. When the recognition confidence level meets the confidence level admission condition, an admission judgment result is generated that allows the semantic segment to participate in subsequent processing. When the recognition confidence level does not meet the confidence level admission condition, an admission judgment result is generated that prohibits the semantic segment from participating in subsequent processing. Finally, the admission judgment result and the corresponding semantic segment are output together to provide a constraint basis for the initial generation of the semantic state.

[0033] In a preferred embodiment of the present invention, the semantic state initial generation unit includes: First, the system receives the admission determination result, corresponding semantic fragment, and semantic category identifier from the confidence admission determination unit. When the admission determination result indicates that the semantic fragment is allowed to participate in subsequent processing, the system determines the initial state type corresponding to the semantic fragment according to the semantic category identifier, so that semantic fragments of different semantic categories are mapped to different types of semantic states. Then, the system obtains the recognition confidence level corresponding to the semantic fragment, and converts the recognition confidence level into a state value that represents the initial influence level of the semantic state according to the preset influence level assignment rule. The conversion process includes determining the corresponding influence level according to the preset interval in which the recognition confidence level is located. Finally, the system combines the initial state type, the initial influence level, and the time identifier corresponding to the semantic fragment to generate the initial semantic state for subsequent time-series update processing.

[0034] In a preferred embodiment of the present invention, the semantic state solidification output unit includes: First, the system continuously receives updated semantic states from the semantic state time-series update unit and records the current impact level of the updated semantic states and their corresponding time information. Then, it compares the changes in the impact level of the updated semantic states with preset solidification conditions to determine whether the semantic state has entered a stable state. When the determination result indicates that the updated semantic state meets the solidification conditions, the semantic state is marked as a solidified state, and subsequent impact level rollback or weakening operations are prohibited for the semantic state. Finally, the semantic states in the solidified state, along with their corresponding recognition confidence, are output as semantic state data for the cumulative processing of the time window semantic evolution analysis module.

[0035] In a preferred embodiment of the present invention, a preset solidification condition is used to determine whether the updated semantic state needs to be set to an irreversible state, and the method for setting the condition includes: First, based on the category of semantic states, corresponding solidification judgment rules are set for different semantic categories, so that semantic states related to urgency, physical condition, and the intention to seek help have different solidification requirements. Second, the solidification conditions include at least the persistence condition that the semantic state maintains its influence without decreasing within multiple consecutive update time windows, and the intensity condition that the influence of the semantic state reaches or exceeds a preset stability threshold. Among them, the persistence condition is used to judge the stability of the semantic state in the time dimension, and the intensity condition is used to judge the certainty of the semantic state in semantic expression. When the updated semantic state satisfies both the persistence condition and the intensity condition, it is determined that the semantic state meets the solidification conditions. By setting the preset solidification conditions in the above manner, the semantic state is solidified only when it is sufficiently stable and has a clear direction, thereby providing reliable input for subsequent semantic evolution analysis and emergency demand determination.

[0036] In a preferred embodiment of the present invention, the time window semantic evolution analysis module includes: The analysis time window generation unit is used to generate at least one time window for semantic evolution analysis based on the current time and a preset time span. The semantic state window mapping unit is used to map semantic state data to a time window according to its corresponding time identifier, and generate a set of semantic states within the window. The semantic state order determination unit is used to sort the semantic state set within the window in a temporal order according to the time identifier corresponding to the semantic state data, and generate an ordered semantic state sequence. The semantic state influence adjustment unit is used to adjust the influence of each semantic state in the sequence accumulation process according to the recognition confidence of each semantic state in the ordered semantic state sequence, so as to generate a controlled accumulation sequence. The semantic state change trajectory generation unit is used to record the state trajectory of the semantic state corresponding to the same semantic category changing over time in a controlled cumulative sequence. The semantic evolution result locking unit is used to generate and lock the semantic evolution result for emergency demand determination based on the state trajectory.

[0037] In this embodiment of the invention, by constructing a time window for semantic evolution analysis and mapping semantic state data to the corresponding time window according to time identifiers, the analysis of the emergency state of trapped personnel is limited to a clear time range, which is beneficial to controlling the timeliness of semantic evolution analysis. After chronologically sorting the set of semantic states within the time window, an ordered semantic state sequence is formed, enabling subsequent processing to analyze the semantic states based on the actual order of occurrence. By introducing identification confidence to adjust the degree of influence of semantic states in cumulative processing, the contribution paths of semantic states from different confidence sources to the cumulative results are differentiated within the same time window, thereby ensuring that the semantic evolution analysis results are more closely aligned with the actual expressions of the trapped personnel. Furthermore, by recording the state trajectory formed by the changes of semantic states under the same semantic category over time, and generating semantic evolution results for emergency demand determination based on the state trajectory, the output results reflect the changing process of emergency demands within the time window, rather than just the judgment result at a single point in time.

[0038] In a preferred embodiment of the present invention, the semantic evolution result locking unit includes: First, the semantic state change trajectory output by the semantic state change trajectory generation unit is received. The semantic state change trajectory contains continuous change information of semantic states in the time dimension under the same semantic category. Then, the integrity of the semantic state change trajectory is checked to confirm that the trajectory covers valid semantic state data within a preset time window. After the check is completed, the overall change trend of the semantic category within the time window is summarized and analyzed according to the segmented change characteristics corresponding to each state change segment in the semantic state change trajectory. The analysis process includes judging whether the direction of change of the degree of influence of the semantic state is consistent and whether there is a rollback in the change process. When the judgment result indicates that the semantic state change trend remains continuous and consistent within the time window, the corresponding semantic evolution result is generated. Finally, the generated semantic evolution result is set to a locked state, and rollback or regeneration processing of the semantic evolution result is prohibited, thereby ensuring the stability of the semantic evolution result output to the emergency demand determination stage.

[0039] In a preferred embodiment of the present invention, a preset time span is used to limit the time range covered by the semantic evolution analysis of the time window, and the method for setting the preset time span includes: First, based on the continuity of the voice expressions of trapped individuals in elevator emergency scenarios, a basic time span for semantic evolution analysis is set to cover the time required for a trapped individual to complete a full expression of help. Second, the basic time span is categorized according to the type of semantic state, with different time span ranges corresponding to semantic states related to urgency, physical state, and the intention to seek help. Urgency-related semantic states correspond to shorter time spans to quickly reflect changes in urgency, while physical state and intention-related semantic states correspond to longer time spans to observe continuous changes in state. In practice, the start time of the time window is determined by tracing back the time span from the current moment, and the end time of the time window is set to the current moment. By setting the preset time span in this way, the semantic evolution analysis is both time-sensitive and can cover the continuous changes in the semantic state of trapped individuals, providing a reliable time basis for subsequent emergency demand determination.

[0040] In a preferred embodiment of the present invention, the emergency demand determination triggering module includes: The demand determination candidate generation unit is used to parse and process the semantic evolution results, extract the determination elements that reflect changes in urgency, physical state, and intention to seek help, and match the determination elements with the preset emergency demand determination dimensions. Based on the corresponding results, demand candidate information for emergency demand determination is generated. The demand candidate threshold comparison unit is used to compare the demand candidate information with the preset multi-level emergency demand judgment thresholds and generate the corresponding threshold matching results. The pre-triggering state generation unit is used to generate a pre-triggering state identifier corresponding to the emergency demand type when the threshold matching result meets the pre-triggering condition. The confirmation condition continuous verification unit is used to continuously compare and verify the semantic evolution results obtained after the pre-trigger state identifier is generated, and generate confirmation verification results. The emergency demand type determination unit is used to determine the corresponding emergency demand type based on the pre-trigger status identifier when the verification result meets the confirmation conditions, and set the emergency demand type as a non-reversible determination result.

[0041] In this embodiment of the invention, by parsing the semantic evolution results and extracting judgment elements reflecting changes in urgency, physical condition, and intent to seek help, the determination of emergency needs is based on multi-dimensional semantic changes, rather than a single semantic indicator. After mapping the judgment elements to preset emergency need judgment dimensions, candidate need information is generated, ensuring a clear mapping relationship between different types of emergency needs during the judgment process. By comparing the candidate need information with multi-level emergency need judgment thresholds and generating threshold matching results, the emergency need judgment process is given staged constraints, enabling the differentiation of need states under different severity levels. A pre-triggering state identifier is generated when pre-triggering conditions are met, and new semantic evolution results are continuously verified in subsequent processing, ensuring that the final determination of emergency needs depends on the stability of continuous semantic evolution. When confirmation conditions are met, the emergency need type is set as an irreversible determination, which helps avoid frequent changes in need types during emergency response, thereby providing a stable control basis for subsequent interactive execution.

[0042] In a preferred embodiment of the present invention, the demand determination candidate generation unit includes: First, the semantic evolution results output by the semantic evolution result locking unit are received. These results include, at least, the changing trends of semantic states related to urgency, physical condition, and the intention to seek help within a time window. Then, the semantic evolution results are analyzed item by item to identify judgment elements reflecting changes in urgency, physical condition, and the intention to seek help. These judgment elements include the direction of change in the degree of influence of the semantic state, the duration of the change, and whether the change reaches a preset intensity level. After extracting the judgment elements, each element is categorized according to its corresponding semantic category, forming a structured set of judgment elements. Finally, the set of judgment elements is matched with preset emergency demand judgment dimensions to generate candidate demand information for subsequent threshold comparison.

[0043] In a preferred embodiment of the present invention, the candidate threshold comparison unit includes: First, the system receives the demand candidate information output by the demand determination candidate generation unit and reads the determination elements corresponding to different emergency demand determination dimensions. Then, for each emergency demand determination dimension, the corresponding determination element is compared with multi-level emergency demand determination thresholds in a preset comparison order. The step-by-step comparison includes determining whether the change intensity represented by the determination element reaches the condition limited by the corresponding level threshold. During the comparison process, when the determination element meets the threshold condition of a certain level, a threshold matching result corresponding to that level threshold is generated. Finally, the threshold matching results of each emergency demand determination dimension are summarized and output for subsequent pre-trigger state generation.

[0044] In a preferred embodiment of the present invention, the emergency demand determination dimension is used to determine the demand from different perspectives based on the semantic evolution results, and its setting method includes: First, based on the types of needs that trapped individuals may express in elevator emergency scenarios, at least three dimensions are defined: urgency level assessment, physical condition assessment, and assistance intent assessment. The urgency level assessment dimension reflects changes in the urgency of the trapped individual's situation; the physical condition assessment dimension reflects changes in the trapped individual's physical condition; and the assistance intent assessment dimension reflects the clarity of the trapped individual's proactive assistance behavior. Second, for each assessment dimension, a type of assessment element is defined that can be used for assessment, ensuring that the assessment of different dimensions has corresponding semantic evolution basis. By setting emergency need assessment dimensions in the above manner, the need assessment process can comprehensively reflect the true emergency need status of trapped individuals from multiple semantic perspectives.

[0045] In a preferred embodiment of the present invention, a multi-level emergency demand determination threshold is used to classify and determine emergency demands of different degrees, and the method for setting the threshold includes: First, based on the response levels of the emergency response process, multiple judgment levels are set for each emergency demand judgment dimension, enabling emergency demands to be distinguished into different degrees of severity. Second, corresponding threshold conditions are set for each judgment level. These threshold conditions are used to limit the requirements that the judgment elements must meet in terms of change intensity, change persistence, and change consistency. Among them, lower-level thresholds correspond to basic response demands, and higher-level thresholds correspond to emergency response demands. In specific implementation, the judgment elements in the demand candidate information are compared with the above-mentioned multi-level threshold conditions level by level, and the highest level that is met is taken as the judgment result for that dimension. By setting multi-level emergency demand judgment thresholds in the above manner, the judgment of emergency demand types has hierarchy and controllability, which facilitates the implementation of corresponding level of response measures in the subsequent emergency interaction execution phase.

[0046] In a preferred embodiment of the present invention, the pre-triggered state generation unit includes: First, the threshold matching results output by the candidate threshold comparison unit are received, and the judgment level corresponding to each emergency demand judgment dimension is identified. Then, based on whether the pre-triggering condition is met in the threshold matching results, it is determined whether to enter the pre-triggering process. When the judgment level of at least one emergency demand judgment dimension meets the pre-triggering condition, the emergency demand type corresponding to that judgment dimension is determined, and a pre-triggering status identifier corresponding to that emergency demand type is generated. The pre-triggering status identifier is used to indicate that the emergency demand has entered the pending confirmation state. Finally, the pre-triggering status identifier is bound to the time information of the current semantic evolution result and output as the input for subsequent continuous verification processing.

[0047] In a preferred embodiment of the present invention, the pre-triggering condition is used to define the rules for determining whether an emergency need moves from the candidate state to the pre-triggering state, and the method for setting the pre-triggering condition includes: First, based on the timeliness requirements of the emergency response process, corresponding pre-trigger conditions are set for different types of emergency needs. These pre-trigger conditions include at least a threshold matching level condition and a trend consistency condition. The threshold matching level condition limits the minimum judgment level that candidate information for the need must reach, while the trend consistency condition ensures that the semantic evolution results maintain a consistent direction of change within the time window. Second, in actual judgment, the pre-trigger condition is only satisfied when the threshold comparison result of the candidate need meets the corresponding threshold matching level condition and the semantic evolution result does not show a reverse change within the current time window. By setting pre-trigger conditions in this way, emergency needs have a buffer and verification phase before formal confirmation, avoiding direct triggering of emergency response due to a single semantic fluctuation.

[0048] In a preferred embodiment of the present invention, the verification condition continuity verification unit includes: First, after generating the pre-trigger status identifier, the system continuously receives subsequent semantic evolution results output by the time window semantic evolution analysis module. Then, it compares the changing trend of the corresponding emergency demand type in the subsequent semantic evolution results with the semantic evolution results when the pre-trigger status identifier was generated, determining whether the changing trend remains consistent or is further strengthened. During continuous verification, the system records the number of times or duration that the subsequent semantic evolution results meet the confirmation conditions. When the subsequent semantic evolution results meet the confirmation conditions within multiple consecutive time windows, a confirmation verification result indicating that the confirmation conditions are met is generated. When a subsequent semantic evolution result fails to meet the confirmation conditions, a confirmation verification result indicating that the confirmation conditions are not met is generated, and the current confirmation process is terminated.

[0049] In a preferred embodiment of the present invention, the emergency demand type determination unit includes: First, the system receives the confirmation verification result output by the confirmation condition continuous verification unit. When the confirmation verification result indicates that the confirmation condition is met, the system reads the pre-trigger status identifier corresponding to the confirmation verification result. Then, based on the emergency demand type information recorded in the pre-trigger status identifier, the system determines the final emergency demand type. After determining the emergency demand type, the system marks the emergency demand type as confirmed and prohibits re-judgment or downgrade processing for this emergency demand type. Finally, the confirmed emergency demand type is output to the emergency interaction execution module as the sole control basis for subsequent interaction and handling operations.

[0050] In a preferred embodiment of the present invention, the semantic state time-series update unit includes: The update time window construction unit is used to construct a discrete update time window for semantic state update based on the time identifier corresponding to the semantic fragment, and generate the corresponding update time window identifier. The update trigger determination unit is used to determine whether there is a new semantic fragment entering the update time window based on the update time window identifier, and to generate an update trigger determination result; The semantic category conflict determination unit is used to compare the semantic category corresponding to the newly added semantic fragment with the semantic category corresponding to the existing semantic state when the update triggers the determination result representation to perform update processing, and generate a semantic category conflict determination result. The update method selection unit is used to compare the recognition confidence of the new semantic segment with the recognition confidence of the existing semantic state when the semantic category conflict determination result indicates that the new semantic segment and the existing semantic state belong to the same semantic category. If the comparison result meets the preset coverage condition, it determines to perform a coverage update on the existing semantic state. If the comparison result does not meet the preset coverage condition, it determines to perform a suppression update on the existing semantic state. The update state generation unit is used to update the influence of the existing semantic state according to the method of overwrite update or suppress update, and generate an updated semantic state as input for subsequent processing.

[0051] In this embodiment of the invention, by constructing discrete update time windows and combining them with the time identifiers of semantic segments to determine whether to trigger semantic state updates, the updates of semantic states no longer depend on a single voice input, but are controlled by the time window boundary, thereby avoiding frequent modifications to the semantic state due to high-frequency voice input. By introducing semantic category conflict determination when triggering updates, new semantic segments and existing semantic states are distinguished at the semantic category level, ensuring that subsequent update processing is only performed under the premise of semantic category consistency, which helps maintain the category consistency of semantic states. On this basis, by comparing the recognition confidence corresponding to the new semantic segment with the recognition confidence contained in the existing semantic state, two update methods, overlay update and suppression update, are distinguished, allowing the semantic state update process to be selectively executed according to the expression confidence, avoiding undue overlay of the formed semantic state by low-confidence expressions, thereby ensuring the stability and continuity of the semantic state in the process of temporal evolution.

[0052] In a preferred embodiment of the present invention, the update state generation unit includes: First, the update method selection unit outputs the update method determination result, along with relevant information on the existing semantic state and the newly added semantic fragment. When the update method determination result indicates that a coverage update is to be performed, the identification confidence and semantic category information corresponding to the newly added semantic fragment are obtained. Based on the preset influence degree update rule, the influence degree corresponding to the identification confidence of the newly added semantic fragment is used as the influence degree of the updated semantic state. During the coverage update process, the time identifier of the existing semantic state is retained, and the update time identifier is updated to the time corresponding to the current update time window. When the update method determination result indicates that a suppression update is to be performed, the influence degree of the existing semantic state remains unchanged, and only the time identifier of the semantic state is updated to indicate that the semantic state has been reconfirmed within the current update time window. Finally, the semantic state after the update process is completed is output as the updated semantic state for subsequent semantic state solidification judgment or time window semantic evolution analysis.

[0053] In a preferred embodiment of the present invention, a preset coverage condition is used to determine whether a newly added semantic segment is qualified to cover an existing semantic state. The method for setting this condition includes: First, based on the reliability requirements during the semantic state update process, basic conditions are set for the coverage update to ensure that the recognition confidence of the newly added semantic fragment is not lower than the recognition confidence of the existing semantic state. Second, a consistency condition is introduced to limit the consistency between the newly added semantic fragment and the existing semantic state in terms of semantic category, ensuring that the coverage update is only allowed when the semantic categories are the same. Third, a time-related condition is introduced to limit the newly added semantic fragment to occur within the current update time window, so as to avoid retrospective coverage of the formed semantic state by historical semantic fragments. In specific implementation, the preset coverage condition is determined to be met only when the newly added semantic fragment simultaneously meets the above recognition confidence condition, semantic category consistency condition, and time-related condition. By setting the preset coverage condition in the above manner, the coverage update of the semantic state is based on reliability, consistency, and temporal rationality, thereby ensuring the stability and controllability of the semantic state update process.

[0054] In a preferred embodiment of the present invention, the semantic state influence adjustment unit includes: The confidence segmentation unit is used to divide the ordered semantic state sequence into multiple confidence influence segments based on the recognition confidence contained in each semantic state in the ordered semantic state sequence, and generate segment identifiers. The segment participation eligibility determination unit is used to determine whether the corresponding semantic state is eligible to participate in the cumulative processing based on the preset participation rules according to each confidence level of the affected segment, and to generate the participation eligibility determination result. The cross-segment impact blocking determination unit is used to generate a blocking determination result to block the impact of a low-confidence impact segment on the semantic state of a high-confidence impact segment when the semantic state of the low-confidence impact segment does not meet the participation qualification determination result. The accumulation method determination unit is used to determine the participation method of each semantic state in the accumulation process based on the participation qualification determination result and the blocking determination result; The controlled accumulation result generation unit is used to perform accumulation processing on semantic states that have passed the participation qualification determination and have not been blocked, according to the participation method, to generate a controlled accumulation sequence as input for subsequent trajectory analysis.

[0055] In this embodiment of the invention, the ordered semantic state sequence is divided into multiple confidence-influenced segments based on the identification confidence level. This ensures that semantic states from different confidence sources have clear segment affiliations during the accumulation processing stage, facilitating subsequent hierarchical control of semantic state participation. By performing participation eligibility determination on each confidence-influenced segment, clear admission conditions are established for whether a semantic state can participate in the accumulation processing, thus preventing the direct introduction of semantic states that do not meet the participation rules into the accumulation process. When a semantic state in a low-confidence-influenced segment does not meet the participation eligibility determination result, a cross-segment influence blocking determination result is generated to restrict its influence on the semantic states in high-confidence-influenced segments, ensuring that high-confidence semantic states maintain their dominant role in the accumulation processing. Furthermore, by combining the participation eligibility determination result and the blocking determination result, the participation mode of each semantic state in the accumulation processing is determined, ensuring that the accumulation process executes according to the expected rules, ultimately generating a controlled accumulation sequence and providing structured input for subsequent semantic trajectory analysis.

[0056] In a preferred embodiment of the present invention, the segment participation qualification determination unit includes: First, the system receives the confidence-affected segment information output by the confidence segment division unit and obtains the set of semantic states contained in each confidence-affected segment. Then, for each confidence-affected segment, it reads the identification confidence, semantic category, and current influence level corresponding to each semantic state within the segment. Based on this, it calls the preset participation rule corresponding to the confidence-affected segment to perform participation qualification determination processing on each semantic state within the segment. The determination processing includes judging whether the identification confidence of the semantic state meets the minimum requirement for participation in the cumulative processing of that segment, and judging whether the semantic state is in a fixed or blocked state. When the semantic state simultaneously meets all the conditions in the preset participation rule, a participation qualification determination result representing that the semantic state is allowed to participate in the cumulative processing is generated. When the semantic state does not meet any condition, a participation qualification determination result representing that the semantic state is not allowed to participate in the cumulative processing is generated. Finally, the participation qualification determination result is associated with the corresponding semantic state and output for subsequent cross-segment influence blocking determination and accumulation method determination processing.

[0057] In a preferred embodiment of the present invention, a preset participation rule is used to limit whether semantic states within different confidence level influence segments are eligible to participate in semantic state accumulation processing. The method for setting this rule includes: First, based on the role of semantic states in the cumulative processing, corresponding participation rules are set for different confidence levels of influence segments, so that high-confidence influence segments and low-confidence influence segments adopt different admission criteria. Second, each preset participation rule includes at least identification confidence requirements, semantic state stability requirements, and update state restriction requirements. The identification confidence requirements limit the identification confidence of the semantic state to the minimum level allowed for participation in that segment. The semantic state stability requirements limit the semantic state from regressing or undergoing abnormal changes within the current time window. The update state restriction requirements limit the semantic state from being marked as inhibited or blocked. In practice, the current attributes of the semantic state are compared with the preset participation rules of the corresponding segment item by item. Only when all requirements are met is the semantic state deemed eligible to participate in the cumulative processing. By setting preset participation rules in the above manner, semantic states from different confidence levels receive differentiated participation control in the cumulative processing stage, thereby ensuring the reliability and controllability of the semantic evolution analysis results.

[0058] In a preferred embodiment of the present invention, the semantic state change trajectory generation unit includes: The category state sequence extraction unit is used to extract semantic states belonging to the same semantic category based on the semantic categories corresponding to the semantic states in the controlled cumulative sequence, and generate a semantic category state sequence. The state change inflection point recognition unit is used to compare the degree of influence of adjacent semantic states in the semantic category state sequence, identify the position where the change direction changes or the change amplitude reaches the preset amplitude change threshold, and generate trajectory inflection point identifiers. The trajectory segmentation generation unit is used to divide the semantic category state sequence into multiple state change segments with the trajectory inflection point marker as the boundary, and generate a segmented trajectory sequence. The segmented change feature generation unit is used to extract segmented change features that characterize the semantic state change trend within each state change segment based on each state change segment. The trajectory result combination unit is used to combine segmented change features in chronological order to generate a semantic state change trajectory that serves as the basis for generating semantic evolution results.

[0059] In this embodiment of the invention, semantic states corresponding to the same semantic category are extracted from the controlled cumulative sequence to form a semantic category state sequence. This allows subsequent analysis to be performed within a single semantic category dimension, avoiding the difficulty in interpreting state changes caused by the mixing of different semantic categories. By comparing the changes in the degree of influence of adjacent semantic states in the semantic category state sequence, the positions where the direction of change changes or the magnitude of change reaches a preset magnitude change threshold are identified, thereby accurately locating key nodes of state change in the time series. The semantic category state sequence is segmented based on trajectory inflection point markers, dividing the semantic state change process into multiple continuous state change segments. By extracting trend features from each state change segment, the overall trend of semantic state change over time can be clearly expressed. By combining the segmented change features in chronological order, a semantic state change trajectory is generated for generating semantic evolution results. This is beneficial for understanding the development process of the semantic state of trapped personnel from an overall change perspective during the emergency demand determination process.

[0060] In a preferred embodiment of the present invention, the segmented change feature generation unit includes: First, the system receives a segmented trajectory sequence output by the trajectory segmentation generation unit. This segmented trajectory sequence consists of multiple state change segments arranged in chronological order. Then, for each state change segment, the system obtains the influence degree of the first and last semantic states within that segment, along with their corresponding time markers. Based on this, the system determines the direction of change in the influence degree of the semantic states within the state change segment to identify whether the segment is continuously increasing, continuously decreasing, or remaining relatively stable. Simultaneously, the system describes the persistence of the semantic state change by considering the time length covered by the state change segment. Finally, the system combines the change direction determination result with the persistence description result to generate segmented change features that characterize the overall change trend of the state change segment, and outputs these features for subsequent trajectory result combination processing.

[0061] In a preferred embodiment of the present invention, the trajectory result combination unit includes: First, the system receives multiple segmented change features output by the segmented change feature generation unit and sorts them according to their corresponding time identifiers. Then, following the sorted time order, the segmented change features are sequentially connected to form a complete semantic state change description sequence. During the connection process, the continuity of change between adjacent segmented change features is verified to confirm the coherence of semantic state changes in the time dimension. When the verification result indicates continuous change, the segmented change features are directly combined to form a semantic state change trajectory. When the verification result indicates discontinuity, the corresponding change boundary information is recorded in the semantic state change trajectory. Finally, the generated semantic state change trajectory is output as input data for subsequent semantic evolution result locking processing.

[0062] In a preferred embodiment of the present invention, a preset amplitude change threshold is used to determine whether a change in the degree of influence of semantic state constitutes a significant change. The method for setting the threshold includes: First, a reference interval is set to describe the magnitude of change in the degree of influence of semantic states within the system. Second, based on the characteristics of changes in the semantic expressions of trapped individuals under normal fluctuation conditions, a magnitude limit is determined within the reference interval to distinguish between normal and significant changes. This magnitude limit is used to determine whether the change in the degree of influence between adjacent semantic states exceeds the acceptable range of normal fluctuations. In specific implementation, by comparing the change in the degree of influence of adjacent semantic states with the magnitude limit, when the change exceeds the limit, it is determined that a preset magnitude change threshold is met. By setting the preset magnitude change threshold in the above manner, the identification of state change inflection points can capture key changes in semantic states while avoiding misjudging subtle fluctuations as important changes, thus providing a stable basis for constructing the semantic state change trajectory.

[0063] In a preferred embodiment of the present invention, the cross-segment impact blocking determination unit includes: The segment state acquisition unit is used to acquire the participation qualification determination results corresponding to each semantic state in the low-confidence influence segment and the high-confidence influence segment before performing semantic state accumulation processing; The cross-segment influence relationship determination unit is used to determine whether the semantic state in the low-confidence influence segment constitutes a cross-segment influence on the semantic state in the high-confidence influence segment when the semantic state in the low-confidence influence segment does not meet the participation qualification determination result. The blocking determination generation unit is used to generate a blocking determination result when the cross-segment influence relationship determination is established, which is used to block the influence of the semantic state of the low-confidence influence segment on the semantic state of the high-confidence influence segment.

[0064] In this embodiment of the invention, by obtaining the participation eligibility determination results for each semantic state in the low-confidence influence segment and the high-confidence influence segment before semantic state accumulation processing, the judgment of cross-segment influence is based on a clear participation status. Furthermore, when a semantic state in the low-confidence influence segment does not meet the participation eligibility determination results, it is determined that it constitutes a cross-segment influence on the semantic state in the high-confidence influence segment, thus logically clarifying the interference relationship between the low-confidence semantic state and the high-confidence semantic state. When the cross-segment influence relationship is established, a blocking determination result is generated, restricting the semantic state in the low-confidence influence segment from affecting the semantic state in the high-confidence influence segment during accumulation processing. This helps maintain the dominant position of the high-confidence semantic state during the semantic state accumulation stage, thereby ensuring that subsequent semantic evolution analysis results can more accurately reflect the actual emergency status changes of the trapped personnel.

[0065] In a preferred embodiment of the present invention, the accumulation mode determination unit includes: The cumulative judgment result collection unit is used to collect the participation qualification judgment results and cross-segment impact blocking judgment results corresponding to each semantic state before performing semantic state accumulation processing, and generate the cumulative judgment input results for accumulation method judgment. The cumulative participation mode determination unit is used to perform the following determination processes on each semantic state in sequence based on the cumulative determination input result: When the corresponding semantic state has cross-segment impact on the blocking determination result, the cumulative participation method of the semantic state is determined to be prohibited from participating in the cumulative processing; When the corresponding semantic state does not have a cross-segment impact on the blocking determination result and the participation qualification determination result indicates that participation is allowed, the cumulative participation method of the semantic state is determined to be the cumulative processing of allowed participation. When the corresponding semantic state does not affect the blocking determination result across segments and the participation qualification determination result indicates that participation is not allowed, the cumulative participation method of the semantic state is determined to be restricted participation cumulative processing. The restricted participation cumulative processing includes restricting the semantic state to participate in the cumulative processing without affecting other semantic states.

[0066] In this embodiment of the invention, by aggregating the participation eligibility determination results and cross-segment impact blocking determination results corresponding to each semantic state before semantic state accumulation processing, a unified accumulation determination input result is formed, ensuring that the determination process of semantic state participation mode has a complete determination basis. Based on the accumulation determination input result, the accumulation participation mode determination is performed on each semantic state sequentially, so that the participation mode of each semantic state in the accumulation processing is determined by explicit rules. When a semantic state has a cross-segment impact blocking determination result, its participation mode is determined to be prohibited from participating in the accumulation processing, which helps to avoid its impact on other semantic states. When a semantic state does not have a blocking determination result and the participation eligibility determination result indicates that participation is allowed, its participation mode is determined to be allowed to participate in the accumulation processing, allowing it to participate normally in semantic state accumulation. When a semantic state does not have a blocking determination result but the participation eligibility determination result indicates that participation is not allowed, its participation mode is determined to be restricted participation in the accumulation processing, and it is restricted to only participating in the accumulation processing that does not affect other semantic states, thereby achieving fine control over the degree of semantic state participation and ensuring the stability and controllability of the accumulation result.

[0067] In a preferred embodiment of the present invention, the state change inflection point identification unit includes: The change comparison object construction unit is used to construct adjacent state comparison objects for change analysis in a semantic category state sequence for adjacent semantic states. The adjacent state comparison objects contain the degree of influence of adjacent semantic states and the corresponding time identifier. The change direction determination unit is used to determine the change direction between adjacent semantic states based on the influence degree change relationship between adjacent semantic states in the adjacent state comparison object, and generate the change direction determination result. The change magnitude determination unit is used to determine whether the change amount meets the preset change magnitude threshold based on the change in the influence degree of adjacent semantic states in the adjacent state comparison object, and to generate the change magnitude determination result. The inflection point determination and generation unit is used to determine the corresponding position as the state change inflection point when the change direction determination result represents a change in the direction of change of the adjacent semantic state, or when the change magnitude determination result represents a change in the amount of change of the adjacent semantic state that meets the preset magnitude change threshold, and to generate a trajectory inflection point identifier corresponding to the state change inflection point.

[0068] In this embodiment of the invention, by constructing adjacent state comparison objects containing influence degree and time identifiers for adjacent semantic states in the semantic category state sequence, the state change analysis has clear comparison objects and time references. By analyzing the relationship between changes in the influence degree of adjacent semantic states and generating a change direction determination result, the change direction of semantic states in the time series can be accurately identified; simultaneously, by judging the change amount of the influence degree of adjacent semantic states and generating a change amplitude determination result, the intensity of state change can be quantitatively described. When the change direction changes or the change amount meets a preset amplitude change threshold, the corresponding position is determined as the state change inflection point and a trajectory inflection point identifier is generated, so that the key nodes in the semantic state change process are clearly marked. This is beneficial for accurately grasping the turning point of emergency state change in the subsequent trajectory segmentation and semantic evolution result generation process, thereby improving the interpretability of the overall semantic evolution analysis.

[0069] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A voice interaction system for an elevator emergency evacuation robot, characterized in that, The system includes: The voice signal segmentation acquisition module is used to acquire continuous voice signals of people trapped in the elevator car in an emergency situation, and to segment the continuous voice signals according to a preset time interval to generate multiple voice segments. The speech-text and confidence generation module is used to perform speech recognition processing on each speech segment, convert the speech segment into corresponding text information, and generate a corresponding recognition confidence score for each piece of text information during the conversion process. The emergency semantic state construction module is used to extract semantic elements representing the degree of urgency, physical state, and intention to seek help from text information, and combine the semantic elements with the corresponding recognition confidence to generate semantic state data that represents the emergency state of trapped personnel and evolves over time. The time window semantic evolution analysis module is used to accumulate multiple semantic state data in chronological order within a preset time window, and to differentiate the influence of different semantic state data on the accumulation process based on the recognition confidence. Based on the changes in semantic state data during the accumulation process, it generates semantic evolution results that reflect the development trend of emergency needs. An emergency demand determination triggering module is used to compare the semantic evolution result with preset emergency demand determination conditions. The emergency demand determination conditions include triggering conditions corresponding to at least one determination stage. When the emergency demand determination conditions are met, the corresponding emergency demand type is determined. The emergency interaction execution module is used to control the elevator emergency evacuation robot to perform corresponding voice feedback, emergency information reporting, or evacuation guidance operations according to the type of emergency need.

2. The voice interaction system of the elevator emergency evacuation robot according to claim 1, characterized in that, The emergency semantic state construction module includes: The semantic fragment localization and generation unit is used to locate and process the content related to urgency, physical condition and intention to seek help in the text information, and generate corresponding semantic fragments and time markers that correspond one-to-one with the semantic fragments. The semantic category mapping generation unit is used to perform category mapping processing on the semantic fragments based on their content, and generate corresponding semantic category identifiers; The confidence level admission determination unit is used to compare the recognition confidence level corresponding to the semantic segment with the preset confidence level admission conditions to generate an admission determination result that represents whether the semantic segment is allowed to participate in subsequent processing. The semantic state initial generation unit is used to determine the corresponding initial state type based on the semantic category identifier when the admission judgment result characterization allows participation in subsequent processing, and to assign an initial influence degree to the initial state type in combination with the recognition confidence of the semantic segment, thereby generating the initial semantic state. The semantic state temporal update unit is used to update the initial semantic state over time according to the time identifier, and generate an updated semantic state that reflects the current influence of the semantic state. The semantic state solidification output unit is used to set the updated semantic state to an irreversible state and use it as semantic state data when the updated semantic state meets the preset solidification conditions. The semantic state data includes the corresponding recognition confidence level.

3. The voice interaction system of the elevator emergency evacuation robot according to claim 1, characterized in that, The time window semantic evolution analysis module includes: The analysis time window generation unit is used to generate at least one time window for semantic evolution analysis based on the current time and a preset time span. The semantic state window mapping unit is used to map semantic state data to a time window according to its corresponding time identifier, and generate a set of semantic states within the window. The semantic state order determination unit is used to sort the semantic state set within the window in a temporal order according to the time identifier corresponding to the semantic state data, and generate an ordered semantic state sequence. The semantic state influence adjustment unit is used to adjust the influence of each semantic state in the sequence accumulation process according to the recognition confidence of each semantic state in the ordered semantic state sequence, so as to generate a controlled accumulation sequence. The semantic state change trajectory generation unit is used to record the state trajectory of the semantic state corresponding to the same semantic category changing over time in a controlled cumulative sequence. The semantic evolution result locking unit is used to generate and lock the semantic evolution result for emergency demand determination based on the state trajectory.

4. The voice interaction system of the elevator emergency evacuation robot according to claim 1, characterized in that, The emergency demand determination and triggering module includes: The demand determination candidate generation unit is used to parse and process the semantic evolution results, extract the determination elements that reflect changes in urgency, physical state, and intention to seek help, and match the determination elements with the preset emergency demand determination dimensions. Based on the corresponding results, demand candidate information for emergency demand determination is generated. The demand candidate threshold comparison unit is used to compare the demand candidate information with the preset multi-level emergency demand judgment thresholds and generate the corresponding threshold matching results. The pre-triggering state generation unit is used to generate a pre-triggering state identifier corresponding to the emergency demand type when the threshold matching result meets the pre-triggering condition. The confirmation condition continuous verification unit is used to continuously compare and verify the semantic evolution results obtained after the pre-trigger state identifier is generated, and generate confirmation verification results. The emergency demand type determination unit is used to determine the corresponding emergency demand type based on the pre-trigger status identifier when the verification result meets the confirmation conditions, and set the emergency demand type as a non-reversible determination result.

5. The voice interaction system of the elevator emergency evacuation robot according to claim 2, characterized in that, The semantic state time-series update unit includes: The update time window construction unit is used to construct a discrete update time window for semantic state update based on the time identifier corresponding to the semantic fragment, and generate the corresponding update time window identifier. The update trigger determination unit is used to determine whether there is a new semantic fragment entering the update time window based on the update time window identifier, and to generate an update trigger determination result; The semantic category conflict determination unit is used to compare the semantic category corresponding to the newly added semantic fragment with the semantic category corresponding to the existing semantic state when the update triggers the determination result representation to perform update processing, and generate a semantic category conflict determination result. The update method selection unit is used to compare the recognition confidence of the new semantic segment with the recognition confidence of the existing semantic state when the semantic category conflict determination result indicates that the new semantic segment and the existing semantic state belong to the same semantic category. If the comparison result meets the preset coverage condition, it determines to perform a coverage update on the existing semantic state. If the comparison result does not meet the preset coverage condition, it determines to perform a suppression update on the existing semantic state. The update state generation unit is used to update the influence of the existing semantic state according to the method of overwrite update or suppress update, and generate an updated semantic state as input for subsequent processing.

6. The voice interaction system of the elevator emergency evacuation robot according to claim 3, characterized in that, The semantic state influence adjustment unit includes: The confidence segmentation unit is used to divide the ordered semantic state sequence into multiple confidence influence segments based on the recognition confidence contained in each semantic state in the ordered semantic state sequence, and generate segment identifiers. The segment participation eligibility determination unit is used to determine whether the corresponding semantic state is eligible to participate in the cumulative processing based on the preset participation rules according to each confidence level of the affected segment, and to generate the participation eligibility determination result. The cross-segment impact blocking determination unit is used to generate a blocking determination result to block the impact of a low-confidence impact segment on the semantic state of a high-confidence impact segment when the semantic state of the low-confidence impact segment does not meet the participation qualification determination result. The accumulation method determination unit is used to determine the participation method of each semantic state in the accumulation process based on the participation qualification determination result and the blocking determination result; The controlled accumulation result generation unit is used to perform accumulation processing on semantic states that have passed the participation qualification determination and have not been blocked, according to the participation method, to generate a controlled accumulation sequence as input for subsequent trajectory analysis.

7. The voice interaction system of the elevator emergency evacuation robot according to claim 3, characterized in that, The semantic state change trajectory generation unit includes: The category state sequence extraction unit is used to extract semantic states belonging to the same semantic category based on the semantic categories corresponding to the semantic states in the controlled cumulative sequence, and generate a semantic category state sequence. The state change inflection point recognition unit is used to compare the degree of influence of adjacent semantic states in the semantic category state sequence, identify the position where the change direction changes or the change amplitude reaches the preset amplitude change threshold, and generate trajectory inflection point identifiers. The trajectory segmentation generation unit is used to divide the semantic category state sequence into multiple state change segments with the trajectory inflection point marker as the boundary, and generate a segmented trajectory sequence. The segmented change feature generation unit is used to extract segmented change features that characterize the semantic state change trend within each state change segment based on each state change segment. The trajectory result combination unit is used to combine segmented change features in chronological order to generate a semantic state change trajectory that serves as the basis for generating semantic evolution results.

8. The voice interaction system of the elevator emergency evacuation robot according to claim 6, characterized in that, The cross-segment impact closure determination unit includes: The segment state acquisition unit is used to acquire the participation qualification determination results corresponding to each semantic state in the low-confidence influence segment and the high-confidence influence segment before performing semantic state accumulation processing; The cross-segment influence relationship determination unit is used to determine whether the semantic state in the low-confidence influence segment constitutes a cross-segment influence on the semantic state in the high-confidence influence segment when the semantic state in the low-confidence influence segment does not meet the participation qualification determination result. The blocking determination generation unit is used to generate a blocking determination result when the cross-segment influence relationship determination is established, which is used to block the influence of the semantic state of the low-confidence influence segment on the semantic state of the high-confidence influence segment.

9. The voice interaction system of the elevator emergency evacuation robot according to claim 6, characterized in that, The accumulation method determination unit includes: The cumulative judgment result collection unit is used to collect the participation qualification judgment results and cross-segment impact blocking judgment results corresponding to each semantic state before performing semantic state accumulation processing, and generate the cumulative judgment input results for accumulation method judgment. The cumulative participation mode determination unit is used to perform the following determination processes on each semantic state in sequence based on the cumulative determination input result: When the corresponding semantic state has cross-segment impact on the blocking determination result, the cumulative participation method of the semantic state is determined to be prohibited from participating in the cumulative processing; When the corresponding semantic state does not have a cross-segment impact on the blocking determination result and the participation qualification determination result indicates that participation is allowed, the cumulative participation method of the semantic state is determined to be the cumulative processing of allowed participation. When the corresponding semantic state does not affect the blocking determination result across segments and the participation qualification determination result indicates that participation is not allowed, the cumulative participation method of the semantic state is determined to be restricted participation cumulative processing. The restricted participation cumulative processing includes restricting the semantic state to participate in the cumulative processing without affecting other semantic states.

10. The voice interaction system of the elevator emergency evacuation robot according to claim 7, characterized in that, The state change inflection point identification unit includes: The change comparison object construction unit is used to construct adjacent state comparison objects for change analysis in a semantic category state sequence for adjacent semantic states. The adjacent state comparison objects contain the degree of influence of adjacent semantic states and the corresponding time identifier. The change direction determination unit is used to determine the change direction between adjacent semantic states based on the influence degree change relationship between adjacent semantic states in the adjacent state comparison object, and generate the change direction determination result. The change magnitude determination unit is used to determine whether the change amount meets the preset change magnitude threshold based on the change in the influence degree of adjacent semantic states in the adjacent state comparison object, and to generate the change magnitude determination result. The inflection point determination and generation unit is used to determine the corresponding position as the state change inflection point when the change direction determination result represents a change in the direction of change of the adjacent semantic state, or when the change magnitude determination result represents a change in the amount of change of the adjacent semantic state that meets the preset magnitude change threshold, and to generate a trajectory inflection point identifier corresponding to the state change inflection point.

Images