Generative rainstorm and flood intelligent risk-avoiding dialogue method, system and equipment and medium

Abstract

The invention relates to the technical field of intelligent emergency management, in particular to a generative rainstorm and flood intelligent risk avoiding dialogue method, system and device and a medium, and the method comprises the steps: loading an expert network model under a dynamic hybrid expert architecture according to the operation state of a portable edge device; preprocessing the voice stream data; predicting an interrupted voice content of the complemented voice stream data; uploading crowd source data of a region where the user is located, and assimilating the verified crowd source data to a preset flood forecasting model; outputting the confidence of the standard text; generating user semantic information and environment data meeting a preset high-reliability condition; and according to the flood and rainstorm early warning level and the position coordinates of the user, generating a dynamic risk avoiding dialogue strategy of the rainstorm and flood. According to the method, adaptive interaction and dynamic allocation of edge equipment computing resources are realized in a flood scene, real-time assimilation and dynamic correction are performed on the flood model through public source data of community interaction, and life and property safety of the aged facing flood extreme weather is guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of intelligent emergency management technology, and in particular to a generative rainstorm and flood intelligent risk avoidance dialogue method, system, equipment and medium. Background Art

[0002] The development of voice interaction technology offers new insights for improving the ability of elderly individuals to respond to disasters. In recent years, smart devices (such as smartphones and smart speakers) have been able to provide assistance and guidance through voice conversations. However, practical applications still face technical challenges: First, insufficient dialect recognition capabilities. Elderly individuals are highly dependent on dialects, and if the system cannot recognize their local dialect, it will be unable to understand distress messages. Second, the intermittent nature of speech makes it difficult for elderly individuals to communicate smoothly with edge devices. Third, the high noise levels and risk of communication interruptions in flood disaster scenarios require voice systems to be highly robust and adaptable.

[0003] Existing large models, such as ChatGPT (Chat Generative Pre-trained Transformer), support multi-round conversations, but their high-frequency invocation is extremely costly. During a rainstorm disaster, millions of interaction requests can be triggered daily, easily leading to a "burst" of computing resources. The Claude model's specialized training cycle on dialect data is too long, making it difficult to meet the emergency management standard of "system readiness within 30 minutes before a disaster." Therefore, handling high-concurrency interactions on edge devices with limited computing power is an urgent technical challenge.

[0004] Furthermore, traditional flood forecasting relies primarily on fixed monitoring stations, which have limited spatial coverage. Crowdsourcing experiments have demonstrated that public on-site observations can significantly fill the information gaps caused by sparse instrumentation, particularly in urban waterlogging and rural flash flooding scenarios. However, current intelligent interactive systems have yet to integrate the closed loop of "community interaction-data assimilation-dynamic correction" at the edge, resulting in a disconnect between forecasts and actual flood conditions, which urgently needs improvement. Summary of the Invention

[0005] The present invention provides a generative rainstorm and flood intelligent risk avoidance dialogue method, system, device and medium to solve the extremely high high-frequency call cost of existing large models. During rainstorm disasters, millions of interaction requests may be triggered daily, which easily causes an "explosion" of computing resources. The Claude model has a long special training cycle for dialect data, making it difficult to meet the emergency management requirement of "system readiness within 30 minutes before a disaster." Traditional flood forecasting mainly relies on fixed monitoring stations with limited spatial coverage. The current intelligent interactive system has not yet integrated the "community interaction-data assimilation-dynamic correction" closed loop into the edge end, resulting in a disconnect between predictions and actual flood conditions.

[0006] The first embodiment of the present invention provides a generative rainstorm and flood intelligent risk avoidance dialogue method, comprising the following steps: initializing the data structure of the prefix tree, and loading the expert network model under the dynamic hybrid expert architecture according to the operating state of the portable edge device; using the expert network model to obtain the user's voice stream data in real time according to the portable edge device when the user is in a flood disaster emergency interaction scenario, and preprocessing the voice stream data to extract voice features that meet preset clarity conditions; analyzing the voice features that meet the preset clarity conditions to predict and complete the interrupted voice content of the voice stream data to generate a complete voice sequence; uploading the crowdsource data of the user's area, and verifying the crowdsource data. The method assimilates the verified multi-source data into a preset flood forecast model to generate a corrected flood level and flow forecast, and generates flood risk information based on the corrected flood level and flow forecast; based on the flood risk information, decodes the dialect speech in the complete speech sequence into a standard text, and outputs the confidence of the standard text; based on the confidence of the standard text, asks the user to guide the user to clarify the ambiguous semantics, and generates user semantic information and environmental data that meet the preset high-credibility conditions; based on the user semantic information and the environmental data that meet the preset high-credibility conditions, generates a dynamic risk avoidance dialogue strategy for rainstorms and floods according to the flood and rainstorm warning level and the user's location coordinates.

[0007] Optionally, in one embodiment of the present invention, the loading of the expert network model under the dynamic hybrid expert architecture according to the operating status of the portable edge device includes: determining the dialect recognition expert model, semantic clarification expert model, risk avoidance decision expert model, device optimization expert model and data assimilation and alignment expert model in the flood disaster expert network model according to the operating status of the portable edge device; when the remaining power of the portable edge device is greater than a first preset threshold, loading the dialect recognition expert model, the semantic clarification expert model, the risk avoidance decision expert model, the device optimization expert model and the data assimilation and alignment expert model at the same time; when the remaining power is greater than a second preset threshold and less than or equal to the first preset threshold, loading the device optimization expert model; when the remaining power is less than or equal to the second preset threshold, loading the dialect recognition expert model, the semantic clarification expert model and the risk avoidance decision expert model.

[0008] Optionally, in one embodiment of the present invention, after obtaining the user's voice stream data in real time according to the portable edge device, it also includes: when it is detected that there are preset keywords in the voice stream data, pausing non-critical background tasks and increasing the weight of resource allocation to a preset percentage; when it is detected that the modifiers in the voice stream data are greater than a preset threshold, calling the prefix tree algorithm to analyze the syntax to generate the core semantic components of the sentence.

[0009] Optionally, in one embodiment of the present invention, the analysis of the speech features that meet the preset clarity conditions to predict and complete the interrupted speech content of the speech stream data to generate a complete speech sequence includes: when a long period of interruption or ambiguous phoneme fragment is detected in the speech stream data, using the prefix tree algorithm to trigger a node backtracking mechanism to count the phoneme path frequency of the speech features; backtracking and selecting a phoneme sequence path according to the phoneme path frequency to predict and complete the interrupted speech content according to the phoneme sequence path to generate the complete speech sequence.

[0010] Optionally, in one embodiment of the present invention, uploading the crowd-source data of the user's area and verifying the crowd-source data to assimilate the verified crowd-source data into a preset flood forecast model to generate a corrected flood water level and flow forecast, and generating flood risk information based on the corrected flood water level and flow forecast, includes: uploading the crowd-source data of the user's area and preprocessing the crowd-source data to generate preprocessed crowd-source data; cross-validating the preprocessed crowd-source data to generate the verified crowd-source data; assimilating the verified crowd-source data into the preset flood forecast model to generate the corrected flood water level and flow forecast data; identifying the magnitude of flood risk changes based on the corrected flood water level and flow forecast data, and adjusting the flood threshold and warning level according to the magnitude of flood risk changes to generate the flood risk information.

[0011] Optionally, in one embodiment of the present invention, the credibility calculation formula of the standard text is:

[0012] in, C Indicates the credibility value of the spatiotemporal alignment of multi-source data in the system, represents the time dimension weight, represents the spatial dimension weight, represents the semantic dimension weight, Indicates the credibility score corresponding to the time dimension, represents the credibility score corresponding to the spatial dimension, represents the credibility score corresponding to the semantic dimension, Represents the time dimension index, represents the spatial dimension index, Represents the semantic dimension index; The calculation formula for the credibility score corresponding to the time dimension is:

[0013] in, Indicates the absolute value difference between the voice interaction time and the sensor recording time, Indicates the maximum allowed time difference under the warning level; The calculation formula for the credibility score corresponding to the spatial dimension is:

[0014] in, Indicates the user's location and i The distance between monitoring points, Indicates the i The water level status of each monitoring point, Indicates the monitoring point index, Indicates the number of monitoring points; The calculation formula for the credibility score corresponding to the semantic dimension is:

[0015] in, n Indicates the number of semantic clarification questions asked.

[0016] The second embodiment of the present invention provides a generative rainstorm and flood intelligent risk avoidance dialogue system, including: a prefix tree initialization module, which is used to initialize the data structure of the prefix tree, and load the expert network model under the dynamic hybrid expert architecture according to the operating status of the portable edge device; a preprocessing module, which is used to use the expert network model to obtain the user's voice stream data in real time according to the portable edge device when the user is in a flood disaster emergency interaction scenario, and preprocess the voice stream data to extract voice features that meet preset clarity conditions; a prediction module, which is used to analyze the voice features that meet the preset clarity conditions to predict and complete the interrupted voice content of the voice stream data and generate a complete voice sequence; a crowdsource data uploading module, which is used to upload crowdsource data in the area where the user is located, and verify The multi-source data is verified to assimilate the verified multi-source data into a preset flood forecasting model to generate a corrected flood water level and flow forecast, and generate flood risk information based on the corrected flood water level and flow forecast; a decoding module is used to decode the dialect speech in the complete speech sequence into a standard text based on the flood risk information, and output the confidence of the standard text; an interaction module is used to ask the user questions based on the confidence of the standard text to guide the user to clarify ambiguous semantics and generate user semantic information and environmental data that meet preset high-credibility conditions; a risk avoidance dialogue module is used to generate a dynamic risk avoidance dialogue strategy for rainstorm and flooding based on the user semantic information and the environmental data that meet the preset high-credibility conditions according to the flood and rainstorm warning level and the user's location coordinates.

[0017] Optionally, in one embodiment of the present invention, the prefix tree initialization module includes: a determination unit, used to determine the dialect recognition expert model, semantic clarification expert model, risk avoidance decision expert model, device optimization expert model and data assimilation and alignment expert model in the flood disaster expert network model according to the operating status of the portable edge device; a first loading unit, used to simultaneously load the dialect recognition expert model, the semantic clarification expert model, the risk avoidance decision expert model, the device optimization expert model and the data assimilation and alignment expert model when the remaining power of the portable edge device is greater than a first preset threshold; a second loading unit, used to load the device optimization expert model when the remaining power is greater than a second preset threshold and less than or equal to the first preset threshold; a third loading unit, used to load the dialect recognition expert model, the semantic clarification expert model and the risk avoidance decision expert model when the remaining power is less than or equal to the second preset threshold.

[0018] Optionally, in one embodiment of the present invention, it also includes: a pause module for pausing non-critical background tasks and increasing the weight of resource allocation to a preset percentage after obtaining the user's voice stream data in real time according to the portable edge device, when preset keywords are detected in the voice stream data; a calling module for calling the prefix tree algorithm to analyze the syntax to generate the core semantic components of the sentence when it is detected that the modifiers in the voice stream data are greater than a preset threshold.

[0019] Optionally, in one embodiment of the present invention, the prediction module includes: a statistical unit, which is used to trigger a node backtracking mechanism using the prefix tree algorithm to count the phoneme path frequency of the speech feature when a long interruption or ambiguous phoneme segment is detected in the speech stream data; and a prediction unit, which is used to backtrack and select a phoneme sequence path according to the phoneme path frequency, so as to complete the interrupted speech content according to the prediction of the phoneme sequence path and generate the complete speech sequence.

[0020] Optionally, in one embodiment of the present invention, the crowd-source data uploading module includes: an uploading unit for uploading the crowd-source data of the user's area and preprocessing the crowd-source data to generate preprocessed crowd-source data; a cross-validation unit for cross-validating the preprocessed crowd-source data to generate the verified crowd-source data; an assimilation unit for assimilating the verified crowd-source data into the preset flood forecasting model to generate the corrected flood water level and flow forecast data; an adjustment unit for identifying the magnitude of flood risk changes based on the corrected flood water level and flow forecast data, and adjusting the flood threshold and warning level according to the magnitude of flood risk changes to generate the flood risk information.

[0021] Optionally, in one embodiment of the present invention, the credibility calculation formula of the standard text is:

[0022] in, C Indicates the credibility value of the spatiotemporal alignment of multi-source data in the system, represents the time dimension weight, represents the spatial dimension weight, represents the semantic dimension weight, Indicates the credibility score corresponding to the time dimension, represents the credibility score corresponding to the spatial dimension, represents the credibility score corresponding to the semantic dimension, Represents the time dimension index, represents the spatial dimension index, Represents the semantic dimension index; The calculation formula for the credibility score corresponding to the time dimension is:

[0023] in, Indicates the absolute value difference between the voice interaction time and the sensor recording time, Indicates the maximum allowed time difference under the warning level; The calculation formula for the credibility score corresponding to the spatial dimension is:

[0024] in, Indicates the user's location and i The distance between monitoring points, Indicates the i The water level status of each monitoring point, Indicates the monitoring point index, Indicates the number of monitoring points; The calculation formula for the credibility score corresponding to the semantic dimension is:

[0025] in, n Indicates the number of semantic clarification questions asked.

[0026] A third aspect of the present invention provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the generative rainstorm and flood intelligent risk avoidance dialogue method as described in the above embodiment.

[0027] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-mentioned generative rainstorm and flood intelligent risk avoidance dialogue method.

[0028] A fifth aspect of the present invention provides a computer program product, which stores a computer program that, when executed by a processor, implements the above-mentioned generative rainstorm and flood intelligent risk avoidance dialogue method.

[0029] In the embodiment of the present invention, the MoE architecture collaborates through multiple "expert networks," ensuring model capacity while only activating experts relevant to the current task, thereby improving computational efficiency. A flood emergency interactive system for the elderly is designed based on a dynamic MoE architecture, deploying expert networks for dialect recognition, semantic clarification, risk avoidance decision-making, device optimization, and cross-modal alignment on edge devices. The system dynamically calls on experts as needed during the interaction process: It balances local computing power and battery life through device optimization experts, switching to cloud-based collaborative computing when necessary to ensure real-time response under peak interaction volumes, and utilizing cross-modal alignment experts to fuse sensor data to verify the reliability of voice information, providing timely and accurate risk avoidance guidance for elderly users. This solves the problem of extremely high high-frequency call costs of existing large models. During rainstorm disasters, millions of interaction requests may be triggered every day, which can easily cause an "explosion" of computing resources. The Claude model's special training cycle for dialect data is too long, making it difficult to meet the emergency management requirement of "system readiness within 30 minutes before a disaster." Traditional flood forecasting mainly relies on fixed monitoring stations with limited spatial coverage. The current intelligent interactive system has not yet integrated the "community interaction-data assimilation-dynamic correction" closed loop into the edge, resulting in problems such as disconnection between predictions and actual flood conditions.

[0030] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a flow chart of a generative rainstorm and flood intelligent risk avoidance dialogue method provided according to an embodiment of the present invention; Figure 2 A network diagram of flood experts based on a dynamic MoE architecture according to an embodiment of the present invention; Figure 3 A schematic diagram of the overall structure of a system according to an embodiment of the present invention; Figure 4 is a data assimilation flow chart according to one embodiment of the present invention; Figure 5 is an emergency interaction flow chart according to an embodiment of the present invention; Figure 6 A schematic diagram of the structure of a generative rainstorm and flood intelligent risk avoidance dialogue system provided according to an embodiment of the present invention; Figure 7 A schematic structural diagram of an electronic device provided according to an embodiment of the present invention.

[0032] Among them, 10-generative rainstorm and flood intelligent risk avoidance dialogue system; 100-prefix tree initialization module, 200-preprocessing module, 300-prediction module, 400-crowd source data upload module, 500-decoding module, 600-interaction module, 700-risk avoidance dialogue module; 701-memory, 702-processor, 703-communication interface. DETAILED DESCRIPTION

[0033] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0034] The following describes, with reference to the accompanying drawings, a generative rainstorm and flood intelligent risk avoidance dialogue method, apparatus, device, and medium according to an embodiment of the present invention. The high cost of high-frequency invocation of existing large models mentioned in the background art above can trigger millions of interaction requests daily during rainstorm disasters, easily leading to an "explosion" of computing resources. The Claude model's long training cycle for dialect data makes it difficult to meet the emergency management standard of "system readiness within 30 minutes before a disaster." Traditional flood forecasting primarily relies on fixed monitoring stations with limited spatial coverage. Current intelligent interactive systems have not yet integrated the "community interaction-data assimilation-dynamic correction" closed loop into the edge, resulting in a disconnect between predictions and actual flood conditions. The present invention provides a generative rainstorm and flood intelligent risk avoidance dialogue method. In this method, a MoE architecture collaborates through multiple "expert networks" to ensure model capacity while only activating experts relevant to the current task, thereby improving computational efficiency. A flood emergency interactive system for the elderly is designed based on a dynamic MoE architecture, deploying expert networks for dialect recognition, semantic clarification, risk avoidance decision-making, device optimization, and cross-modal alignment on edge devices. The system dynamically calls on experts on demand during the interaction process. It uses device optimization experts to balance local computing power and battery life, switching to cloud-based collaborative computing when necessary to ensure real-time response under peak interaction volumes. It also leverages cross-modal alignment experts to fuse sensor data to verify the reliability of voice information, providing timely and accurate risk avoidance guidance for elderly users. This addresses the extremely high cost of high-frequency calls to existing large models. During a rainstorm disaster, millions of interaction requests may be triggered daily, easily causing an "explosion" of computing resources. The Claude model's specialized training cycle for dialect data is too long, making it difficult to meet the emergency management standard of "system readiness within 30 minutes before a disaster." Traditional flood forecasting primarily relies on fixed monitoring stations with limited spatial coverage. Current intelligent interactive systems have yet to integrate the "community interaction-data assimilation-dynamic correction" closed loop at the edge, resulting in a disconnect between predictions and actual flood conditions.

[0035] Specifically, Figure 1 A flow chart of a generative rainstorm and flood intelligent risk avoidance dialogue method provided by an embodiment of the present invention.

[0036] like Figure 1 As shown, the generative rainstorm and flood intelligent risk avoidance dialogue method includes the following steps: In step S101 , the data structure of the prefix tree is initialized, and the expert network model under the dynamic hybrid expert architecture is loaded according to the operating state of the portable edge device.

[0037] During actual implementation, the present invention can initialize the system and load the expert network. This system activates the flood emergency response interactive system, initializes the prefix tree data structure, which includes information such as the phoneme segments and their frequencies stored in the tree nodes, and loads the expert network model based on the dynamic MoE (Mixture-of-Experts) architecture based on the operating status of the portable edge device.

[0038] Optionally, in one embodiment of the present invention, an expert network model under a dynamic hybrid expert architecture is loaded according to the operating status of the portable edge device, including: determining the dialect recognition expert model, semantic clarification expert model, risk avoidance decision expert model, device optimization expert model and data assimilation and alignment expert model in the flood disaster expert network model according to the operating status of the portable edge device; when the remaining power of the portable edge device is greater than a first preset threshold, the dialect recognition expert model, semantic clarification expert model, risk avoidance decision expert model, device optimization expert model and data assimilation and alignment expert model are loaded at the same time; when the remaining power is greater than a second preset threshold and less than or equal to the first preset threshold, the device optimization expert model is loaded; when the remaining power is less than or equal to the second preset threshold, the dialect recognition expert model, semantic clarification expert model and risk avoidance decision expert model are loaded.

[0039] It is understandable that if Figure 2As shown, the flood disaster emergency interactive system with a dynamic MoE architecture in the embodiment of the present invention constructs five types of flood expert network models by realizing dynamic activation of MoE experts: dialect recognition expert model, semantic clarification expert model, risk avoidance decision-making expert model, equipment optimization expert model, and data assimilation and alignment expert model (including crowdsourcing data assimilation and multimodal spatiotemporal alignment). The dialect recognition expert model takes voice signals as input and outputs standard text decoded from the dialect. The semantic clarification expert model is responsible for implementing human-computer dialogue interaction, guiding users to clarify ambiguous semantics through proactive questioning. The risk avoidance decision-making expert model generates escape routes based on Beidou satellites, takes as input the flood and rainstorm warning level and the user's location coordinates, and outputs a dynamic risk avoidance dialogue strategy including evacuation directions. The device optimization expert model is used to resolve the contradiction between limited device resources and the computing power requirements of AI (artificial intelligence) in disaster scenarios, ensuring battery life and performance, and ensuring that local computing power can still support critical operations even in the event of sudden cloud computing tasks such as sudden outages. The data assimilation and alignment expert model converts multi-source observations such as water level photos and subjective water depth descriptions uploaded by community users into flow / water level inputs after quality assessment, and assimilates them into the hydrological model in real time. It is also used for spatiotemporal alignment of multi-source data and crisis credibility assessment, eliminating deviations caused by sensor data and voice descriptions. In this embodiment of the present invention, the first preset threshold can be 40%, and the second preset threshold can be 20%.

[0040] Specifically, the embodiment of the present invention can determine the dialect recognition expert model, semantic clarification expert model, risk avoidance decision expert model, device optimization expert model and data assimilation and alignment expert model in the flood disaster expert network model according to the operating status of the portable edge device, and adopt structured pruning to retain the key dialect phoneme recognition path for the above-mentioned dialect recognition expert model. Based on the dynamic MoE loading strategy, the expert network is triggered according to the remaining power of the device: when the power is greater than 40%, the dialect recognition expert model, semantic clarification expert model, risk avoidance decision expert model, device optimization expert model and data assimilation and alignment expert model are loaded at the same time; when the power is 20% less than or equal to 40%, other local networks except the device optimization expert are closed, and cloud-based collaborative reasoning is called; when the power is less than or equal to 20%, only the dialect recognition expert model, semantic clarification expert model and risk avoidance decision expert model are retained, among which the semantic clarification expert model enables semantic compression and only retains noun phrases and verbs.

[0041] In addition, the system can connect to portable edge devices (such as smart bracelets) through the communication unit. When it detects that the battery level of the main device is ≤10%, it automatically wakes up the bracelet and vibrates to remind the user to enter low-power mode and charge to ensure the continuous operation of the system. The overall structure diagram of the system is shown in the figure. Figure 3 shown.

[0042] In the embodiment of the present invention, the system completes the preparation of the core data structure and model through initialization, and dynamically configures the expert network according to the device status to provide guarantee for the subsequent flood emergency interaction of the elderly.

[0043] In step S102, the expert network model is used to obtain the user's voice stream data in real time based on the portable edge device when the user is in a flood disaster emergency interaction scenario, and preprocess the voice stream data to extract voice features that meet preset clarity conditions.

[0044] It is understandable that the user in the embodiment of the present invention may be an elderly user; and the voice feature that meets the preset clarity condition may be a clear voice feature.

[0045] In actual implementation, embodiments of the present invention can perform voice acquisition and preprocessing. When entering a flood disaster emergency interaction scenario, the system activates the voice acquisition module and uses the portable edge device's microphone to acquire real-time voice stream data from elderly users. To account for background interference, such as ambient noise, in the elderly user's voice signal, the system first performs information preprocessing, including noise reduction filtering and audio calibration, to extract clear voice features.

[0046] The embodiments of the present invention can realize adaptive interaction and dynamic allocation of edge device computing resources in flood scenarios, ensuring the safety of life and property of the elderly facing extreme weather such as floods.

[0047] It should be noted that the preset clear conditions can be set by those skilled in the art according to actual conditions and are not specifically limited here.

[0048] Optionally, in one embodiment of the present invention, after obtaining the user's voice stream data in real time based on the portable edge device, it also includes: when preset keywords are detected in the voice stream data, pausing non-critical background tasks and increasing the weight of resource allocation to a preset percentage; when modifiers are detected in the voice stream data greater than a preset threshold, calling the prefix tree algorithm to analyze the syntax to generate the core semantic components of the sentence.

[0049] It can be understood that the preset keywords in the embodiment of the present invention may be disaster keywords, the preset percentage may be 70%, and the preset threshold may be 30%.

[0050] During the actual implementation process, the device optimization expert model in the embodiment of the present invention can monitor the voice content in real time to determine the degree of emergency. When disaster keywords such as "flood", "rainstorm" and "submergence" are detected in the user's voice, the device optimization module immediately triggers resource reallocation, suspends non-critical background tasks, and increases the CPU (Central Processing Unit) computing resource allocation weight to 70% to ensure that relevant calculations during emergency interactions can be given priority. In addition, if it is detected that the user's sentence contains more than 30% uncertainty modifiers, such as the frequent appearance of words such as "probably" and "possibly", indicating that there may be obstacles in the user's description, the device optimization model will automatically call the prefix tree algorithm for syntactic analysis to extract the core semantic components of the sentence.

[0051] In the embodiment of the present invention, the system can obtain the pre-processed speech signal, providing conditions for subsequent speech recognition and analysis.

[0052] It should be noted that the preset keywords, preset percentages and preset thresholds can be set by those skilled in the art according to actual conditions and are not specifically limited here.

[0053] In step S103, speech features that meet preset clarity conditions are analyzed to predict the interrupted speech content of the completed speech stream data and generate a complete speech sequence.

[0054] Among them, the embodiment of the present invention can perform discontinuous speech reconstruction and real-time analysis, input the pre-processed clear speech features into the prefix tree algorithm module for real-time analysis, thereby predicting and completing the interrupted speech content of the speech stream data and generating a complete speech sequence.

[0055] The present invention provides a prefix tree algorithm for edge devices to reconstruct speech streams based on the discontinuous speech characteristics of the elderly. Specifically, it includes: Prefix tree initialization module: Initializes the prefix tree structure, including defining tree node storage units. Each node contains the current phoneme segment, frequency statistics, and a set of child nodes.

[0056] Voice collection and preprocessing module: The module collects voice stream data of elderly users in real time through edge devices, extracts intermittent voice signals, and performs signal preprocessing to remove environmental noise and device background interference and extract effective phoneme features.

[0057] Real-time speech matching module: Based on the preprocessed phoneme features, it gradually performs comparison and matching through the prefix tree, traversing layer by layer from the root node, and updating the statistical frequency of the phoneme nodes in real time to reflect the frequency of phoneme usage.

[0058] Speech Interruption Prediction Module: This module addresses the slow speech speed and frequent speech interruptions typical of the elderly. When it detects that the speech interruption time exceeds a set threshold (the threshold is set to 3 seconds) or an ambiguous phoneme recognition result occurs, it automatically triggers the prefix tree node backtracking mechanism. The backtracking path selects the phoneme path with the highest frequency to achieve speech content prediction and completion.

[0059] Dialect Recognition and Calibration Module: The dialect recognition expert module is used to identify the predicted completed speech sequence and verify the accuracy of the completion. When the recognition confidence is lower than the preset threshold (the confidence is set to 80%), the cross-modal alignment expert module is automatically called to perform secondary calibration in combination with environmental sensor data.

[0060] Interactive response module: The complete speech stream that has been calibrated and confirmed is handed over to the semantic clarification expert module for subsequent interactive response, and the risk avoidance decision module is driven to formulate specific emergency strategies based on the response results.

[0061] Through the above modular design, the present invention realizes efficient and real-time voice stream reconstruction for the intermittent and fuzzy voice characteristics of the elderly group in flood disaster scenarios, thereby improving the reliability and effectiveness of the disaster emergency interaction system.

[0062] It should be noted that the preset clear conditions can be set by those skilled in the art according to actual conditions and are not specifically limited here.

[0063] Optionally, in one embodiment of the present invention, speech features that meet preset clarity conditions are analyzed to predict and complete the interrupted speech content of the speech stream data to generate a complete speech sequence, including: when a long period of interruption or ambiguous phoneme fragment is detected in the speech stream data, a prefix tree algorithm is used to trigger a node backtracking mechanism to count the phoneme path frequency of the speech feature; a phoneme sequence path is backtracked and selected according to the phoneme path frequency to predict and complete the interrupted speech content according to the phoneme sequence path to generate a complete speech sequence.

[0064] During the actual implementation process, after the prefix tree is initialized, the embodiment of the present invention gradually maps the phoneme stream to the prefix tree structure, matches the phoneme segments layer by layer starting from the root node, and dynamically updates the frequency statistics of the response nodes. Through the real-time voice matching process, the pauses and repetitions in the elderly's voice can be effectively captured. When a long time interruption (such as the interruption time exceeds the preset threshold of 3s) or an ambiguous and difficult-to-distinguish phoneme segment is detected in the voice, the prefix tree algorithm will trigger the node backtracking mechanism. According to the accumulated phoneme path frequency statistics, the most likely phoneme sequence path is backtracked to predict and complete the voice content interrupted by the user, thereby reconstructing the complete voice sequence.

[0065] In the embodiment of the present invention, the system can compensate for the information loss caused by the elderly's slow speaking speed or incoherent speech, thereby generating a relatively complete and coherent speech signal.

[0066] In step S104, the crowd-source data of the user's area is uploaded and verified to assimilate the verified crowd-source data into a preset flood forecasting model to generate a corrected flood level and flow forecast, and generate flood risk information based on the corrected flood level and flow forecast.

[0067] It is understandable that the embodiments of the present invention can introduce a community-interactive crowdsourced data upload module to give full play to the role of public participation in rainstorm and flood warning. This module is designed for elderly users and provides a simple interface and dialogue interaction, enabling them to timely upload disaster information in their area, including on-site photos, subjective estimates of hydrological depth, text descriptions and other data. The system can accept these real-time information from the community and provide a supplementary data source for flood forecasting, especially in areas where official observations are scarce. The present invention establishes community-participated data collection and assimilation processing as one of the core technical paths for accurate flood warning and intelligent risk avoidance decision-making.

[0068] During the actual implementation process, the embodiment of the present invention can upload crowd-source data in the user's area and verify the crowd-source data to assimilate the verified crowd-source data into a preset flood forecasting model, generate corrected flood water level and flow forecasts, and generate flood risk information based on the corrected flood water level and flow forecasts.

[0069] The embodiments of the present invention can assimilate and dynamically correct flood models in real time through crowd-sourced data from community interactions, further protecting the lives and property of the elderly facing extreme flood weather.

[0070] Optionally, in one embodiment of the present invention, crowd-source data in the user's area is uploaded and verified to assimilate the verified crowd-source data into a preset flood forecast model to generate a corrected flood level and flow forecast, and flood risk information is generated based on the corrected flood level and flow forecast, including: uploading crowd-source data in the user's area and preprocessing the crowd-source data to generate preprocessed crowd-source data; cross-validating the preprocessed crowd-source data to generate verified crowd-source data; assimilating the verified crowd-source data into a preset flood forecast model to generate corrected flood level and flow forecast data; identifying the magnitude of flood risk changes based on the corrected flood level and flow forecast data, and adjusting the flood threshold and warning level according to the magnitude of flood risk changes to generate flood risk information.

[0071] In the actual implementation process, Figure 4As shown, the embodiment of the present invention can report and receive data: elderly users can report the real-time flood situation at their location through the voice, text or photo function of the dialogue terminal. Users can describe "water over the calf" in text or upload photos of the water on site. The system's community interaction module receives the information and automatically records data such as the upload time and user location.

[0072] Furthermore, embodiments of the present invention can perform data preprocessing and quality assessment: The system preprocesses and performs quality checks on received user data. For text descriptions, it parses and extracts key information such as qualitative water level height (e.g., "calf" corresponds to approximately 0.5 meters). For photos, it uses built-in image analysis algorithms to estimate the water depth or affected area reflected in the image. The system then cross-validates data from different sources. If multiple users report water levels in adjacent areas, this information is mutually verified to enhance credibility. If a piece of data significantly deviates from common sense or surrounding data, it is flagged as an anomaly and ignored.

[0073] Furthermore, embodiments of the present invention can perform data conversion and assimilation: For observation data that has passed quality control, the system performs format conversion and unit conversion as needed. The data assimilation module of the flood forecasting model, based on the Ensemble Kalman Filter (EnKF) algorithm, is activated. This module includes a prediction phase and an assimilation phase, integrating the latest observation data into the model state. Through iterative calculations, it generates corrected flood level and flow forecasts, significantly improving forecast accuracy.

[0074] (1) Prediction stage: Use the model state evolution equation to advance the analysis state set of the previous moment to the current moment to obtain the predicted set state. The state evolution equation can be expressed as:

[0075] in, express The predicted state vector at time , f (•) represents the model evolution operator, represents process noise, characterizes model error, represents the analysis status of k-1 step (previous step), Indicates that this quantity is the predicted value obtained by model evolution before assimilation. is the time step index (indicating the kth moment), Indicates that the quantity is the optimal estimate obtained by fusing observations.

[0076] The corresponding model observation equation is:

[0077] in, express The observed state vector at time , is the observation matrix, which is used to map the model state to the observation space (such as water level, flow), represents the observation noise, Represents the predicted state vector at time .

[0078] (2) Assimilation stage: When the Actual observation at the moment Finally, the predicted set state is corrected through the Kalman filter update formula.

[0079] First, calculate the sample covariance matrix based on the forecast set:

[0080] in, represents the sample covariance matrix, N Indicates the number of members in the set. represents the average value of the ensemble prediction state, represents the predicted state vector of the i-th member at time k, Indicates that this quantity is the predicted value obtained by model evolution before assimilation. is the time step index, , z is the index of the set member.

[0081] Next, calculate the Kalman gain:

[0082] in, R represents the observation error covariance matrix, represents the Kalman gain, represents the sample covariance matrix, represents the transpose of H (i.e., projecting the observation space information back to the model space in the calculation), is the observation matrix.

[0083] Finally, the Kalman gain is used to integrate the observation information into each ensemble member, and the update formula is: i =1,… N in, represents the updated analysis state vector at time k after the assimilation observation, and the i-th set member corresponds to a set of independent updates, represents the predicted state vector obtained by the same member at time k before the assimilation prediction based solely on model evolution, represents the Kalman gain matrix, express Actual observation of the moment, is the observation matrix, is the time step index, The index of the collection member.

[0084] This gives the time k The assimilation update process minimizes the deviation between the model's predicted values ​​and the measured values, thereby correcting the model's state. The updated set of states serves as the initial condition for the next forecast time, and the prediction-assimilation process is repeated iteratively. Through this data assimilation module, the flood forecasting system can promptly incorporate actual observation information, making real-time corrections to predicted water levels and flows, effectively improving forecast accuracy.

[0085] Furthermore, embodiments of the present invention enable forecast correction and decision adjustments: updated model forecast results are fed back to the risk avoidance decision-making expert model in real time, enabling automatic updates of warning levels through a dynamic threshold adjustment algorithm. Whenever the hydrological model generates a new water level or risk forecast through data assimilation, the system's decision-making module instantly compares the old and new forecasts, identifies the magnitude of the flood risk change, and adjusts the thresholds and alert levels accordingly. If the assimilated model indicates that the water level in a particular area is rising faster than originally predicted and is projected to exceed a preset safety threshold, the system will raise the alert level for that area, enabling more timely evacuation or risk avoidance measures. Conversely, if the threshold is not exceeded, the current alert level is maintained and monitoring continues. For areas no longer at risk, the alert level is lowered and prompts adjusted accordingly. The risk avoidance decision-making expert model also updates its user dialogue accordingly, ensuring that elderly users receive guidance based on the latest context. The entire threshold update process is completed by the system in real time, requiring no human intervention. With the help of an intelligent edge computing architecture, the above model updates and threshold adjustments can be run efficiently at edge nodes close to the site, minimizing cloud transmission delays. Even if the central network is damaged in a disaster, local devices can autonomously reassess thresholds based on sensor data to ensure the continuity and reliability of early warning services.

[0086] Furthermore, embodiments of the present invention enable information feedback and a cyclical improvement process: the system promptly provides updated flood risk information and mitigation plans to users via a dialogue interface, and based on user feedback, determines whether to issue further prompts. Simultaneously, all new user observation data and corrected model states are stored in the system database for subsequent analysis and model improvement. As more community data is uploaded, the aforementioned quality control, assimilation, and correction process is repeated, gradually optimizing flood forecast accuracy and forming a virtuous cycle combining public participation with intelligent early warning.

[0087] In step S105 , based on the flood risk information, the dialect speech in the complete speech sequence is decoded into standard text, and the confidence of the standard text is output.

[0088] During the actual implementation process, the embodiment of the present invention can perform dialect recognition and multi-source data alignment calibration. For the reconstructed speech sequence obtained in the above steps, the dialect recognition expert model is called to perform speech recognition or text writing, the dialect speech is decoded into standard text, and the recognition confidence is output. When the recognition confidence result is higher than the preset threshold (80%), it means that the speech content is reliable, and the next step of semantic clarification interaction is directly entered; when the recognition confidence structure is lower than the preset threshold (80%), the speech content may be ambiguous or erroneous, and the cross-modal alignment expert module will be automatically activated for secondary calibration. The cross-modal alignment expert will obtain multi-source sensor data of the current environment, such as water level sensor data at the disaster site, the geographical location coordinates of the user, the flood warning level of the corresponding area, etc., and perform spatiotemporal consistency verification on the speech information.

[0089] This invention provides a method for spatiotemporal alignment and crisis credibility assessment of multi-source data. The credibility assessment includes temporal credibility, spatial credibility, and semantic credibility. Temporal credibility is used to calibrate timing consistency. When there may be a time difference between the user's voice interaction with the system and the data from the water level sensor, it is used to verify whether the two belong to the same event cycle. Spatial credibility is used to calibrate the mapping relationship between geographic locations. By verifying the spatial relationship between the location described by the voice interactor and the sensor monitoring point, it can avoid cross-regional forecast misjudgments. Semantic credibility is calculated based on the number of dialogue turns and semantic entropy to clarify any ambiguous semantics that may be expressed by the user.

[0090] In one embodiment of the present invention, the cross-modal alignment expert performs weighted fusion of temporal, spatial and semantic credibility according to preset weights to calculate the credibility value of multi-source data. C。 The formula for calculating the credibility of the standard text is:

[0091] in, C Indicates the credibility value of the spatiotemporal alignment of multi-source data in the system, represents the time dimension weight, represents the spatial dimension weight, represents the semantic dimension weight, Indicates the credibility score corresponding to the time dimension, represents the credibility score corresponding to the spatial dimension, represents the credibility score corresponding to the semantic dimension, Represents the time dimension index, represents the spatial dimension index, Represents the semantic dimension index; First, calculate the time credibility. Based on the difference between the voice interaction time and the sensor recording time, determine whether the two are in the same event cycle. The calculation formula for the credibility score corresponding to the time dimension is:

[0092] in, Indicates the credibility score corresponding to the time dimension, Indicates the absolute value difference between the voice interaction time and the sensor recording time, Indicates the maximum allowed time difference under the warning level; Then, the spatial credibility is calculated. Based on the distance between the user's location and each monitoring point and the abnormal state of the corresponding water level, the degree of consistency between the location described by the user and the data monitored by the sensor is evaluated. The closer the user is to the sensor monitoring point, the higher the credibility, and vice versa. The calculation formula for the credibility score corresponding to the spatial dimension is:

[0093] in, represents the credibility score corresponding to the spatial dimension, Indicates the user's location and i The distance between monitoring points, Indicates the i The water level status of each monitoring point (0 means normal, 1 means abnormal), Indicates the number of monitoring points, Indicates the monitoring point index; Finally, the semantic credibility is calculated based on the conversation rounds to quantify the clarity of the user's description of the semantics and the number of clarification questions. n The more or the higher the semantic uncertainty, the lower the semantic credibility. The calculation formula for the credibility score corresponding to the semantic dimension is:

[0094] in, represents the credibility score corresponding to the semantic dimension, n Indicates the number of semantic clarification questions. To avoid fatigue, n ≤5.

[0095] The mathematical constraints for the weight control of spatiotemporal credibility are:

[0096] To improve the accuracy of credibility assessment, the system dynamically adjusts the weight of each dimension based on the environment, including: (1) When the user positioning signal is weak (positioning error exceeds 50 meters), the spatial dimension weight is reduced to less than 20%, and the importance of the time dimension is enhanced to compensate for positioning uncertainty with time correlation; (2) When the external sensor data update delay is large (delay exceeds 10 minutes), the time dimension weight is reduced to less than 20%, and the spatial dimension weight is increased instead, compensating for the lack of timeliness of sensor data through spatial correlation; (3) When the current meteorological warning level is red (highest alert), in order to avoid over-reliance on semantic interaction and missed opportunities, the weight of time, space, and semantic dimensions can be balanced to 0.2, that is, more emphasis is placed on directly using regional monitoring data to assist decision-making, reducing reliance on user semantic confirmation; (4) When it is detected that the user's voice has obvious ambiguity (the speaking speed exceeds 180 words / minute or the voice interruption exceeds 3 seconds), the weight of the semantic dimension is increased to 0.4, and the overall credibility assessment is improved by strengthening semantic analysis first.

[0097] After the above multi-source data alignment and credibility calculation, the system obtains a comprehensive credibility value C , and adjust the emergency strategy accordingly: If C If the value is higher (e.g. not less than 0.8), the system considers that the user's voice matches the environmental conditions well and can directly formulate an emergency plan based on this information. C If it is too low, the system will rely more on objective sensor data or adopt a conservative strategy when generating risk avoidance strategies, and if necessary, proceed to the next step to interact with the user to obtain more information.

[0098] In step S106 , based on the confidence of the standard text, questions are asked to the user to guide the user to clarify the fuzzy semantics, and generate user semantic information and environmental data that meet the preset high confidence conditions.

[0099] It can be understood that, in the embodiment of the present invention, the user semantic information and environmental data that meet the preset high credibility condition may be highly credible user semantic information and environmental data.

[0100] In actual implementation, embodiments of the present invention can perform semantic clarification interactions and enhance semantic credibility. After determining the approximate textual representation of the speech content, the system enters the semantic clarification interaction phase. Based on the confidence level of the standard text, the semantic clarification expert model conducts multiple rounds of dialogue with the user to guide the user in clarifying ambiguous semantics and generate user semantic information and environmental data that meet preset high-confidence criteria. This eliminates potential ambiguity and ambiguous semantics, further improving semantic credibility. The system proactively asks questions regarding ambiguities in the content identified in the previous phase, guiding the user to clarify key information, such as location details, level of hazard, or special needs. After each round of clarification inquiry, the system records the number of dialogue turns n and calculates the current semantic entropy, comprehensively assessing the changing trend of semantic credibility. After several rounds of questioning and answering (excluding fatigue considerations, no more than five rounds), the user's intentions and descriptions become clearer, the semantic entropy significantly decreases, and the semantic credibility correspondingly improves. The semantic clarification phase ends when the preset upper limit of dialogue turns (e.g., five rounds) is reached or the system determines that the semantics are sufficiently clear. Through this interactive process, the system ensures that it understands the user's intentions and on-site conditions accurately to the greatest extent possible, laying a semantic foundation for generating the optimal emergency risk avoidance strategy.

[0101] It should be noted that the preset high reliability condition can be set by those skilled in the art according to actual conditions and is not specifically limited here.

[0102] In step S107 , based on the user semantic information and environmental data that meet the preset high credibility conditions, a dynamic risk avoidance dialogue strategy for heavy rain and floods is generated according to the flood and heavy rain warning level and the user's location coordinates.

[0103] Specifically, embodiments of the present invention enable risk avoidance decision-making and result output. After obtaining highly reliable user semantic information and environmental data, the system generates a specific flood disaster emergency risk avoidance dialogue strategy using a risk avoidance decision-making expert model. The risk avoidance decision-making expert model combines authoritative warning information with the user's current geographic location and utilizes Beidou satellite positioning data to plan a safe escape route. Based on this integrated multi-source information, a dynamic risk avoidance plan is developed, including evacuation directions, route selection, and possible safe havens. For example, if the user's area is monitored to have continuously rising water levels and a high warning level, the user can immediately move to higher ground along a predetermined route, with simultaneous voice or on-screen guidance on evacuation directions. If the credibility assessment indicates a discrepancy between the user's reported risk and the sensor data, the system will adopt a conservative strategy, such as reminding the user to remain vigilant and seek a nearby safe area pending further confirmation. During the decision-making process, the device optimization expert model continues to ensure high-performance and low-energy system operation, and if necessary, utilizes cloud computing resources to obtain more refined route planning. Finally, the generated emergency risk avoidance strategy is fed back to the user via speech synthesis or terminal display.

[0104] Through the above steps, the system of the present invention recognizes and reconstructs the dialect and intermittent speech features of elderly users, integrates multi-source environmental data to calibrate the credibility of information, optimizes equipment resources in real time and outputs personalized risk avoidance decisions, realizing full-process intelligent interaction from voice input to emergency strategy output, and improving the reliability and timeliness of human-computer interaction in emergency situations of flood disasters.

[0105] Specifically, it can be combined Figure 5 As shown, the working principle of the generative rainstorm and flood intelligent risk avoidance dialogue method in the embodiment of the present invention is described in detail with a specific embodiment.

[0106] like Figure 5 As shown, the embodiment of the present invention may include the following steps: Step S501: Voice input from an elderly user.

[0107] Step S502: Dialect recognition.

[0108] Step S503: Crowdsourcing data assimilation.

[0109] Step S504: semantic clarification.

[0110] Step S505: Crowd-source data credibility assessment (time, space, semantics).

[0111] Step S506: Generate risk hedging decision.

[0112] According to the generative intelligent risk avoidance method for rainstorms and floods proposed in an embodiment of the present invention, the MoE architecture works collaboratively through multiple "expert networks" to ensure model capacity while only activating experts relevant to the current task, thereby improving computing efficiency. A flood emergency interaction system for the elderly is designed based on a dynamic MoE architecture, that is, expert networks such as dialect recognition, semantic clarification, risk avoidance decision-making, device optimization, and cross-modal alignment are deployed on edge devices. The system dynamically calls experts on demand during the interaction process: it can balance local computing power and battery life through device optimization experts, switch to cloud-based collaborative computing when necessary, and ensure real-time response under peak interaction volume. It can also use cross-modal alignment experts to fuse sensor data to verify the reliability of voice information, providing timely and accurate risk avoidance guidance for elderly users. This solves the problem of extremely high high-frequency call costs of existing large models. During rainstorm disasters, millions of interaction requests may be triggered every day, which can easily cause an "explosion" of computing resources. The Claude model's special training cycle for dialect data is too long, making it difficult to meet the emergency management requirement of "system readiness within 30 minutes before a disaster." Traditional flood forecasting mainly relies on fixed monitoring stations with limited spatial coverage. The current intelligent interactive system has not yet integrated the "community interaction-data assimilation-dynamic correction" closed loop into the edge, resulting in a disconnect between predictions and actual flood conditions.

[0113] Next, the generative rainstorm and flood intelligent risk avoidance dialogue system proposed in an embodiment of the present invention will be described with reference to the accompanying drawings.

[0114] Figure 6 It is a structural diagram of the generative rainstorm and flood intelligent risk avoidance dialogue system according to an embodiment of the present invention.

[0115] like Figure 6 As shown, the generative rainstorm and flood intelligent risk avoidance dialogue system 10 includes: a prefix tree initialization module 100, a preprocessing module 200, a prediction module 300, a crowd-source data uploading module 400, a decoding module 500, an interaction module 600 and a risk avoidance dialogue module 700.

[0116] Specifically, the prefix tree initialization module 100 is used to initialize the data structure of the prefix tree and load the expert network model under the dynamic hybrid expert architecture according to the operating status of the portable edge device.

[0117] The preprocessing module 200 is used to use the expert network model to obtain the user's voice stream data in real time based on the portable edge device when the user is in a flood disaster emergency interaction scenario, and preprocess the voice stream data to extract voice features that meet preset clarity conditions.

[0118] The prediction module 300 is used to analyze speech features that meet preset clarity conditions to predict the interrupted speech content of the completed speech stream data and generate a complete speech sequence.

[0119] The crowd-source data uploading module 400 is used to upload crowd-source data in the user's area and verify the crowd-source data, so as to assimilate the verified crowd-source data into a preset flood forecasting model, generate a corrected flood water level and flow forecast, and generate flood risk information based on the corrected flood water level and flow forecast.

[0120] The decoding module 500 is used to decode the dialect speech in the complete speech sequence into standard text based on the flood risk information, and output the confidence level of the standard text.

[0121] The interactive module 600 is used to ask questions to the user based on the confidence of the standard text to guide the user to clarify the fuzzy semantics and generate user semantic information and environmental data that meet the preset high confidence conditions.

[0122] The risk avoidance dialogue module 700 is used to generate a dynamic risk avoidance dialogue strategy for rainstorms and floods based on user semantic information and environmental data that meet preset high-credibility conditions, according to the flood and rainstorm warning level and the user's location coordinates.

[0123] Optionally, in one embodiment of the present invention, the prefix tree initialization module 100 includes: a determination unit, a first loading unit, a second loading unit, and a third loading unit.

[0124] Among them, the determination unit is used to determine the dialect recognition expert model, semantic clarification expert model, risk avoidance decision-making expert model, equipment optimization expert model and data assimilation and alignment expert model in the flood disaster expert network model according to the operating status of the portable edge device.

[0125] The first loading unit is used to simultaneously load the dialect recognition expert model, the semantic clarification expert model, the risk avoidance decision expert model, the device optimization expert model and the data assimilation and alignment expert model when the remaining power of the portable edge device is greater than a first preset threshold.

[0126] The second loading unit is used to load the device optimization expert model when the remaining power is greater than the second preset threshold and less than or equal to the first preset threshold.

[0127] The third loading unit is used to load the dialect recognition expert model, the semantic clarification expert model and the risk avoidance decision expert model when the remaining power is less than or equal to the second preset threshold.

[0128] Optionally, in one embodiment of the present invention, the generative rainstorm and flood intelligent risk avoidance dialogue system 10 further includes: a pause module and a call module.

[0129] Among them, the pause module is used to pause non-critical background tasks and increase the weight of resource allocation to a preset percentage after obtaining the user's voice stream data in real time based on the portable edge device, if preset keywords are detected in the voice stream data.

[0130] The calling module is used to call the prefix tree algorithm to analyze the syntax when it is detected that the modifier in the voice stream data is greater than a preset threshold, so as to generate the core semantic components of the sentence.

[0131] Optionally, in one embodiment of the present invention, the prediction module 300 includes: a statistics unit and a prediction unit.

[0132] Among them, the statistical unit is used to trigger the node backtracking mechanism using the prefix tree algorithm to count the phoneme path frequency of the speech feature when a long interruption or ambiguous phoneme segment is detected in the voice stream data.

[0133] The prediction unit is used to backtrack and select a phoneme sequence path according to the phoneme path frequency, so as to predict and complete the interrupted speech content according to the phoneme sequence path and generate a complete speech sequence.

[0134] Optionally, in one embodiment of the present invention, the crowd-source data uploading module 400 includes: an uploading unit, a cross-validation unit, an assimilation unit, and an adjustment unit.

[0135] Among them, the uploading unit is used to upload the crowd-source data in the user's area and pre-process the crowd-source data to generate pre-processed crowd-source data.

[0136] The cross-validation unit is used to cross-validate the preprocessed crowd-source data to generate verified crowd-source data.

[0137] The assimilation unit is used to assimilate the verified multi-source data into the preset flood forecast model to generate corrected flood water level and flow forecast data.

[0138] The adjustment unit is used to identify the magnitude of flood risk changes based on the corrected flood water level and flow forecast data, and adjust the flood threshold and warning level according to the magnitude of flood risk changes to generate flood risk information.

[0139] Optionally, in one embodiment of the present invention, the formula for calculating the credibility of the standard text is:

[0140] in, C Indicates the credibility value of the spatiotemporal alignment of multi-source data in the system, represents the time dimension weight, represents the spatial dimension weight, represents the semantic dimension weight, Indicates the credibility score corresponding to the time dimension, represents the credibility score corresponding to the spatial dimension, represents the credibility score corresponding to the semantic dimension, Represents the time dimension index, represents the spatial dimension index, Represents the semantic dimension index; The calculation formula for the credibility score corresponding to the time dimension is:

[0141] in, Indicates the absolute value difference between the voice interaction time and the sensor recording time, Indicates the maximum allowed time difference under the warning level; The calculation formula for the credibility score corresponding to the spatial dimension is:

[0142] in, Indicates the user's location and i The distance between monitoring points, Indicates the i The water level status of each monitoring point, Indicates the monitoring point index, Indicates the number of monitoring points; The calculation formula for the credibility score corresponding to the semantic dimension is:

[0143] in, n Indicates the number of semantic clarification questions asked.

[0144] It should be noted that the aforementioned explanation of the embodiment of the generative rainstorm and flood intelligent risk avoidance dialogue method is also applicable to the generative rainstorm and flood intelligent risk avoidance dialogue system of this embodiment, and will not be repeated here.

[0145] According to the generative rainstorm and flood intelligent risk avoidance dialogue system proposed in an embodiment of the present invention, the MoE architecture works collaboratively through multiple "expert networks". While ensuring model capacity, it only activates experts related to the current task, thereby improving computing efficiency. Based on the dynamic MoE architecture, a flood emergency interaction system for the elderly is designed, that is, expert networks such as dialect recognition, semantic clarification, risk avoidance decision-making, device optimization, and cross-modal alignment are deployed on edge devices. During the interaction process, the system dynamically calls experts on demand: it can balance local computing power and battery life through device optimization experts, switch to cloud-based collaborative computing when necessary, and ensure real-time response under peak interaction volume. It can also use cross-modal alignment experts to fuse sensor data to verify the reliability of voice information, providing timely and accurate risk avoidance guidance for elderly users. This solves the problem of extremely high high-frequency call costs of existing large models. During rainstorm disasters, millions of interaction requests may be triggered every day, which can easily cause an "explosion" of computing resources. The Claude model's special training cycle for dialect data is too long, making it difficult to meet the emergency management requirement of "system readiness within 30 minutes before a disaster." Traditional flood forecasting mainly relies on fixed monitoring stations with limited spatial coverage. The current intelligent interactive system has not yet integrated the "community interaction-data assimilation-dynamic correction" closed loop into the edge, resulting in a disconnect between predictions and actual flood conditions.

[0146] Figure 7 This is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. The electronic device may include: Memory 701 , processor 702 , and computer programs stored in the memory 701 and executable on the processor 702 .

[0147] When the processor 702 executes the program, the generative rainstorm and flood intelligent risk avoidance dialogue method provided in the above embodiment is implemented.

[0148] Furthermore, the electronic device further includes: The communication interface 703 is used for communication between the memory 701 and the processor 702 .

[0149] The memory 701 is used to store computer programs that can be run on the processor 702 .

[0150] The memory 701 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0151] If the memory 701, processor 702, and communication interface 703 are implemented independently, the communication interface 703, memory 701, and processor 702 can be connected to each other via a bus and communicate with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus. Buses can be divided into address buses, data buses, control buses, etc. For ease of representation, Figure 7 Only one thick line is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0152] Optionally, in a specific implementation, if the memory 701, the processor 702 and the communication interface 703 are integrated on a chip, the memory 701, the processor 702 and the communication interface 703 can communicate with each other through an internal interface.

[0153] The processor 702 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention.

[0154] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned generative rainstorm and flood intelligent risk avoidance dialogue method.

[0155] An embodiment of the present invention also provides a computer program product, which stores a computer program. When the program is executed by a processor, it implements the above-mentioned generative rainstorm and flood intelligent risk avoidance dialogue method.

[0156] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples without contradiction.

[0157] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Thus, a feature specified as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "N" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0158] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or N executable instructions for implementing a custom logical function or step of a process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may be performed out of the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention pertain.

[0159] The logic and / or steps represented in a flowchart or otherwise described herein, for example, can be considered a sequenced list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" is any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (not exhaustive) of computer-readable media include: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and programmable read-only memory (EPROM or flash memory), fiber optic devices, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program can be obtained electronically by optically scanning the paper or other medium and then editing, interpreting or processing it in other suitable ways as necessary, and then storing it in a computer memory.

[0160] It should be understood that various components of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiment, the N steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one or combination of the following technologies known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits having suitable combinational logic gate circuits, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0161] Those skilled in the art will understand that all or part of the steps in the method of the above embodiment can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiment.

[0162] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing module, or each unit may exist physically separately, or two or more units may be integrated into a single module. The aforementioned integrated modules may be implemented in the form of hardware or in the form of software functional modules. If the integrated modules are implemented in the form of software functional modules and sold or used as independent products, they may also be stored in a computer-readable storage medium.

[0163] The storage medium mentioned above may be a read-only memory, a magnetic disk, or an optical disk, etc. Although the embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and are not to be construed as limiting the present invention. Persons skilled in the art may make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A generative rainstorm and flood intelligent risk avoidance dialogue method, characterized by: The following steps are involved: Initialize the data structure of the prefix tree and load the expert network model under the dynamic hybrid expert architecture according to the operating status of the portable edge device; Using the expert network model, when the user is in a flood disaster emergency interaction scenario, the portable edge device obtains the user's voice stream data in real time, and pre-processes the voice stream data to extract voice features that meet preset clarity conditions; Analyzing the speech features that meet the preset clarity conditions to predict and complete the interrupted speech content of the speech stream data to generate a complete speech sequence; Uploading crowd-source data in the user's area and verifying the crowd-source data to assimilate the verified crowd-source data into a preset flood forecast model to generate a corrected flood level and flow forecast, and generating flood risk information based on the corrected flood level and flow forecast; Based on the flood risk information, decoding the dialect speech in the complete speech sequence into standard text, and outputting the confidence level of the standard text; Based on the confidence level of the standard text, asking the user questions to guide the user to clarify ambiguous semantics and generate user semantic information and environmental data that meet preset high-confidence conditions; Based on the user semantic information that meets the preset high credibility condition and the environmental data, a dynamic risk avoidance dialogue strategy for heavy rain and floods is generated according to the flood and heavy rain warning level and the user's location coordinates.

2. The generative rainstorm and flood intelligent risk avoidance dialogue method according to claim 1 is characterized in that: The method of loading the expert network model under the dynamic hybrid expert architecture according to the operating state of the portable edge device includes: Determining, according to the operating status of the portable edge device, a dialect recognition expert model, a semantic clarification expert model, a risk avoidance decision expert model, a device optimization expert model, and a data assimilation and alignment expert model in the flood disaster expert network model; When the remaining power of the portable edge device is greater than a first preset threshold, the dialect recognition expert model, the semantic clarification expert model, the risk avoidance decision expert model, the device optimization expert model, and the data assimilation and alignment expert model are loaded simultaneously; When the remaining power is greater than a second preset threshold and less than or equal to the first preset threshold, loading the device optimization expert model; When the remaining power is less than or equal to the second preset threshold, the dialect recognition expert model, the semantic clarification expert model, and the risk avoidance decision expert model are loaded.

3. The generative rainstorm and flood intelligent risk avoidance dialogue method according to claim 1 is characterized in that: After acquiring the user's voice stream data in real time according to the portable edge device, the method further includes: When a preset keyword is detected in the voice stream data, suspending non-critical background tasks and increasing the weight of resource allocation to a preset percentage; When it is detected that the number of modifiers in the voice stream data is greater than a preset threshold, a prefix tree algorithm is called to analyze the syntax to generate the core semantic components of the sentence.

4. The generative rainstorm and flood intelligent risk avoidance dialogue method according to claim 3 is characterized in that: The analyzing the speech features that meet the preset clarity conditions to predict and complete the interrupted speech content of the speech stream data to generate a complete speech sequence includes: When a long pause or an ambiguous phoneme segment is detected in the voice stream data, triggering a node backtracking mechanism using the prefix tree algorithm to count the phoneme path frequency of the voice feature; A phoneme sequence path is selected based on the phoneme path frequency, so as to predict and complete the interrupted speech content based on the phoneme sequence path to generate the complete speech sequence.

5. The generative rainstorm and flood intelligent risk avoidance dialogue method according to claim 1 is characterized in that: The uploading of crowd-source data in the user's area and verification of the crowd-source data, assimilating the verified crowd-source data into a preset flood forecast model to generate a corrected flood level and flow forecast, and generating flood risk information based on the corrected flood level and flow forecast, includes: Uploading crowd-source data of the user's area, and preprocessing the crowd-source data to generate preprocessed crowd-source data; Cross-validating the preprocessed crowd-source data to generate the validated crowd-source data; Assimilating the verified multi-source data into the preset flood forecast model to generate the corrected flood level and flow forecast data; Based on the corrected flood level and flow forecast data, the magnitude of flood risk change is identified, and the flood threshold and warning level are adjusted according to the magnitude of flood risk change to generate the flood risk information.

6. The generative rainstorm and flood intelligent risk avoidance dialogue method according to claim 1 is characterized in that: The formula for calculating the credibility of the standard text is: in, C Indicates the credibility value of the spatiotemporal alignment of multi-source data in the system, represents the time dimension weight, represents the spatial dimension weight, represents the semantic dimension weight, Indicates the credibility score corresponding to the time dimension, represents the credibility score corresponding to the spatial dimension, represents the credibility score corresponding to the semantic dimension, Represents the time dimension index, represents the spatial dimension index, Represents the semantic dimension index; The calculation formula for the credibility score corresponding to the time dimension is: in, Indicates the absolute value difference between the voice interaction time and the sensor recording time, Indicates the maximum allowed time difference under the warning level; The calculation formula for the credibility score corresponding to the spatial dimension is: in, Indicates the user's location and i The distance between monitoring points, Indicates the i The water level status of each monitoring point, Indicates the monitoring point index, Indicates the number of monitoring points; The calculation formula for the credibility score corresponding to the semantic dimension is: in, n Indicates the number of semantic clarification questions asked.

7. A generative rainstorm and flood intelligent risk avoidance dialogue system, characterized by: include: A prefix tree initialization module is used to initialize the data structure of the prefix tree and load the expert network model under the dynamic hybrid expert architecture according to the operating status of the portable edge device; a preprocessing module, configured to utilize the expert network model to obtain, in real time, voice stream data of the user from the portable edge device when the user is in a flood disaster emergency interaction scenario, and preprocess the voice stream data to extract voice features that meet preset clarity conditions; A prediction module, configured to analyze the speech features that meet the preset clarity conditions to predict and complete the interrupted speech content of the speech stream data and generate a complete speech sequence; a crowdsource data uploading module, configured to upload crowdsource data in the user's area and verify the crowdsource data, so as to assimilate the verified crowdsource data into a preset flood forecasting model, generate a corrected flood level and flow forecast, and generate flood risk information based on the corrected flood level and flow forecast; A decoding module, configured to decode the dialect speech in the complete speech sequence into standard text based on the flood risk information, and output a confidence score of the standard text; An interactive module, configured to ask the user questions based on the confidence level of the standard text, so as to guide the user to clarify ambiguous semantics and generate user semantic information and environmental data that meet preset high-confidence conditions; The risk avoidance dialogue module is used to generate a dynamic risk avoidance dialogue strategy for rainstorms and floods based on the user semantic information and the environmental data that meet the preset high credibility conditions, according to the flood and rainstorm warning level and the user's location coordinates.

8. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the generative rainstorm and flood intelligent risk avoidance dialogue method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: The program is executed by a processor to implement the generative rainstorm and flood intelligent risk avoidance dialogue method as described in any one of claims 1-6.

10. A computer program product comprising a computer program, characterized in that The computer program is executed to implement the generative rainstorm and flood intelligent risk avoidance dialogue method according to any one of claims 1 to 6.

Images