Risk assessment of target person and recommendation systems and methods thereof
An AI-based system for mountain rescue operations integrates multiple data sources to provide accurate and timely risk assessments and recommendations, addressing information gaps and improving rescue efficiency.
Patent Information
- Application Number
- PCT/CN2025/084741
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2024-03-26
- Filing Date
- 2025-03-25
- Publication Date
- 2025-10-02
AI Technical Summary
Mountain rescue operations face challenges due to a lack of accurate information about a hiker's health status, location, and route, leading to delays and increased risks for both the hiker and rescuers, compounded by the need to sift through vast amounts of contradictory intelligence.
A system utilizing AI-based risk assessment and recommendation, integrating mobile, environmental, and open data sources, with a big-data database and machine learning algorithms to analyze hiker data, environmental conditions, and historical information to provide timely and precise risk assessments and recommendations.
Enhances the efficiency and accuracy of rescue operations by providing timely and actionable insights, reducing costs and risks, and improving the chances of successful rescue through enhanced data processing and precise risk assessment.
Smart Images

Figure CN2025084741_02102025_PF_FP_ABST
Abstract
Description
RISK ASSESSMENT OF TARGET PERSON AND RECOMMENDATION SYSTEMS AND METHODS THEREOFCROSS-REFERENCE TO RELATED APPLICATION
[0001] This application claims priority to, and the benefit of, U.S. Provisional Application Serial No. 63 / 570,230 filed Mar. 26, 2024, entitled System and Computing Device for Strategic Algorithm Model with Attention-Weighted Actionable Rescue Intelligence (STAARI) . The entire contents of the foregoing application are hereby incorporated by reference for all purposes.FIELD OF INVENTION
[0002] This invention relates to risk assessment technologies, and in particular, relates to risk assessment of a target person and recommendation systems and methods thereof.BACKGROUND OF INVENTION
[0003] Mountain Rescue Operation is facing numerous challenges that are hindering the efforts to retrieve or discover survivors and to provide aid to those affected.
[0004] One of the primary challenges is the dearth of information about the hiker’s situation. Rescuers are often in the dark about the hiker’s heath status, their exact location, and the route they took while hiking. The lack of information can lead to a delay in rescue efforts and increase the risk of injury or harm to both the hiker and the rescue team.
[0005] Another significant challenge in mountain rescue is the vast amount of intelligence that rescuers must sift through, which often contains contradictory information and signals. Rescuers must piece together numerous small bits of information from a variety of sources, which can be a daunting task. Moreover, rescuers must quickly determine which pieces of information are reliable and take timely action to execute a successful rescue operation.
[0006] Therefore, there is urgent need for new and efficient intelligent solutions to provide risk assessment of a target person (such as a hiker) and recommendation.SUMMARY OF INVENTION
[0007] In the light of the foregoing background, in one aspect, provided is a system of a target person risk assessment and recommendation with AI based on big data from multiple sources.
[0008] Accordingly, an example embodiment of the present invention relates to a target person risk assessment and recommendation system, comprising a plurality of data collecting modules; a big-data database module; a data analysis module; and a user interface module, wherein the plurality of data collecting modules, the big-data database module, the data analysis module and user interface module are operatively connected with each other, wherein the plurality of data collecting modules comprise: a mobile data collector that is configured to obtain at least one mobile data from a mobile device of the target person; an environmental data collector that is configured to obtain at least one environmental data from at least one environmental data source; and an open data collector that is configured to obtain at least one open data from internet, wherein the big-data database module is configured to store at least one historical data, wherein the data analysis module comprises: a plurality of estimators that are configured to analyze the at least one mobile data, the at least one environmental data, the at least one open data from the plurality of data collecting modules, and / or the at least one historical data from the big-data database to obtain a plurality of feature estimate data; at least one general embedding layer that is configured to analyze the at least one feature estimate data from the plurality of estimators to obtain at least one general feature data; and a classifier that is configured to analyze the at least one general feature data from the at least one general embedding layer to obtain the risk assessment of the target person, and at least one recommendation, and wherein the user interface module is configured to output the risk assessment, and the at least one recommendation from the classifier. In another example embodiment, provided is a target person risk assessment and recommendation method, comprising the steps of: providing a system as claimed in any one of the preceding claims; obtaining, by the mobile data collectors, at least one mobile data; obtaining, by the environmental data collector, at least one environmental data; obtaining, by the open data collector, at least one open data; storing, by the big-data database, at least one historical data; analyzing, by the plurality of estimators, the at least one mobile data, the at least one environmental data, the at least one open data and / o r the at least one historical data to obtain at least one feature estimate data; analyzing, by the at least one general embedding layer, the at least one feature estimate data to obtain at least one general feature data; analyzing, by the classifier, at least one general feature data to obtain the risk assessment of the target person and at least one recommendation; and outputting, by the user interface module, the risk assessment and the at least one recommendation.
[0009] The above example embodiments have benefits and advantages over conventional technologies. For example, in some embodiments, the provided systems and methods address one or more challenges described such as by analyzing and inferring walking patterns of a target person (such as a hiker) using the limited data from the target person’s mobile data (e.g., by a mobile APP) and the open source data, and the AI transformer technology that generates text-based recommendations for the user such as a rescue team, providing them with crucial information to ensure an cost efficient, accurate and successful rescue operation. In some embodiments, the provided systems and methods increase the chance of successful rescue of the target person, reduce the cost and risk of the user (e.g., a rescue team) , and provide warnings and relevant information to related personals (such as the outdoor clubs, government, forest and mountain management, etc. ) . In some embodiments, the provided systems and methods increase processing speed of the data, shorten the turnaround time of the analysis of huge amount and various types of data and / or provide accurate, efficient assessment of the risk level of the target person, and precise and useful recommendation to the user by using the algorithms and AI models as described herein. BRIEF DESCRIPTION OF FIGURES
[0010] FIG. 1 is a schematic diagram of an example system, according to an example embodiment.
[0011] FIG. 2 is another schematic diagram of an example system, according to another example embodiment.
[0012] FIG. 3 is a schematic diagram showing an exhaustion level estimator 331 of an example system, according to another example embodiment.
[0013] FIG. 4 is a schematic diagram showing a reasonable location estimator 432 of an example system, according to another example embodiment.
[0014] FIG. 5 is a schematic diagram showing a zone estimator 533 and a signal landscape estimator 534 of an example system, according to another example embodiment.
[0015] FIG. 6 is a schematic diagram showing a signal abnormality estimator 635 of an example system, according to another example embodiment.
[0016] FIG. 7 is a schematic diagram showing general embedding layers 738 of an example system, according to another example embodiment.
[0017] FIG. 8 is a schematic diagram showing a classifier 839 of an example system, according to another example embodiment.
[0018] FIG. 9 is a schematic diagram showing an example method 900 for target person risk assessment and recommendation.DETAILED DESCRIPTIONDEFINITIONS
[0019] As used herein and in the claims, the terms “comprising” (or any related form such as “comprise” and “comprises” ) , “including” (or any related forms such as “include” or “includes” ) , “containing” (or any related forms such as “contain” or “contains” ) , means including the following elements but not excluding others. It shall be understood that for every embodiment in which the term “comprising” (or any related form such as “comprise” and “comprises” ) , “including” (or any related forms such as “include” or “includes” ) , or “containing” (or any related forms such as “contain” or “contains” ) is used, this disclosure / application also includes alternate embodiments where the term “comprising” , “including, ” or “containing, ” is replaced with “consisting essentially of” or “consisting of” . These alternate embodiments that use “consisting of” or “consisting essentially of” are understood to be narrower embodiments of the “comprising” , “including, ” or “containing, ” embodiments.
[0020] For the sake of clarity, “comprising” , including, and “containing” , and any related forms are open-ended terms which allows for additional elements or features beyond the named essential elements, whereas “consisting of” is a closed end term that is limited to the elements recited in the claim and excludes any element, step, or ingredient not specified in the claim.
[0021] For the sake of clarity, “characterized by” or “characterized in” (together with their related forms as described above) , does not limit or change the nature of whether the list of terms following it are open or closed. For example, in a claim directed towards “acomposition comprising A, B, C, and characterized in D, E, and F” , the elements D, E, and F are still open-ended terms and the claim is meant to include other elements due to the use of the word “comprising” earlier in the claim.
[0022] As used herein and in the claims, the singular forms “a, ” “an, ” and “the” are intended to include the plural forms as well unless the context clearly indicates otherwise. Where a range is referred to in the specification and the claims, the range is understood to include each discrete point within the range. For example, 1-7 means 1, 2, 3, 4, 5, 6, and 7. Another example is that (a) - (d) means (a) , (b) , (c) , and (d) .
[0023] As used herein and in the claims, the terms “general” or “generally” , or “substantial” or “substantially” mean that the recited characteristic, shape, state, structure, or value need not be achieved exactly, but that deviations or variations, including for example, tolerances, measurement error, measurement accuracy limitations and other factors known to those of skill in the art, may occur in amounts that do not preclude the effect the characteristic was intended to provide. For example, an object that has a “generally” cylindrical shape would mean that the object has either an exact cylindrical shape or a nearly exact cylindrical shape.
[0024] As used herein, the term “about” is understood as within a range of normal tolerance in the art and not more than ±10%of a stated value. By way of example only, about 50 means from 45 to 55 including all values in between. As used herein, the phrase “about” a specific value also includes the specific value, for example, about 50 includes 50.
[0025] As used herein and in the claims, “operatively connect” refers to a functional or operational connection between two elements such as components, modules or systems that allows them to communicate, or interact with each other. Such connection may be direct or indirect. In some examples, the elements are in communication connection with each other.
[0026] As used herein and in the claims, “in communication connection with” refers to a connection between two elements such as components, modules or systems that allows them to exchange data, signals, or information, enabling coordination or interaction between them. Such connection may be direct or indirect, wired or wireless.
[0027] As used herein and in the claims, “a target person” refers to a subject or individual of interest for risk assessment. In some examples, the target person has an associated mobile device such as mobile cell phone with a mobile APP installed to track and transceiver data related to the target person to the system. In some examples, the target person is a hiker or a potentially missing person.
[0028] As used herein and in the claims, “feature data” refers to the specific attributes, characteristics, or variables extracted or processed from raw data that are used as inputs for machine learning models or data analysis to identify patterns, make predictions, or perform classifications.
[0029] As used herein and in the claims, “risk assessment” refers to a process of identifying, analyzing, and evaluating potential risks or hazards of the target person to determine the likelihood and impact, enabling to provide recommendation to mitigate or manage them.
[0030] As used herein and in the claims, “recommendation” refers to providing suggestion or proposal generated by the system to guide a user toward specific actions, such as based on the risk assessment.
[0031] As used herein and in the claims, “big-data database” refers to a specialized system to store, manage and analyze large and complex datasets, such as by leveraging distributed computing hardware, such as clustered memory and multi-core processors, to handle massive volumes of data.
[0032] As used herein and in the claims, “module” refers to is a hardware and / or software unit that contributes to the overall functionality of a system such as a computer system.
[0033] As used herein and in the claims, “estimator” refers to a software, module component or algorithm that estimates or predicts values, outcomes, based on the input of the data. In some examples, an estimator includes an AI model for implementing the functions thereof.
[0034] As used herein and in the claims, “embedding layer” refers to neural network module or algorithm that transforms high-dimensional, discrete input data into lower-dimensional, continuous vector representations, capturing meaningful relationships and patterns in the data. In some examples, an embedding layer includes one or more concatenation layers, one or more fully-connected layers, one or more normalization layers, and / or one or more splitting layers.
[0035] As used herein and in the claims, “classifier” refers to a machine learning model or algorithm that categorizes input data into predefined classes or labels based on learned patterns from training data. In some examples, a classifier includes an AI Natural Language Model to generate text data based on the input data.
[0036] As used herein and in the claims, “GNSS” refers to Global Navigation Satellite System, a satellite-based positioning technology that provides global coverage for determining precise location, velocity, and timing data. GNSS encompasses multiple satellite constellations, including the Global Positioning System (GPS) (United States) , GLONASS (Russia) , Galileo (European Union) , and BeiDou (China) , among others.
[0037] As used herein and in the claims, “processing” (or other forms such as “process” ) refers to the systematic execution of operations on data to transform, analyze, or manipulate it for a specific purpose. In some embodiments, this includes collecting, organizing, storing, analyzing, and interpreting data using computational techniques such as filtering, classification, pattern recognition, and machine learning.
[0038] As used herein and in the claims, “analyzing” (or other forms such as “analyze” ) refers to the process of systematically examining and interpreting data to extract meaningful insights, detect patterns, or make informed decisions if not stated otherwise.
[0039] As used herein and in the claims, “open data” refers to publicly available data that is freely accessible, usable, and shareable by anyone. In some embodiments, open data is structured and provided in machine-readable formats, enabling automated collection, processing, and analysis by computer systems.
[0040] Although the description referred to particular embodiments, the disclosure should not be construed as limited to the embodiments set forth herein. NUMBERED EMBODIMENTSSet I
[0041] Embodiment 1.1 A system for determining the journey risk level of the hiker (or the missing person) and provides textual recommendations to Police Rescue operation, comprising: a Hiker Data Collector (such as Hong Kong Police Force Mountain Rescue Mobile App) ; an Environmental Data Collector (such as weather, temperature, fogginess, humidity) ; an Open Data Collector (such as social media on Hiking Hot-Spot Trends) ; a Big Data Database on Historical Hiker-Experience and Landscape Information; a recommendation output data result; a user interface system (such as Mobile Application or Computer System) ; an artificial intelligence algorithm for receiving the information from hiker data collector, environmental data collector, open data collector, the artificial intelligence algorithm being configured to: determine the journey risk level of the hiker (or missing person) ; provide textual actionable recommendations as output; having capability to enhance the algorithm accuracy by actions taken, outcome result in human-feedback loop.
[0042] Embodiment 1.2 The system of embodiment 1.1 wherein the Artificial Intelligence Algorithm is configured to accept data from Hiker Data Collector (hiker’s experience level, hiker’s exhaustion level, elevation gain / loss, walking speed, cell-phone GPS data) .
[0043] Embodiment 1.3 The system of any one of embodiments 1.1 and 1.2 wherein the Artificial Intelligence Algorithm is configured to accept data from Environmental Data Collector (such as online weather data, nearby temperature, geographical landscape data) .
[0044] Embodiment 1.4 The system of any one of embodiments 1.1 to 1.3 wherein the Artificial Intelligence Algorithm is configured to accept web and social media data from Open Data Collector (such as social media Hiking Hot-spot trends data) .
[0045] Embodiment 1.5 The system of any one of embodiments 1.1 to 1.4 wherein the Artificial Intelligence Algorithm is configured to query and accept data from big data database on historical hiker-experience and landscape information.
[0046] Embodiment 1.6 The system of any one of embodiments 1.1 to 1.5 wherein the Artificial Intelligence Algorithm is configured with customized AI Transformer Model (AI Large Language Model) to analyse and retrieve essential information (Including social media Hiking Hot-spot locations) from open data sources.
[0047] Embodiment 1.7 The system of any one of embodiments 1.1 to 1.6 wherein the Artificial Intelligence Algorithm is configured with algorithms for Accident Detection Mechanism.
[0048] Embodiment 1.8 The system of any one of embodiments 1.1 to 1.7 wherein the Artificial Intelligence Algorithm is configured with algorithms to estimate the exhaustion level by using elevation gain / loss, walking speed and amateur level.
[0049] Embodiment 1.9. The system of any one of embodiments 1.1 to 1.8 wherein the Artificial Intelligence Algorithm is configured with algorithms to estimate the reasonable location by using landscape data, hiker GPS coordinates, hiker speed and hiker proficiency level.
[0050] Embodiment 1.10 The system of any one of embodiments 1.1 to 1.9 wherein the Artificial Intelligence Algorithm is configured with algorithms to classify the zone by using the cell-site signal. The zone is classified into 3 categories: Sea, Urban and Mountain.
[0051] Embodiment 1.11 The system of any one of embodiments 1.1 to 1.10 wherein the Artificial Intelligence Algorithm is configured with algorithms to estimate the landscape understanding by aggregating multiple cell-site signals to estimate the Black-Spot and White-Spot with seasonal overgrowth factor.
[0052] Embodiment 1.12 The system of any one of embodiments 1.1 to 1.11 wherein the Artificial Intelligence Algorithm is configured with algorithms to estimate 3 cell signals abnormality include “No Signal” , “Unreasonable Staying on Mountain” , “Forget to Stop Tracking” by using output from zone classification in claim 10 and landscape understanding in claim 11.
[0053] Embodiment 1.13 The system of any one of embodiments 1.1 to 1.12 wherein the Artificial Intelligence Algorithm is configured with algorithms to estimate the environmental impact include but not limit to the temperature impact to the hiker speed and behaviour.
[0054] Embodiment 1.14 The system of any one of embodiments 1.1 to 1.13 wherein the Artificial Intelligence Algorithm is configured with algorithms to retrieve trending locations that is considered high-risk to less experienced hiker.
[0055] Embodiment 1.15 The system of any one of embodiment 1.1 to 1.14 wherein the Artificial Intelligence Algorithm is configured with big data database for aggregating historical hiker routes and walker behaviour.
[0056] Embodiment 1.16 The system of any one of embodiments 1.1 to 1.15 wherein the Artificial Intelligence Algorithm is configured with big embedding feature vector layer to represent all compressed factors for risk level estimation.
[0057] Embodiment 1.17 The system of any one of embodiments 1.1 to 1.16 wherein the Artificial Intelligence Algorithm is configured with probabilistic classifier to calculate the Journey Risk Level of Hiker. The algorithm utilize attention-weighted mathematics to weight different factors in risk level.
[0058] Embodiment 1.18. The system of any one of embodiments 1.1 to 1.17 wherein the Artificial Intelligence Algorithm utilize the AI technique include but not limited to the AI NLP transformer architecture to provide textual recommendations as the output of the system.
[0059] Embodiment 1.19. The system of any one of embodiments 1.1 to 1.18 wherein the Artificial Intelligence Algorithm accepts the human feedback data as control feedback loop to fine-tune the embedding feature vectors and AI probability classifier. It acts as self-learning mechanism in the AI system.
[0060] Embodiment 1.20. The system of embodiment 1.19 wherein the feedback data is saved and indexed into the Big Data Database in Claim 15.
[0061] Embodiment 1.21 The system of any one of embodiments 1.1 to 1.20 wherein the user interface system displays the output of analysis results, recommendations, dashboard. The user interfaced system allows user feedback to the AI system.
[0062] Embodiment 1.22. The system of Embodiment 1.1 to 1.21 wherein the user interface system can include one or plurality of hardware devices. The hardware device includes but not limited to any computing machine and mobile device.Set II
[0063] Embodiment 2.1. A target person risk assessment and recommendation system, comprising: a plurality of data collecting modules; a big-data database module; a data analysis module; and a user interface module, wherein the plurality of data collecting modules, the big-data database module, the data analysis module and user interface module are operatively connected with each other, wherein the plurality of data collecting modules comprise: a mobile data collector that is configured to obtain at least one mobile data from a mobile device of the target person; an environmental data collector that is configured to obtain at least one environmental data from at least one environmental data source; and an open data collector that is configured to obtain at least one open data from internet, wherein the big-data database module is configured to store at least one historical data, wherein the data analysis module comprises: a plurality of estimators that are configured to analyze the at least one mobile data, the at least one environmental data, the at least one open data from the plurality of data collecting modules, and / or the at least one historical data from the big-data database to obtain a plurality of feature estimate data; at least one general embedding layer that is configured to analyze the at least one feature estimate data from the plurality of estimators to obtain at least one general feature data; and a classifier that is configured to analyze the at least one general feature data from the at least one general embedding layer to obtain the risk assessment of the target person, and at least one recommendation, and wherein the user interface module is configured to output the risk assessment, and the at least one recommendation from the classifier.
[0064] Embodiment 2.2. The system of embodiment 2.1, wherein the open data collector is configured to collect at least one online text that is associated with at least one hot spot location, and wherein the plurality of estimators comprise: at least one artificial intelligence (AI) natural language processing (NLP) estimator that is configured to analyze the at least one online text from the open data collector based on the at least one hot spot location to obtain at least one hot-spot feature estimate data; and / or an environmental impact estimator that is configured to analyze the at least one environmental data from the environmental data collector to obtain at least one environmental feature estimate data.
[0065] Embodiment 2.3. The system of embodiment 2.1 or 2.2, wherein the big-data database module is configured to receive and process the at least one mobile data from the mobile data collector, the at least one environmental data from the environmental data collector, the at least one open data from the open data collector, and / or the plurality of feature estimate data from the plurality of estimators to update the at least one historical data.
[0066] Embodiment 2.4. The system of any one of the preceding embodiments, wherein the mobile data collector is configured to collect at least one GNSS data, at least one coordinate data, at least one elevation change data, at least one signal characteristics data, and at least one amateur level data from the mobile device of the target hiker, wherein the big-data database module is configured to store at least one historical elevation change feature data, at least one historical walking speed feature data, at least one historical hiking location feature data, at least one historical hiking speed feature data, and / or at least one historical signal landscape feature data, wherein the plurality of estimators comprise: an exhaustion level estimator that is configured to analyze the GNSS data, the at least one elevation change data, at least one amateur level data, the at least one historical elevation change feature data, and the at least one historical walking speed data to obtain at least one exhaustion feature estimate data; a reasonable location estimator that is configured to analyze the at least one coordinate data, at least one amateur level data, at least one historical hiking location feature data, and / or at least one historical hiking speed feature data to obtain at least one reasonable location feature estimate data; a zone estimator that is configured to analyze the at least one signal characteristics data to obtain at least one zone feature estimate data; a signal landscape estimator that is configured to analyze the at least one signal characteristics data and / or the historical signal landscape data to obtain at least one signal landscape feature estimate data; and / or a signal abnormality estimator that is configured to analyze the at least one zone feature estimate data from the zone estimator, the at least one signal landscape feature estimate data from the signal landscape estimator, and the at least one amateur level data from the mobile data collector to obtain at least one signal abnormality feature estimate data.
[0067] Embodiment 2.5. The system of any one of the preceding embodiments, wherein the at least one general embedding layer comprises: a general embedding concatenation submodule that is configured to concatenate the plurality of feature estimate data from the plurality of estimators to obtain at least one pre-general abnormality feature data; a general embedding fully-connected layer submodule that is configured to analyze the at least one pre-general abnormality feature data from the general embedding concatenation submodule to obtain at least one analyzed pre-general abnormality feature data; and a general embedding normalization layer submodule that is configured to analyze the at least one analyzed pre-general abnormality feature data from the general embedding fully-connected layer submodule to obtain the at least one general feature data.
[0068] Embodiment 2.6. The system of any one of the preceding embodiments, wherein the classifier comprises: a classifier AI attention-weighted risk calculation submodule comprising: a classifier self-attention layer that is configured to analyze the at least one general feature data from the at least one general embedding layer to obtain at least one self-attention feature data; and a classifier two-layers-fully-connected layer that is configured to analyze the at least one self-attention feature data from the classifier self-attention layer to obtain the risk assessment; and a classifier transformer-based recommendation submodule comprising: a classifier transformer-based Large Language Model (LLM) that is configured to analyze the at least one general feature data from the at least one general embedding layer, at least one self-attention feature data from the classifier self-attention layer, and the risk assessment from the classifier two-layers-fully-connected layer to obtain the at least one recommendation.
[0069] Embodiment 2.7. The system of any one of the preceding embodiments, wherein the user interface module is configured to further receives at least one action outcome data of the hiker, and wherein the user interface is configured to send the at least one action outcome data to the big-data database to update the at least one historical data.
[0070] Embodiment 2.8. A risk assessment of a target person and recommendation system, comprising: a mobile data collector configured to obtain at least one mobile data comprising at least one GNSS data, at least one coordinate data, at least one elevation change data, at least one signal characteristics data, and at least one amateur level data from the mobile device of the target hiker from the mobile device of the target person; an environmental data collector configured to obtain at least one environmental data from at least one environmental data source; an open data collector configured to obtain at least one open data comprising at least one online text that is associated with at least one hot spot location from internet; a big-data database configured to store at least one historical data comprising at least one historical elevation change feature data, at least one historical walking speed feature data, at least one historical hiking location feature data, at least one historical hiking speed feature data, and at least one historical signal landscape feature data; at least one artificial intelligence (AI) natural language processing (NLP) estimator that is configured to analyze the at least one online text from the open data collector based on the at least one hot spot location to obtain at least one hot-spot feature estimate data; an environmental impact estimator that is configured to analyze the at least one environmental data from the environmental data collector to obtain at least one environmental feature estimate data; an exhaustion level estimator that is configured to analyze the GNSS data, the at least one elevation change data, at least one amateur level data, the at least one historical elevation change feature data, and the at least one historical walking speed data to obtain at least one exhaustion feature estimate data; a reasonable location estimator that is configured to analyze the at least one coordinate data, at least one amateur level data, at least one historical hiking location feature data, and at least one historical hiking speed feature data to obtain at least one reasonable location feature estimate data; a zone estimator that is configured to analyze the at least one signal characteristics data to obtain at least one zone feature estimate data; a signal landscape estimator that is configured to analyze the at least one signal characteristics data and the historical signal landscape data to obtain at least one signal landscape feature estimate data; a signal abnormality estimator that is configured to analyze the at least one zone feature estimate data from the zone estimator, the at least one signal landscape feature estimate data from the signal landscape estimator, and the at least one amateur level data from the mobile data collector to obtain at least one signal abnormality feature estimate data; a general embedding concatenation submodule that is configured to concatenate the at least one hot-spot feature estimate data from the at least one artificial intelligence (AI) natural language processing (NLP) estimator, the at least one environmental feature estimate data from the environmental impact estimator, the at least one exhaustion feature estimate data from the exhaustion level estimator, the at least one reasonable location feature estimate data from the reasonable location estimator, the at least one zone feature estimate data from the zone estimator, the at least one signal landscape feature estimate data from the signal landscape estimator, and the at least one signal abnormality feature estimate data from the signal abnormality estimator to obtain at least one pre-general abnormality feature data; a general embedding fully-connected layer submodule that is configured to analyze the at least one pre-general abnormality feature data from the general embedding concatenation submodule to obtain at least one analyzed pre-general abnormality feature data; a general embedding normalization layer submodule that is configured to analyze the at least one analyzed pre-general abnormality feature data from the general embedding fully-connected layer submodule to obtain the at least one general feature data; a classifier self-attention layer that is configured to analyze the at least one general feature data from the general embedding normalization layer submodule to obtain at least one self-attention feature data; a classifier two-layers-fully-connected layer that is configured to analyze the at least one self-attention feature data from the self-attention layer submodule to obtain the risk assessment; a classifier transformer-based Large Language Model (LLM) submodule that is configured to analyze the at least one general feature data from the general embedding normalization layer submodule, at least one self-attention feature data from the classifier self-attention layer, and the risk assessment from the classifier two-layers-fully-connected layer to obtain the at least one recommendation; a user interface configured to output the risk assessment from the two-layers-fully- connected layers submodule, and the at least one recommendation from the classifier transformer-based Large Language Model (LLM) submodule, wherein the mobile data collector, the environmental data collector, the open data collector, the big-data database, the at least one artificial intelligence (AI) natural language processing (NLP) estimator, the environmental impact estimator, the exhaustion level estimator, the reasonable location estimator, the zone estimator, the signal landscape estimator, the signal abnormality estimator, the general embedding concatenation submodule, the general embedding fully-connected layer submodule, the general embedding normalization layer submodule, the classifier self-attention layer submodule, the classifier two-layers-fully-connected layer, the classifier transformer-based Large Language Model (LLM) submodule, and the user interface are operatively connected with each other, wherein the big-data database is configured to receive and process at least one mobile data from the mobile data collector, the at least one environmental data from the environmental data collector, the at least at least one open data from the open data collector, the at least one hot-spot feature estimate data from the at least one artificial intelligence (AI) natural language processing (NLP) estimator, the at least one environmental feature estimate data from the environmental impact estimator, the at least one exhaustion feature estimate data from the exhaustion level estimator, the at least one reasonable location feature estimate data from the reasonable location estimator, the at least one zone feature estimate data from the zone estimator, the at least one signal landscape feature estimate data from the signal landscape estimator, and the at least one signal abnormality feature estimate data from the signal abnormality estimator to update the at least one historical data, wherein the user interface module is configured to receive at least one action outcome data of the target person, and wherein the user interface is configured to send the at least one action outcome data to the big-data database to update the at least one historical data.
[0071] Embodiment 2.9. A target person risk assessment and recommendation method, comprising the steps of: providing a system as claimed in any one of the preceding claims; obtaining, by the mobile data collector, at least one mobile data; obtaining, by the environmental data collector, at least one environmental data; obtaining, by the open data collector, at least one open data; storing, by the big-data database, at least one historical data; analyzing, by the plurality of estimators, the at least one mobile data, the at least one environmental data, the at least one open data and / or the at least one historical data to obtain at least one feature estimate data; analyzing, by the at least one general embedding layer, the at least one feature estimate data to obtain at least one general feature data; analyzing, by the classifer, at least one general feature data to obtain the risk assessment of the target person and at least one recommendation; and outputting, by the user interface module, the risk assessment and the at least one recommendation.
[0072] Embodiment 2.10. The method of any one of the preceding embodiments, wherein the step (iv) obtaining at least one open data from Internet comprises the step of collecting, by the open data collector, at least one online text that is associated with at least one hot spot location, wherein the step (vi) analyzing the at least one mobile data, the at least one environmental data, the at least one open data from the plurality of data collectors, and / or the at least one historical data comprises the steps of: analyzing, by the at least one artificial intelligence (AI) natural language processing (NLP) estimator, the at least one online text from the open data collector based on the hot spot location to obtain at least one hot-spot feature estimate data; and / or analyzing, by the environmental impact estimator, the at least one environmental data from the environmental data collector to obtain at least one environmental feature estimate data.
[0073] Embodiment 2.11. The method of any one of the preceding embodiments, further comprising the steps of: receiving and processing, by the big-data database module, the at least one mobile data, the at least one environmental data, the at least one open data, and / or the at least feature estimate data; and updating, by the big-data database module, the at least one historical data.
[0074] Embodiment 2.12. The method of any one of the preceding embodiments, further comprising the steps of: collecting, by the mobile data collector, at least one GNSS data, at least one coordinate data, at least one elevation change data, at least one signal characteristics data, and at least one amateur level data from the mobile device of the target hiker; storing, by the big-data database module, at least one historical elevation change feature data, at least one historical walking speed feature data, at least one historical hiking location feature data, at least one historical hiking speed feature data, and / or at least one historical signal landscape feature data, wherein the step (vi) analyzing the at least one mobile data, the at least one environmental data, the at least one open data from the plurality of data collectors, and / or the at least one historical data further comprises the steps of: analyzing, by the exhaustion level estimator, the GNSS data, the at least one elevation change data, at least one amateur level data, the at least one historical elevation change feature data, and the at least one historical walking speed data to obtain at least one exhaustion feature estimate data; analyzing, by the reasonable location estimator, at least one coordinate data, at least one amateur level data, at least one historical hiking location feature data, and / or at least one historical hiking speed feature data to obtain at least one reasonable location feature estimate data; analyzing, by the zone estimator, the at least one signal characteristics data to obtain at least one zone feature estimate data; analyzing, by the signal landscape estimator, the at least one signal characteristics data and / or the historical signal landscape data to obtain at least one signal landscape feature estimate data; and / or analyzing, by the signal abnormality estimator, the at least one zone feature estimate data from the zone estimator, the at least one signal landscape feature estimate data from the signal landscape estimator, and the at least one amateur level data from the mobile data collector to obtain at least one signal abnormality feature estimate data.
[0075] Embodiment 2.13. The method of any one of the preceding embodiments, wherein the step (vii) analyzing the at least one feature estimate data comprises the steps of: concatenating, by the general embedding concatenation submodule, the plurality of feature estimate data from the plurality of estimators to obtain at least one pre-general abnormality feature data; analyzing, by the general embedding fully-connected layer submodule, the at least one pre- general abnormality feature data from the general embedding concatenation submodule to obtain at least one analyzed pre-general abnormality feature data; and analyzing, by the general embedding normalization layer submodule, the at least one analyzed pre-general abnormality feature data from the general embedding fully-connected layer submodule to obtain the at least one general feature data.
[0076] Embodiment 2.14. The method of any one of the preceding embodiments, wherein the step (viii) analyzing at least one general feature data comprises the steps of: analyzing, by the classifier self-attention layer, the at least one general feature data to obtain at least one self-attention feature data; analyzing, by the classifier two-layers-fully-connected layer, the at least one self-attention feature data to obtain the risk assessment; and analyzing, by the classifier transformer-based Large Language Model (LLM) , the at least one general feature data, at least one self-attention feature data, and the risk assessment to obtain the at least one recommendation.
[0077] Embodiment 2.15. The method of any one of the preceding embodiments, further comprising the steps of: receiving, by the user interface module, at least one action outcome data of the target person; sending, by the user interface module, the at least one action outcome data to the big-data database; and updating, by the big-data database module, the at least one historical data.
[0078] Embodiment 2.16. A method of assessing risk of a target person and providing recommendation, comprising the steps of: providing a system as claimed in any one of the preceding embodiments; obtaining, by the mobile data collector, at least one mobile data comprising at least one GNSS data, at least one coordinate data, at least one elevation change data, at least one signal characteristics data, and at least one amateur level data from the mobile device of the target hiker from the mobile device of the target person; obtaining, by the environmental data collector, at least one environmental data from at least one environmental data source; obtaining, by the open data collector, at least one open data comprising at least one online text that is associated with at least one hot spot location from internet; storing, by the big-data database, at least one historical data comprising at least one historical elevation change feature data, at least one historical walking speed feature data, at least one historical hiking location feature data, at least one historical hiking speed feature data, and at least one historical signal landscape feature data; analyzing, by the at least one artificial intelligence (AI) natural language processing (NLP) estimator, the at least one online text based on the at least one hot spot location to obtain at least one hot-spot feature estimate data; analyzing, by the environmental impact estimator, the at least one environmental data to obtain at least one environmental feature estimate data; analyzing, by the exhaustion level estimator, the GNSS data, the at least one elevation change data, at least one amateur level data, the at least one historical elevation change feature data, and the at least one historical walking speed data to obtain at least one exhaustion feature estimate data; analyzing, by the reasonable location estimator, the at least one coordinate data, at least one amateur level data, at least one historical hiking location feature data, and at least one historical hiking speed feature data to obtain at least one reasonable location feature estimate data; analyzing, by the zone estimator, the at least one signal characteristics data to obtain at least one zone feature estimate data; analyzing, by the signal landscape estimator, the at least one signal characteristics data and the historical signal landscape data to obtain at least one signal landscape feature estimate data; analyzing, by the signal abnormality estimator, the at least one zone feature estimate data from the zone estimator, the at least one signal landscape feature estimate data from the signal landscape estimator, and the at least one amateur level data from the mobile data collector to obtain at least one signal abnormality feature estimate data. concatenating, by the general embedding concatenation submodule, the at least one hot- spot feature estimate data, the at least one environmental feature estimate data, the at least one exhaustion feature estimate data, the at least one reasonable location feature estimate data, the at least one zone feature estimate data, the at least one signal landscape feature estimate data, and the at least one signal abnormality feature estimate data to obtain at least one pre-general abnormality feature data; analyzing, by the general embedding fully-connected layer submodule, the at least one pre- general abnormality feature data to obtain at least one analyzed pre-general abnormality feature data; analyzing, by the a general embedding normalization layer submodule, the at least one analyzed pre-general abnormality feature data to obtain the at least one general feature data; analyzing, by the classifier self-attention layer, the at least one general feature data to obtain at least one self-attention feature data; analyzing, by the classifier two-layers-fully-connected layers submodule, the at least one self-attention feature data to obtain the risk assessment; analyzing, by the classifier transformer-based Large Language Model (LLM) submodule, the at least one general feature data, at least one self-attention feature data and the risk assessment to obtain the at least one recommendation; outputting, by the user interface, the risk assessment from the classifier two-layers-fully- connected layer, and the at least one recommendation from the classifier transformer-based Large Language Model (LLM) submodule; receiving and processing, by the big-data database, at least one mobile data, the at least one environmental data, the at least at least one open, the at least one hot-spot feature estimate data, the at least one environmental feature estimate data, the at least one exhaustion feature estimate data, the at least one reasonable location feature estimate data, the at least one zone feature estimate data, the at least one signal landscape feature estimate data, and the at least one signal abnormality feature estimate data to update the at least one historical data; receiving, by the user interface, at least one action outcome data of the target person; sending, by the user interface, the at least one action outcome data to the big-data database; and updating, by the big-data database, the at least one historical data. EXAMPLES
[0079] The following example embodiments alone or in combination may be practiced to provide a method and a system of risk assessment of a target person and providing recommendation.
[0080] In some examples, provided is a system of risk assessment, comprising one or more processors and one or more memory coupled to the one or more processors and storing instructions that, when executed by the one or more processors, cause the system to perform the methods as described herein.
[0081] In some examples, provided is a non-transitory computer-readable medium storing at least one computer program for implementing any method as described herein.EXAMPLE 1
[0082] Now referring to Fig. 1, the objective of the proposed system 100 with AI model –STAARI -is to assist the call-taker of Police Call Centre to take Rescue Actions with “Timely Actionable Insights” .
[0083] The proposed system and the AI model (STAARI) provide strategic recommendations in Mountain Rescue. By using the environmental data (including geographical landscape data, weather data, nearby temperature) , the open data (social media Hiking Hot-spot trends data) , and the hiker’s data (hiker’s experienced level, exhaustion level, elevation gain / loss, walking speed, cell-phone GPS data) , the AI model determine the journey risk level of the hiker (or the missing person) . The model provides the textual strategic recommendation and the rationale behind to Police Rescue operation.
[0084] Challenges Addressed in the proposed method
[0085] The AI model addresses the primary challenge by analysing and inferring the hiker’s walking patterns using the limited data from the HKPF Mountain Rescue Mobile App data and the open-data. The AI transformer technology then generates text-based recommendations for the rescue team, providing them with crucial information to ensure a successful rescue operation.
[0086] The AI model employs an attention-weighted structure to address the second challenge by understanding the significance of various pieces of information, such as historical walker experience data, weather data and open-source data, weighting them accordingly. This enables the model to give more significance to the most relevant data points, ensuring that the rescue team receives the most accurate and helpful information.
[0087] Various examples of the disclosure are discussed below. While specific implementations are discussed, it should be understood that this is done for illustrative purposes and variations with other components and configurations may be used without departing from the scope of the disclosure as defined by appended claims.
[0088] The present disclosure proposes a system with AI method (STAARI) shown in FIG. 1 to address the challenges.
[0089] At block 1 data is received from Hiker Mountain Rescue Mobile App Data Collector. This encompasses GPS location data, Hiker coordinate, Hiker elevation gain / loss, Hiker Phone cell-site signal and hiker proficiency level (amateur, intermediate, advanced) .
[0090] At block 2 weather data is received from Environmental Data collector. This data encompasses temperature, humidity and fogginess.
[0091] At block 3 data from web and social media platforms is received via Open Data Collector. This includes data on trending locations and locations flagged for high-risk for less experienced hiker.
[0092] At block 4 data from Hiker Mountain Rescue Mobile App Data Collector is received to estimate the exhaustion level. It includes AI probability estimation using elevation gain / loss, walking speed and amateur level. The AI algorithm can be any probabilistic regressor include but not limit to regression, support vector machines, random forest, 2-layers fully-connected neural network layers or Cross-attention layers.
[0093] At block 5 data from Hiker Mountain Rescue Mobile App Data Collector is received to estimate the reasonableness of location. The algorithm estimates the hiker location by using the previous hiker walking speed and the nearby hiking landscape.
[0094] At block 6 data from Hiker Mountain Rescue Mobile App Data Collector is received to estimate the zone including “Sea” Zone, “Urban” Zone, and “Mountain Zone” . If the hiker is at Urban Zone, it is expected that the risk level is very low. If the hiker is at Sea Zone, the time of flight over the sea should be under reasonable limit. If the hiker stays longer than expected at the same coordinate over the sea, the risk level is high. If the mountain zone is detected, other algorithms on calculating risk level is applied.
[0095] At block 7 data from Hiker Mountain Rescue Mobile App Data Collector is received to deduce the Black-Spot and White-Spot. If the signal is always at poor reception in the area after aggregating multiple hikers signal, the area is spotted as Black-Spot. It means that that area is normal for poor phone signal. Thus, a hiker walking near Black-Spot having poor signal cannot imply the High-Risk Level as it may be normal behaviour.
[0096] If the signal is always at the good reception in the area after aggregating multiple hikers signal, the area is spotted as White-Spot. It means that the area is expected to have good phone signal.
[0097] Thus, a hiker walking near White-Spot having poor signal may have a higher probability of risk of missing or out of battery.
[0098] The calculation is also taken the seasonal overgrowth factor into account. Hiker route changes due to overgrowth of grass.
[0099] At block 8 AI estimation of risk level of 3 abnormal cell signals abnormality will be estimated from the assisted data from block 6 and block 7.
[0100] At block 9 AI estimate the environmental impact such as temperature on the hiker speed and exhaustion level.
[0101] At block 10 social media hot spot location is received via AI NLP model for high-risk location.
[0102] At block 11 Big Data Database records the aggregate walker speed, hiker behaviour and hiker route.
[0103] At block 12 All calculated factors from block 4, block 5, block 6, block 7, block 8, block 9, block 10 and block 11 are represented in embedding feature vectors layer.
[0104] At block 13 The embedding feature vectors from block 12 is fed into AI probabilistic classifier to calculate the Journey Risk Level of Hiker. Attention-weight technique in block 14 is used to weight different factors to provide the risk level. The AI algorithm can be any probabilistic regressor include but not limit to regression, support vector machines, random forest or 2-layers fully-connected neural network layers.
[0105] At block 15 AI Textual recommendation using transformer NLP model will be used to generate the textual recommendations and actionable insights.
[0106] At block 16 The analysis is displayed in the user interface of the system in call centre. Police supervisor can take or not take actions according to the recommendations.
[0107] At block 16 Actual actions taken and outcome is fed back to block 11 Big Data Database to fine-tune the future recommendations.
[0108] The above embodiments are described by way of example only. Many variations are possible without departing from the scope of the invention as defined in the appended claims.
[0109] For clarity of explanation, in some instances the present technology has been presented as including individual functional blocks including functional blocks comprising devices, device components, steps or routines in a method embodied in software, or combinations of hardware and software.
[0110] Methods according to the above-described examples can be implemented using computer-executable instructions that are stored or otherwise available from computer readable media. Such instructions can comprise, for example, instructions and data which cause or otherwise configure a general-purpose computer, special purpose computer, or special purpose processing device to perform a certain function or group of functions. Portions of computer resources used can be accessible over a network. The computer executable instructions may be, for example, binaries, intermediate format instructions such as assembly language, firmware, or source code. Examples of computer-readable media that may be used to store instructions, information used, and / or information created during methods according to described examples include magnetic or optical disks, flash memory, Universal Serial Bus (USB) devices provided with non-volatile memory, networked storage devices, and so on.
[0111] Devices implementing methods according to these disclosures can comprise hardware, firmware and / or software, and can take any of a variety of form factors. Typical examples of such form factors include laptops, smart phones, small form factor personal computers, personal digital assistants, and so on. Functionality described herein also can be embodied in peripherals or add-in cards. Such functionality can also be implemented on a circuit board among different chips or different processes executing in a single device, by way of further example.
[0112] The instructions, media for conveying such instructions, computing resources for executing them, and other structures for supporting such computing resources are means for providing the functions described in these disclosures.
[0113] Although a variety of examples and other information was used to explain aspects within the scope of the appended claims, no limitation of the claims should be implied based on particular features or arrangements in such examples, as one of ordinary skill would be able to use these examples to derive a wide variety of implementations. Further and although some subject matter may have been described in language specific to examples of structural features and / or method steps, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to these described features or acts. For example, such functionality can be distributed differently or performed in components other than those identified herein. Rather, the described features and steps are disclosed as examples of components of systems and methods within the scope of the appended claims.
[0114] It is to be understood that any feature described in relation to any one example may be used alone, or in combination with other features described, and may also be used in combination with any features of any other of the examples, or any combination of any other of the examples. EXAMPLE 2 SYSTEM ARCHITECTURE
[0115] Now referring to FIG. 2, showing system 200 for risk assessment of a target person and recommendation. In this example, the target person is a hiker. The system 200 generally includes a plurality of data collecting modules 210, a big-data database module 220, a data analysis module 230 and a user interface module 240, wherein these modules are operatively connected to or in communication connection with each other. In this example, at least one or more of the data collecting module 210, big-data database module 220, data analysis module 230 and the user interface module 240 run on one or separate servers / computers / hardware. In this example, the data collecting modules 210 are configured to collect data from various sources and is in communication with the data analysis module 230. In this example, 210 further contains a mobile data collector 211 that is configured to obtain at least one mobile data 201 from a mobile device of the target person, an environmental data collector 212 that is configured to obtain at least one environmental data 202 from at least one environmental data source, and an open data collector 213 that is configured to obtain at least one open data 203 from the internet. In some examples, the mobile device is a transceiver and contains an application (APP) that collects and transmits mobile data associated with the target person. In some examples, the mobile device can also be any equipment for outdoor life usage with wireless communication functions. The big-data database module 220 is configured to store at least one historical data, which may be obtained by directly or indirectly receiving and processing data from 210, 236, 237 and / or 204.230 is configured to receive the data from 210 and perform data analysis. The data analysis module 230 generally contains a plurality of estimators 231-237, at least one embedding layer 238 and a classifier 239, which are in communication connection with each other sequentially. In this example, seven estimators 231, 232, 233, 234, 235, 236, 237 are provided and configured to analyze the data from 210 and / or historical data from the big-data database 220 to obtain a plurality of feature estimate data. The feature estimate data is received and analyzed by the general embedding layers 238 to obtain one or more general feature data. The classifier 239 includes a classifier AI attention-weighted risk calculation submodule 2391 and a classifier transformer-based recommendation submodule 2392, which are configured to receive and analyze at least one general feature data to obtain a risk assessment 23901 of the target person and at least one recommendation 23902. The user interface module 240 receives and outputs the risk assessment 23901 of the target person and the at least one recommendation 23902 to the user for missions, such as rescue, management, and / or surveillance. The user can also input at least one action outcome data 204 into 240, which further send the at least one action outcome data 204 to the big-data database module 220. Based on the received 204, 220 updates the at least one historical data stored in the database. 2.1 DATA AND DATA SOURCE
[0116] Still referring to FIG. 2, the system 200 collects and analyzes various types of data from various data sources. In this example, data includes mobile data, environmental data (such as weather data) and open data. In some examples, mobile data includes data associated with the target person (e.g., hiker) , such as GPS location or zone data, hiker coordinates, hiker elevation change (e.g., gain / loss) , hiker walking speed, hiker mobile phone cell-site signal intensity and hiker proficiency or amateur level (e.g., amateur, intermediate, advanced) and / or landscape score data. In some examples, the mobile data is sourced from a mobile device of the target person such as by a designated APP. Environmental data includes weather data such as temperature, humidity, fogginess, online weather data, geographical landscape data. In some examples, the environmental data is sourced from one or more environmental data source, such as the observatory and / or the weather department. Open data includes any open-sourced data such as from internet, such as from web and social media platforms. In some examples, the open data includes trending locations for hiking and locations flagged for high-risk for less experienced (amateur) hiker. 2.2 DATA COLLECTING MODULES
[0117] Still referring to FIG. 2, three data collecting modules 210 (including mobile data collector 211, environmental data collector 212 and open data collector 213) are provided in example system 200. In this example, data collecting modules (or data collectors) are computer programs that are configured to receive data from various sources. In this example, the mobile data collector 211 is a Hiker Mountain Rescue Mobile App Data Collector and is configured to receive mobile data from the mobile device of the target person (the hiker) . The mobile data collector 211 is configured to collect at least one GPS data, at least one coordinate data, at least one elevation change data, at least one signal characteristics data, and / or at least one amateur level data from the mobile device of the target hiker. The environmental data collector 212 is configured to receive environmental data such as weather data from one or more environmental data source. The open data collector is configured to receive open-sourced data from internet, such as from web and social media platforms. In some examples, the open data collector is configured to collect at least one online text that is associated with at least one hot spot location for hiking. In some embodiments, data collecting modules (or data collectors) are sensing devices and / or data gathering hardware. 2.3 BIG-DATA DATABASE MODULE
[0118] In this example, the big-data database module 220 is a database stored on one or more server, which is configured to store historical data from various sources. In some examples, the big-data database module is further configured to receive and process the at least one mobile data, the at least one environmental data, the at least one open data, the at least one hot-spot estimate feature data, and / or the at least one environmental feature data to update the at least one historical data. In some examples, the big-data database records feature vectors from one or more of the data collectors, such as those described in Examples 3-6. The big-data database 220 is configured to record Environmental-Feature-Embedding (e.g., 1x128 vector) data from environmental impact estimator 236 and / or SM-Feature-Embedding (e.g., 1x128 vector) data from the at least one artificial intelligence (AI) natural language processing (NLP) estimator 237. In some examples, historical data is pre-stored and / or updated from any data from other modules or any estimate feature data from any estimators 231-237. In some examples, the big-data database module 220 is configured to store historical data such as store at least one historical elevation change feature data, at least one historical walking speed feature data, at least one historical hiking location feature data, at least one historical hiking speed feature data, and / or at least one historical signal landscape feature data. In some examples, the big-data database module 220 is configured to record the aggregate walker speed, hiker behaviour and hiker route by receiving and processing the exhaustion feature estimate data, the reasonable location feature estimate data, the zone feature estimate data, the signal landscape feature estimate data, and the signal abnormality feature estimate data to update the historical data. In some examples, the big-data database module 220 is a cloud-based database and / or a memory. 2.4 DATA ANALYSIS MODULE
[0119] Now referring to FIG. 2, the data analysis module 230 generally contains multiple estimators 231-237, a general embedding layer 238 and a classifier 239, which are in communication connection with each other. In this example, the estimators include exhaustion level estimator 231, reasonable location estimator 232, zone estimator 233, signal landscape estimator 234, signal abnormality estimator 235, environmental impact estimator 236, at least one artificial intelligence (AI) natural language processing (NLP) estimator 237.
[0120] The exhaustion level estimator 231, reasonable location estimator 232, zone estimator 233, and signal landscape estimator 234 are generally configured to receive mobile data 201 from mobile data collector 211, and to analyze the mobile data 201.
[0121] The exhaustion level estimator 231 is configured to receive the mobile data 201 and to analyze the elevation change (e.g., gain / loss) data, the GPS data (walking speed data) and / or amateur level data (from the mobile data 201) based on the historical data such as at least one historical elevation change feature data, and / or the at least one historical walking speed data (from the big-data database 220) to obtain at least one exhaustion feature estimate data of the target person, such that the exhaustion level of the target person is estimated. In one example, it includes AI probability estimation using elevation gain / loss, walking speed and amateur level. The AI algorithm can be any probabilistic regressor include but not limit to regression, support vector machines, random forest, 2-layers fully-connected neural network layers or Cross-attention layers.
[0122] The reasonable location estimator 232 is configured to analyze at least one coordinate data (e.g., landscape data, and hiker coordinate (GPS) , walking speed (going uphill or downhill) and / or at least one amateur level data (from the mobile data collector) based on at least one historical hiking location feature data, and / or at least one historical hiking speed feature data (from the big-data database) to obtain at least one reasonable location feature estimate data, so that the target person’s reasonable location is estimated. In one example, the reasonable location estimator 232 receives data from the mobile data collector 211 (e.g. a Hiker Mountain Rescue Mobile App Data Collector) to estimate the reasonableness of location, for which the algorithm estimates the hiker location by using the previous hiker walking speed and the nearby hiking landscape.
[0123] The zone estimator 233 is configured to analyze the signal characteristic data of the mobile device site signal (from the mobile data collector) to obtain at least one zone feature estimate data. For example, it is configured to classify or estimate whether the mobile device site signal is located in the sea / urban / mountain zone. In one example, the zone estimator 233 receives data from the mobile data collector 211 (e.g. a Hiker Mountain Rescue Mobile App Data Collector) to estimate the zone including “Sea” Zone, “Urban” Zone, and “Mountain Zone” . If the hiker is at Urban Zone, it is expected that the risk level is very low. If the hiker is at Sea Zone, the time of flight over the sea should be under reasonable limit. If the hiker stays longer than expected at the same coordinate over the sea, the risk level is high. If the mountain zone is detected, other algorithms on calculating risk level is applied.
[0124] The signal landscape estimator 234 is configured to analyze the signal characteristics data of the mobile device based on the historical data such as or the historical signal landscape data (such as signal reception black / white spot and seasonal overgrowth factors) to obtain the signal landscape feature estimate data, so as to understand and estimate the landscape of the site signal of the mobile device. In one example, the signal landscape estimator 234 receives data from the mobile data collector 211 (e.g., a Hiker Mountain Rescue Mobile App Data Collector) to deduce the Black-Spot and White-Spot. If the signal is always at poor reception in the area after aggregating multiple hikers signal, the area is spotted as Black-Spot, which means that that area is normal for poor phone signal. Thus, a hiker walking near Black-Spot having poor signal cannot imply the High-Risk Level as it may be normal behaviour. If the signal is always at the good reception in the area after aggregating multiple hikers signal, the area is spotted as White-Spot, which means that the area is expected to have good phone signal. Thus, a hiker walking near White-Spot having poor signal may have a higher probability of risk of missing or out of battery. The calculation is also taken the seasonal overgrowth factor into account, such as hiker route changes due to overgrowth of grass.
[0125] The signal abnormality estimator 235 is configured to analyze the zone feature estimate data from the zone estimator 233, the signal landscape feature estimate data from the signal landscape estimator 234, and the at least one amateur level data from the mobile data collector 210 to obtain at least one signal abnormality feature estimate data, so that the risk level of having abnormal mobile device signals abnormality is estimated. For example, the risk level can be classified as no signal / unreasonable staying on a mountain / forget to stop tracking in the mobile APP. In one example, the signal abnormality estimator 235 estimates with AI for the estimation of risk level of 3 abnormal cell signals abnormality from the assisted data from the zone estimator 233 and the signal landscape estimator 234.
[0126] The environmental impact estimator 236 is configured to receive environmental data 202 from the environmental data collector 212 and to analyze the environmental data to obtain at least one environmental feature estimate data, so that the environmental impact to the risk assessment is estimated. In one example, the environmental impact estimator 236 estimates with AI algorithms the environmental impact such as temperature on the hiker speed and exhaustion level. The environmental impact estimator 236 receives Feature Vector from the mobile data collector 211 and feeds the Feature Vector into a fully-connected layer, which outputs the feature vector (Environmental-Feature-Embedding, e.g., 1x128 Vector) .
[0127] The AI NLP estimator 237 is configured to receive open data 203 from open data collector 213 and to analyze the at least one online text based on the at least one hot spot location to obtain at least one hot-spot feature estimate data, such that any hot-spot locations for hiking are estimated and retrieved and can be saved into the big-data database. In one example, via the AI NLP estimator 237, a social media hot spot location is received for high-risk location. With the AI NLP estimator 237, social media text is appended with question prompt text related to querying hot spot location. Then, the resultant text is fed into the transformer-based large language model, which outputs the numeric feature vector (SM-Feature-Embedding, 1x128 Vector) of hot spot location. In this example, the transformer-based large language model can be any open-source or proprietary large language model based on transformer structure.
[0128] The general embedding layer 238 is configured to analyze the at least one feature estimate data from estimators 231-237 to obtain at least one general feature data. In one example, the general embedding layer 238 is an embedding feature vectors layer, in which all calculated factors from the estimators 231-237 and the big-data database 220 are represented.
[0129] The classifier 239 is configured to analyze the at least one general feature data from the general embedding layer 238 to obtain the risk assessment 23901 of the target person, and at least one recommendation 23902. The classifier 239 includes a classifier AI attention-weighted risk calculation submodule 2391 and a classifier transformer-based recommendation submodule 2392. In one example, the classifier 239 is an AI probabilistic classifier. The embedding feature vectors from 238 is fed into AI probabilistic classifier to calculate the Journey Risk Level of Hiker. The classifier AI attention-weighted risk calculation submodule 2391 uses attention-weight technique (AI Attention-weighted mathematical Weighting of Factors Importance) to weight different factors to provide the risk level. The classifier transformer-based recommendation submodule 2392 includes an AI NLP model that outputs textual recommendation. In some examples, the AI algorithms in 239 can be any probabilistic regressor include but not limit to regression, support vector machines, random forest or 2-layers fully-connected neural network layers. 2.5 USER INTERFACE MODULE
[0130] The user interface module 240 is configured to output the risk assessment 23901, and the at least one recommendation 23902 from the classifier 239. The user interface module 240 is configured to further receives action outcome data 204 of the hiker, and send the action outcome data to the big-data database 220 to update the at least one historical data. In another example, the user interface module 240 displays the analysis in a police call centre. Police supervisor can take or not take actions according to the recommendations. In some examples, the user interface module 240 is a mobile application (APP) . In some examples, the user interface module 240 sends the action outcome data 204 to the big-data database 220 to fine-tune the future recommendations. EXAMPLE 3 EXHAUSTION LEVEL ESTIMATOR
[0131] Now referring to FIG. 3, showing a partial example system including an exhaustion level estimator 331. The exhaustion level estimator 331 generally contains an elevation gain (EG) estimator submodule 3311, a walking speed (WS) estimator submodule 3312, and an exhaustion level estimator Summation submodule 3313, which are in communication connection with each other. In this example, the exhaustion level estimator 331 is configured to analyze the GPS data and the at least one elevation change data, the at least one historical elevation change feature data and / or the at least one historical walking speed data to obtain at least one exhaustion feature estimate data or Exhaustion-Feature-Embedding 331301 such that the target person’s exhaustion level is estimated.
[0132] The Elevation Gain (EG) estimator submodule 3311 contains an EG estimator LSTM layer (or 2-layers fully-connected layer) 33111 and an EG estimator cross-attention layer 33112. In this example, feature vector from the mobile data collector 211 is inputted to 33111. The elevation change data (e.g., time series of elevation gain / loss) contained in the feature vector from 211 is converted to compressed feature vector using a Time Series Regressor, i.e., the EG estimator LSTM layer (or 2-layers fully-connected layers) 33111, for which the output feature vector is denoted as Current-EG-Feature-Embedding (1x128 Vector) . The Historical-EG-Feature-Embedding (1x128 Vector) , i.e. the historical elevation change feature data, is retrieved from the big-data database 220. The Current-EG-Feature-Embedding Vector and Historical-EG-Feature-Embedding Vector are fed into a cross-attention structure, i.e. the EG estimator cross-attention layer 33112, which outputs an EG-Feature-Embedding (1x128 Vector) .
[0133] The Walking Speed (WS) estimator submodule 3312 contains a WS estimator LSTM Layer (or 2-layers Fully-Connected Layer) 33121 and a WS estimator cross-attention Layer 33122. In one example, feature vector from 211 is inputted to 33121. The GPS data (e.g. time series of walking speed) contained in the feature vector from 211 is then converted to compressed feature vector using Time Series Regressor, i.e. the WS estimator LSTM Layer (or 2-layers Fully-Connected Layer) 33121, for which the output feature vector is denoted as Current-WS-Feature-Embedding (1x128 Vector) . The Historical-WS-Feature-Embedding (1x128 Vector) is retrieved from the big-data database 220. The Current-WS-Feature-Embedding Vector and Historical-WS-Feature-Embedding Vector are fed into the WS estimator cross-attention Layer 33122 that output a WS-Feature-Embedding (1x128 Vector) .
[0134] The exhaustion level estimator Summation submodule 3313 contains a concatenation layer 33131, a fully-connected layers 33132, and a normalization layer 33133. In one example, the concatenation layer 33131 is configured to receive EG-Feature-Embedding (1x128 Vector) from 3311, WS-Feature-Embedding (1x128 Vector) from 3312 and Amateur-Feature-Embedding (1x128 Vector, i.e. the amateur level data) from 211, and concatenate the three feature vectors to form the output feature vector Pre-Exhaustion-Feature-Embedding (1x384 Vector) , which is then fed into the fully-connected layers 33132 and the normalization layer 33133 that follows 33132. Then the normalization layer 33133 outputs the Exhaustion-Feature-Embedding 331301 (1x128 Vector) . EXAMPLE 4 REASONABLE LOCATION ESTIMATOR
[0135] Now referring to FIG. 4, showing a partial example system including a reasonable location estimator 432. The reasonable location estimator 432 generally contains a Hiking Location (HL) estimator submodule 4321, a Hiker Speed (HS) estimator submodule 4322, and a reasonable location estimator summation submodule 4323, which are in communication connection with each other. In this example, the reasonable location estimator 432 is configured to analyze at least one coordinate data, at least one amateur level data, at least one historical hiking location feature data, and / or at least one historical hiking speed feature data to obtain at least one reasonable location feature estimate data 432301 or Location-Feature-Embedding such that the reasonableness of the target person’s location can be estimated.
[0136] The Hiking Location (HL) estimator submodule 4321 contains an HL estimator LSTM layer (or 2-layers fully-connected layer) 43211 and an HL estimator Cross-Attention Layer 43212. In one example, feature vector from 211 is inputted to 43211. The at least one coordinate data (e.g. time series of coordinates) of the target person contained in the feature vector is converted to compressed feature vector using Time Series Regressor, i.e. the HL estimator LSTM layer (or 2-layers fully-connected layer) 43211, for which the output feature vector is denoted as Current-HL-Feature-Embedding (1x128 Vector) . The Historical-HL-Feature-Embedding (1x128 Vector) , i.e. historical hiking location feature data, is retrieved from the big-data database 220. The Current-HL-Feature-Embedding and Historical-HL-Feature-Embedding are fed into the HL estimator Cross-Attention Layer 43212 that outputs a HL-Feature-Embedding (1x128 Vector) .
[0137] The Hiker Speed (HS) estimator submodule 4322 contains an HS estimator LSTM layer (or 2-layers fully-connected layer) 43221 and an HS estimator Cross-Attention Layer 43222. In one example, feature vector from 211 is inputted to 43221. The at least one coordinate data (e.g. time series of coordinates) of the target person contained in the feature vector is converted to compressed feature-vector using Time Series Regressor, i.e. HS estimator LSTM layer (or 2-layers fully-connected layer) 43221, for which, the output feature vector is denoted as Current-HS-Feature-Embedding (1x128 Vector) . Historical-HS-Feature-Embedding (1x128 Vector) , i.e. the historical hiking speed feature data, is retrieved from the big-data database 220. The Current-HS-Feature-Embedding Vector and Historical-HS-Feature-Embedding Vector are fed into HS estimator Cross-Attention Layer 43222 that outputs a HS-Feature-Embedding (1x128 Vector) .
[0138] The reasonable location estimator summation submodule 4323 contains a concatenation layer 43231, one or more fully-connected layers 43232, and a normalization layer 43233. In one example, the concatenation layer 43231 is configured to receive the HL-Feature-Embedding (1x128 Vector) from 4321, the HS-Feature-Embedding (1x128 Vector) from 4322 and Amateur-Feature-Embedding (1x128 Vector, i.e. the amateur level data) from 211, and concatenate the three feature vector inputs to form the output feature vector Pre-Location-Feature-Embedding (1x384 Vector) , which is fed into the one or more fully-connected layers 43232. The one or more fully-connected layers 43232 then outputs the Location-Feature-Embedding (1x128 Vector) . EXAMPLE 5 ZONE ESTIMATOR AND SIGNAL LANDSCAPE ESTIMATOR
[0139] Now referring to FIG. 5, showing a partial example system including a zone estimator 533 and a signal landscape estimator 534.
[0140] In this example, the zone estimator 533 is configured to analyze the at least one signal characteristics data from the mobile data collector 211 to obtain at least one zone feature estimate data 53301 or Zone-Feature-Embedding such that whether the mobile device site signal is located in the sea / urban / mountain zone can be classified or estimated. The zone estimator 533 contains a zone estimator LSTM layer (or 2-layers fully-connected layer) 5331. Feature vector from 211 is inputted to 5331, which are in communication connection with each other. The signal characteristics data (e.g. time series of zone) will be converted to compressed feature vector using Time Series Regressor, i.e. the zone estimator LSTM layer (or 2-layers fully-connected layer) 5331, for which, the output feature vector is denoted as Zone-Feature-Embedding (1x128 Vector) .
[0141] The signal landscape estimator 534 contains a signal landscape estimator LSTM Layer (or 2-layers Fully-Connected Layer) 5341, a signal landscape estimator Cross-Attention Layer 5342, and a signal landscape estimator fully-connected layers 5343. In this example, the signal landscape estimator 534 is configured to analyze the at least one signal characteristics data and / or the historical signal landscape data to obtain at least one signal landscape feature estimate data 53401 or CL-Feature-Embedding such that the landscape of the site signal of the mobile device of the target person is estimated or understood. Feature vector from 211 is inputted to 5341, the signal characteristics data (e.g., the time series of cell signal intensity and landscape score) is converted to compressed feature vector using Time Series Regressor, i.e. the signal landscape estimator LSTM Layer (or 2-layers Fully-Connected Layer) 5341, for which, the output feature vector is denoted as Current-CL-Feature-Embedding (e.g., 1x128 Vector) . The historical signal landscape data or historical-CL-Feature-Embedding (e.g., 1x128 Vector) is retrieved from 220. The Current-CL-Feature-Embedding Vector and Historical-CL-Feature-Embedding Vector are fed into the signal landscape estimator Cross-Attention Layer 5342 that outputs a Pre-CL-Feature-Embedding (e.g., 1x128 Vector) . The output (Pre-CL-Feature-Embedding, 1x384 Vector) is fed into the signal landscape estimator fully-connected layers 5343 that output CL-Feature-Embedding (e.g., 1x128 Vector) . EXAMPLE 6 SIGNAL ABNORMALITY ESTIMATOR
[0142] Now referring to FIG. 6, showing a partial example system including a signal abnormality estimator 635. The signal abnormality estimator 635 generally contains a concatenation layer 6351, one or more fully-connected layers 6352, and a normalization layer 6353, which are in communication connection with each other. In this example, the signal abnormality estimator 635 is configured to analyze the at least one zone feature estimate data from the zone estimator 233, the at least one signal landscape feature estimate data from the signal landscape estimator 234, and the at least one amateur level data from the mobile data collector 211 to obtain at least one signal abnormality feature estimate data 63501 or Signal-Abnormality-Feature-Embedding. In this example, the concatenation layer 6351 is configured to receive feature vector Zone-Feature-Embedding (e.g., 1x128 Vector) from 233, feature vector CL-Feature-Embedding (e.g., 1x128 Vector) from 234, and feature vector Amateur-Feature-Embedding (e.g., 1x128 Vector) from 211, and concatenate the three feature vectors to form the output feature vector Pre-Signal-Abnormality-Feature-Embedding (e.g., 3x128 Vector) . The output (Pre-Signal-Abnormality -Feature-Embedding, e.g., 1x384 Vector) is fed into the one or more fully-connected layers 6352 and then the normalization layer 6353 that follows 6352. The normalization layer 6353 outputs the at least one signal abnormality feature estimate data 63501 or Signal-Abnormality-Feature-Embedding (e.g., 1x128 Vector) . EXAMPLE 7 EMBEDDING LAYER
[0143] Now referring to FIG. 7, showing a partial example system including a general embedding layer 738. In this example, the general embedding layer 738 is configured to analyze the at least one feature estimate data from the plurality of estimators 231-237 to obtain at least one general feature data 73801. The general embedding layer 738 generally contains a general embedding layer 7381, a general embedding fully-connected layer submodule 7382 and a general embedding normalization layer submodule 7383, which are in communication connection with each other.
[0144] The general embedding concatenation submodule 7381 is configured to concatenate the plurality of feature estimate data from the plurality of estimators 231-237 to obtain at least one pre-general abnormality feature data. The general embedding fully-connected layer submodule 7382 is configured to analyze the at least one pre-general abnormality feature data from the general embedding concatenation submodule to obtain at least one analyzed pre-general abnormality feature data. The general embedding normalization layer submodule 7383 is configured to analyze the at least one analyzed pre-general abnormality feature data from the general embedding fully-connected layer submodule to obtain the at least one general feature data 73801.
[0145] In one example, the general embedding concatenation submodule 7381 is configured to receive seven feature vectors HS-Feature-Embedding (e.g., 1x128 Vector) from 231, Location-Feature-Embedding (e.g., 1x128 Vector) from 232, Zone-Feature-Embedding (e.g., 1x128 Vector) from 233, CL-Feature-Embedding (1x128 Vector) from 234, Signal-Abnormality-Feature-Embedding (e.g., 1x128 Vector) from 235, Environmental-Feature-Embedding (e.g., 1x128 Vector) from 236, and SM-Feature-Embedding (e.g., 1x128 Vector) from 237, and concatenate the seven feature vector to form the output feature vector Pre-General-Feature-Embedding (e.g., 1x896 Vector) , which is fed into the general embedding fully-connected layer submodule 7382 and the general embedding normalization layer submodule 7383 that follows 7382, outputting the at least one general feature data 73801 (General-Feature-Embedding, e.g., 1x128 Vector) .
[0146] In some examples, there can be one or more general embedding layer 738 in the system. In some examples, the general embedding layer 738 contains one or more general embedding layer 7381, one or more general embedding fully-connected layer submodule 7382 and one or more general embedding normalization layer submodule 7383. EXAMPLE 8 CLASSIFER
[0147] Now referring to FIG 8, showing a partial example system including a classifier 839. In this example, the classifier 839 is configured to analyze the at least one general feature data from the at least one general embedding layer 238 to obtain the risk assessment 83901 of the target person, and at least one recommendation 83902.
[0148] The classifier generally contains a classifier AI attention-weighted risk calculation submodule 8391 and a classifier transformer-based recommendation submodule 8392, which are in communication connection with each other. The classifier AI attention-weighted risk calculation submodule 8391 generally contains a classifier self-attention layer 83911 and a classifier two-layers-fully-connected layer 83912. The classifier self-attention layer 83911 is configured to analyze the at least one general feature data from the at least one general embedding layer 238 to obtain at least one self-attention feature data. The classifier two-layers-fully-connected layer 83912 is configured to analyze the at least one self-attention feature data from the classifier self-attention layer 83911 to obtain the risk assessment 83901. The classifier transformer-based recommendation submodule 8392 generally contains a classifier transformer-based Large Language Model (LLM) 83921 that is configured to analyze the at least one general feature data from the at least one general embedding layer 238, at least one self-attention feature data from the classifier self-attention layer 83911, and the risk assessment from the classifier two-layers-fully-connected layer 83912 to obtain the at least one recommendation 83902.
[0149] In this example, the general feature data, General-Feature-Embedding (e.g., 1x128 Vector) , from 238 is fed into a Self-Attention structure, which is classifier self-attention layer 83911. The self-attention structure retrieves the score of factors SA-Feature-Embedding (e.g., 1x128 Vector) based on the relative importance. The output of the self-attention structure, SA-Feature-Embedding (e.g., 1x128 Vector) , is then fed into two layers of fully-connected layers (i.e., the classifier two-layers-fully-connected layer 83912) and return the risk level of hiker (i.e. the risk assessment 83901) . AI Textual recommendation using transformer NLP model (i.e., classifier transformer-based LLM 83921) is used to generate the textual recommendations and actionable insights (i.e., at least one recommendation 83902) . A customized prompt text is created by filling data from General-Feature-Embedding (e.g., 1x128 Vector) from 238, SA-Feature-Embedding (e.g., 1x128 Vector) and risk level from 8391. The prompt text is then fed into the transformer-based large language model. The large language model can be any open-source or proprietary transformer-based large language model fine-tuned with customized prompt as training data. EXAMPLE 9 METHODS OF TARGET PERSON RISK ASSESSMENT AND RECOMMENDATION
[0150] Now referring to FIG. 9, showing the example method 900 for target person risk assessment and recommendation, including the steps of:
[0151] Step 910: Providing any one of the system as described herein.
[0152] Step 920: Obtaining, by the mobile data collectors, at least one mobile data.
[0153] Step 930: Obtaining, by the environmental data collector, at least one environmental data.
[0154] Step 940: Obtaining, by the open data collector, at least one open data.
[0155] Step 950: Storing, by the big-data database, at least one historical data.
[0156] Step 960: Analyzing, by the plurality of estimators, the at least one mobile data, the at least one environmental data, the at least one open data and / or the at least one historical data to obtain at least one feature estimate data.
[0157] Step 970: Analyzing, by the at least one general embedding layer, the at least one feature estimate data to obtain at least one general feature data.
[0158] Step 980: Analyzing, by the classifier, at least one general feature data to obtain the risk assessment of the target person and at least one rescue recommendation.
[0159] Step 990: Outputting, by the user interface module, the risk assessment and the at least one recommendation.
[0160] In some examples, step 940 further comprises the step of collecting, by the open data collector, at least one online text that is associated with at least one hot spot location.
[0161] In some examples, step 960 further comprises the step of analyzing, by the at least one artificial intelligence (AI) natural language processing (NLP) estimator, the at least one online text from the open data collector based on the hot spot location to obtain at least one hot-spot feature estimate data.
[0162] In some examples, step 960 further comprises the step of analyzing, by the environmental impact estimator, the at least one environmental data from the environmental data collector to obtain at least one environmental feature estimate data.
[0163] In some examples, the method 900 further contains the steps of:
[0164] receiving and processing, by the big-data database module, the at least one mobile data, the at least one environmental data, the at least one open data, and / or the at least feature estimate data; and
[0165] updating, by the big-data database module, the at least one historical data.
[0166] In some examples, the method 900 further contains the steps of:
[0167] collecting, by the mobile data collector, at least one GNSS data, at least one coordinate data, at least one elevation change data, at least one signal characteristics data, and / or at least one amateur level data from the mobile device of the target hiker.
[0168] storing, by the big-data database module, at least one historical elevation change feature data, at least one historical walking speed feature data, at least one historical hiking location feature data, at least one historical hiking speed feature data, and / or at least one historical signal landscape feature data.
[0169] In some examples, the step 960 further comprises the steps of:
[0170] analyzing, by the exhaustion level estimator, the GNSS data, the at least one elevation change data, at least one amateur level data, the at least one historical elevation change feature data, and the at least one historical walking speed data to obtain at least one exhaustion feature estimate data;
[0171] analyzing, by the reasonable location estimator, at least one coordinate data, at least one amateur level data, at least one historical hiking location feature data, and / or at least one historical hiking speed feature data to obtain at least one reasonable location feature estimate data;
[0172] analyzing, by the zone estimator, the at least one signal characteristics data to obtain at least one zone feature estimate data;
[0173] analyzing, by the signal landscape estimator, the at least one signal characteristics data and / or the historical signal landscape data to obtain at least one signal landscape feature estimate data; and / or
[0174] analyzing, by the signal abnormality estimator, the at least one zone feature estimate data from the zone estimator, the at least one signal landscape feature estimate data from the signal landscape estimator, and the at least one amateur level data from the mobile data collector to obtain at least one signal abnormality feature estimate data.
[0175] In some examples, Step 970 further contains the steps of:
[0176] concatenating, by the general embedding concatenation submodule, the plurality of feature estimate data from the plurality of estimators to obtain at least one pre-general abnormality feature data;
[0177] analyzing, by the general embedding fully-connected layer submodule, the at least one pre-general abnormality feature data from the general embedding concatenation submodule to obtain at least one analyzed pre-general abnormality feature data; and / or
[0178] analyzing, by the by the general embedding normalization layer submodule, the at least one analyzed pre-general abnormality feature data from the general embedding fully-connected layer submodule to obtain the at least one general feature data.
[0179] In some examples, the step 980 further comprises the steps of:
[0180] analyzing, by the classifier self-attention layer, the at least one general feature data to obtain at least one self-attention feature data;
[0181] analyzing, by the classifier two-layers-fully-connected layer, the at least one self-attention feature data to obtain the risk assessment; and
[0182] analyzing, by the classifier transformer-based Large Language Model (LLM) , the at least one general feature data, at least one self-attention feature data, and the risk assessment to obtain the at least one recommendation.
[0183] In some examples, the method 900 further contains the steps of:
[0184] receiving, by the user interface module, at least one action outcome data of the target person;
[0185] sending, by the user interface module, the at least one action outcome data to the big-data database; and
[0186] updating, by the big-data database module, the at least one historical data. EXAMPLE 10 MODEL TRAINING
[0187] Now referring to FIG 2, training process for AI models in system 200 is described hereafter. The training process generally includes data collection phase, data annotation phase and data training phase. 10.1 TRAINING DATA
[0188] In this example, the data collection phase and annotation phase as well as the key training data for training the AI models in system 200 are described. In data collection phase, data adapter system program is written to collected data from the APP, internal system and external online API. In order to ensure the data quality, data cleansing is conducted with validation rules including the numerical range checking, categorical type validation, trend and outlier tolerance checking, missing value handling.
[0189] In order to derive important metrics, data transformation including mathematical derivation of speed, elevation gain is derived from collected raw data.
[0190] Different sources of data have different ages of data. The ages of data ranges from one years to ten years. Landscape and Map datasets are collected from government provided sources and open data sources. Environmental datasets are collected from observatory repositories and sensor data. Intelligence datasets are collected from government internal data. Mobile datasets are collected from the mobile app.
[0191] The training data is extracted from mobile APP datasets, environmental weather datasets, landscape feature datasets, map zone datasets, social media datasets and internal incidents experience datasets.
[0192] The mobile APP datasets are collected from the mobile data collector 211 or the APP installed in the mobile devices of the target person. The APP allows the target person or hikers to provide hiking information for emergency and rescue purpose. The mobile datasets include hiker historical and real time data such as journey path GIS coordinates, elevation gain, mobile cell signal, battery level, hiking proficiency level and hiking information.
[0193] Environmental weather datasets include temperature, humidity, wind information and more weather information related to the hiking locations.
[0194] Landscape datasets include geographic landscape information such as landmarks, types of terrains, location and area of resting places and dangerous zones.
[0195] Map zone datasets include map data and location of zoning data.
[0196] Social media datasets and internal incidents experience datasets are textual datasets.
[0197] In one example, on top of the pre-training LLM model, additional data of 12GB data is used.
[0198] In some examples, there are two types of datasets used. The first type of datasets include mobile APP datasets, environmental weather datasets, landscape feature datasets, map zone datasets, social media datasets and internal incidents experience datasets. These datasets provides essential knowledge and experience to mountain rescue training. These datasets provides the baseline reference for numerical calculation. The second type of datasets includes additional domain-specific knowledge and common sense textual datasets. These datasets provides additional common sense and general knowledge in hiking and mountain rescue. These datasets are specifically used to fine-tuning the LLM.
[0199] In some examples, the training dataset can be obtained from other data sources, such as public databases, academic databases, libraries, websites, sensing devices, etc.
[0200] An example of training data details is described below.
[0201] Time Series: Date-Time format: “yyyy / mm / dd hh: mm” (It is the time key for all data in time series such as GPS, Elevation, walking speed etc. )
[0202] GPS Data -Longitude: Numeric, 32-bit floating point number
[0203] GPS Data -Latitude: Numeric, 32-bit floating point number
[0204] Elevation: Numeric, 32-bit floating point number
[0205] Elevation Gain: Numeric, 32-bit floating point number
[0206] Walking Speed: Numeric, 32-bit floating point number
[0207] Amateur Level: Numeric, 32-bit floating point number
[0208] Zone: Numeric; 0 for mountain; 1 for urban; 2 for sea
[0209] Battery Level: Numeric, 32-bit floating point number
[0210] Mobile Cell Signal Strength: Numeric, 32-bit floating point number
[0211] Temperature: Numeric, 32-bit floating point number
[0212] Humidity: Numeric, 32-bit floating point number
[0213] Windspeed: Numeric, 32-bit floating point number
[0214] Wind-direction: Numeric, 32-bit floating point number (Angle of Direction)
[0215] Fogginess: Numeric, 32-bit floating point number
[0216] At-Resting-Place: Numeric, 32-bit floating point number
[0217] At-Dangerous-Zone: Numeric, 32-bit floating point number
[0218] Types-of-Terrain: Numeric, 32-bit floating point number
[0219] Exhaustion-Level: Numeric, 32-bit floating point number
[0220] Social Media Data: Text type
[0221] Online Web Data: Text type
[0222] Police Historical Hiking Case Data: Text type
[0223] Another example of training data details is described below.
[0224] GPS Data -Longitude:
[0225] A number from mobile APP
[0226] GPS Data -Latitude: Numeric, 32-bit floating point number
[0227] A number from mobile APP
[0228] Elevation: Numeric, 32-bit floating point number
[0229] A number from mobile APP
[0230] Elevation Gain: Numeric, 32-bit floating point number
[0231] A number from mobile APP
[0232] Walking Speed: Numeric, 32-bit floating point number
[0233] A number from mobile APP
[0234] Amateur Level: Numeric, 32-bit floating point number
[0235] A number from mobile APP
[0236] The number range from 0 (Beginner) to 100 (Advanced) .
[0237] Zone: Numeric: Numeric, 32-bit floating point number
[0238] A number from mobile APP
[0239] 0 for mountain; 1 for urban; 2 for sea
[0240] Battery Level: Numeric, 32-bit floating point number
[0241] A number from mobile APP
[0242] A number range from 0 to 100
[0243] Mobile Cell Signal Strength: Numeric, 32-bit floating point number
[0244] A number from mobile APP
[0245] Temperature: Numeric, 32-bit floating point number
[0246] A number from mobile APP
[0247] Humidity: Numeric, 32-bit floating point number
[0248] A number from mobile APP
[0249] Windspeed: Numeric, 32-bit floating point number
[0250] A number from mobile APP
[0251] Wind-direction: Numeric, 32-bit floating point number
[0252] A number from mobile APP
[0253] The number range from 0 to 360. It indicates the angle of direction.
[0254] Fogginess: Numeric, 32-bit floating point number
[0255] A number from mobile APP
[0256] At-Resting-Place: Numeric, 32-bit floating point number
[0257] A number from mobile APP
[0258] 1 for resting-place, 0 for non-resting-Place
[0259] At-Dangerous-Zone: Numeric, 32-bit floating point number
[0260] A number from mobile APP
[0261] 1 for dangerous-zone, 0 for non-dangerous-zone
[0262] Types-of-Terrain: Numeric, 32-bit floating point number
[0263] A number from mobile APP
[0264] The number range from 0 to 50. The number indicate a particular type of terrain.
[0265] Social Media Data: Text type
[0266] Text data-type.
[0267] Raw post data subscribed from social media provider feed.
[0268] The raw data include the post title and the post content. Those post data are encapsulated in json-formatted. The json-formatted is fed to the model directly.
[0269] Online Web Data: Text type
[0270] Text data-type.
[0271] Raw post data subscribed from social media provider feed.
[0272] The raw data include the web content title text, the web content text. Those data are encapsulated in json-formatted. The json-formatted is fed to the model directly.
[0273] Police Historical Hiking Case Data: Text type
[0274] Text data-type.
[0275] Raw post data subscribed from police historical textual data.
[0276] The raw data include the case title and case content. Those data are encapsulated in json-formatted. The json-formatted is fed to the model directly.
[0277] Training Data labelling Examples
[0278] Example Input Training Data 1:
[0279] Time Series: ( “2020 / 08 / 02 11: 08” , “2020 / 08 / 02 11: 15” , “2020 / 08 / 02 11: 30” , “2020 / 08 / 02 11: 40” , “2020 / 08 / 02 11: 58” )
[0280] GPS Data -Longitude: (114.16963, 114.16913, 114.16820, 114.16729, 114. 16729)
[0281] GPS Data -Latitude: (22.39388, 22.39372, 22.39375, 22.39360, 22.39360)
[0282] Elevation (m) : (530.5, 550.3, 570.2, 569.1, 569.3)
[0283] Elevation Gain: (10.3, 19.8, 19.9, -1.1, 0.2)
[0284] Walking Speed: (1.23, 0.95, 0.67, 0.82, 0.01)
[0285] Amateur Level: 50
[0286] Zone: (0, 0, 0, 0, 0)
[0287] Battery Level: (0.73, 0.69, 0.0.62, 0.53)
[0288] Mobile Cell Signal Strength (dB) : (-100, -103, -125, -124, -124)
[0289] Temperature: (34, 31, 30, 34, 32)
[0290] Humidity: (0.70, 0.69, 0.7, 0.7, 0.7)
[0291] Windspeed: (0.03, 0.02, 0.10, 0.07, 0.07)
[0292] Wind-direction: (273, 260, 265, 265, 266)
[0293] Fogginess: (0.07, 0.05, 0.03, 0.03, 0.05)
[0294] At-Resting-Place: (0, 0, 0, 0, 0)
[0295] At-Dangerous-Zone: (0, 0, 0, 0, 1)
[0296] Types-of-Terrain: (13, 13, 13, 13, 13) (13 means steep and unstable)
[0297] Exhaustion-Level: (83, 86, 88, 92, 95)
[0298] Social Media Data: “Our trip started from Shek Mum Au 9: 00am. …”
[0299] Online Web Data: “Shek Mum Au
[0300] Police Historical Hiking Case Data: “Case ID #184302 Date: 10-09-2010 Event: Missing and Dehydration –Shek Mum Au”
[0301] Example Output Training Data 1:
[0302] Risk Level: 92
[0303] Recommendation Output:
[0304] “Summary of Hiker: High Risk
[0305] The hiker has been moving at an unusually slow pace for the past 30 minutes.
[0306] The hiker is experiencing high exhaustion due to significant elevation gain.
[0307] The high temperature increases the hiker's risk level.
[0308] The hiker is unlikely to remain in the area given the steep and unstable landscape.
[0309] The location is a black-spot for hiking incident.
[0310] Recommendation:
[0311] [Keep Monitoring] Continue monitoring for the next 30 minutes. If the situation does not improve in 30 minutes, a mobile phone call to the hiker is advised. ”
[0312] Example Input Training Data 2:
[0313] Time Series: ( “2021 / 07 / 23 13: 10” , “2021 / 07 / 23 13: 15” , “2021 / 07 / 23 13: 32” , , “2021 / 07 / 23 14: 35” , “2021 / 07 / 23 15: 20” )
[0314] GPS Data -Longitude: (114.226320, 114.229050, 114.232929, 114.232929, 114. 232929)
[0315] GPS Data -Latitude: (22.372008, 22.372441, 22.372960, 22.372960, 22. 372960)
[0316] Elevation (m) : (420.5, 510.3, 604.2, 604.3, 604.2)
[0317] Elevation Gain: (20.3, 89.8, 93.9, 0.1, -0.1)
[0318] Walking Speed: (1.23, 1.5, 1.62, 0.0, 0.0)
[0319] Amateur Level: 30
[0320] Zone: (0, 0, 0, 0, 0)
[0321] Battery Level: (0.05, 0.03, 0.03, 0.02, 0.01)
[0322] Mobile Cell Signal Strength (dB) : (-101, -102, -105, -114, -110)
[0323] Temperature: (28, 29, 30, 28, 28)
[0324] Humidity: (0.62, 0.63, 0.67, 0.67, 0.67)
[0325] Windspeed: (0.12, 0.13, 0.12, 0.13, 0.13)
[0326] Wind-direction: (269, 268, 268, 268, 268)
[0327] Fogginess: (0.03, 0.02, 0.03, 0.02, 0.03)
[0328] At-Resting-Place: (0, 0, 0, 0, 0)
[0329] At-Dangerous-Zone: (0, 0, 1, 1, 1)
[0330] Types-of-Terrain: (12, 12, 12, 12, 12) (12 means bushland)
[0331] Exhaustion-Level: (92, 95, 96, 94, 93)
[0332] Social Media Data: “”
[0333] Online Web Data: “”
[0334] Police Historical Hiking Case Data: “Case ID #182103 Date: 10-09-2015 Event: Missing –Wong Ngau Shan”
[0335] Example Output Training Data 2:
[0336] Risk Level: 92
[0337] Recommendation Output:
[0338] “Summary of Hiker: High Risk
[0339] The hiker has been moving at an unusually slow pace and stop for the past 2 hours.
[0340] The hiker is experiencing high exhaustion due to significant elevation gain.
[0341] The hiker has entered a dangerous zone with frequent hiking accidents.
[0342] The hiker has little experience and limited ability to handle challenging or adverse situations.
[0343] The cell signal strength is normal and good for reception. However, The battery level of the cell phone is extremely low, the hiker may have difficulty in calling for help with cell phone.
[0344] Recommendation:
[0345] [Action Required] A phone call or a text message to the hiker is recommended. If the hiker is unresponsive, escalation to the mountain team for further assessment is necessary. ”
[0346] Example Input Training Data 3:
[0347] Time Series: ( “2022 / 08 / 03 17: 20” , “2022 / 08 / 03 17: 50” , “2022 / 08 / 03 18: 00” , , “2022 / 08 / 03 18: 05” , “2022 / 08 / 03 18: 10” )
[0348] GPS Data -Longitude: (114.25167, 114.25246, 114.25105, 114.25513, 114. 25311)
[0349] GPS Data -Latitude: (22.362223, 22.36130, 22.35674, 22.343920, 22.34465)
[0350] Elevation (m) : (5.1, 5.2, 7.6, 6.1, 5.1)
[0351] Elevation Gain: (0.3, 0.1, 2.4, -1.5, -1.0)
[0352] Walking Speed: (1.5, 1.72, 12.6, 24.2, 3.2)
[0353] Amateur Level: 40
[0354] Zone: (0, 0, 1, 1, 1)
[0355] Battery Level: (0.70, 0.69, 0.65, 0.65, 0.64)
[0356] Mobile Cell Signal Strength (dB) : (-82, -83, -90, -96, -90)
[0357] Temperature: (34, 32, 32, 33, 34)
[0358] Humidity: (0.51, 0.52, 0.51, 0.51, 0.51)
[0359] Windspeed: (0.22, 0.25, 0.27, 0.22, 0.22)
[0360] Wind-direction: (135, 137, 134, 134, 134)
[0361] Fogginess: (0.03, 0.02, 0.03, 0.02, 0.03)
[0362] At-Resting-Place: (1, 0, 0, 0, 0)
[0363] At-Dangerous-Zone: (0, 0, 0, 0, 0)
[0364] Types-of-Terrain: (9, 9, 9, 9, 9) (9 means flat sandy terrain)
[0365] Exhaustion-Level: (32, 28, 12, 15, 12)
[0366] Social Media Data: “”
[0367] Online Web Data: “”
[0368] Police Historical Hiking Case Data: “”
[0369] Sample Output Training Data 3:
[0370] Risk Level: 08
[0371] Recommendation Output: ” Summary of Hiker: Very Low Risk
[0372] The hiker left the resting place of mountain area. The hiker is now travelling in urban area.
[0373] Recommendation:
[0374] No Action Required. ” 10.2 TRAINING PROCESS
[0375] The model training is divided into 4 phases.
[0376] The first phase involves the fine-tuning of large language models in 237 using mentioned training datasets. Depends on the amount of data, this process typically requires computational powers of 3.5 billion Tera-flops. The computational requirement is typically equivalent to around 120 training-days of using NVIDIA GPU 4090. Typically, 192GB GPU memory is recommended for the first phase LLM fine-tuning process.
[0377] The second phase involves the training of 231-237 respectively. Depends on the amount of training data available, this process may require computational power of 3.5 billion Tera-flops. The computational requirement is typically equivalent to around 120 training-days of using NVIDIA GPU 4090.192GB GPU is used in our training.
[0378] The third phase involves the training of 2391. Depends on the amount of training data available, this process may require computational power of 3.5 billion Tera-flops. The computational requirement is typically equivalent to around 120 training-days of using NVIDIA GPU 4090.192GB GPU is used in our training.
[0379] The fourth phase involves the fine-tuning of large language models in classifier transformer-based recommendation submodule (block-15) . The computational requirement is typically around 3.5 billion Tera-flops. It is equivalent to around 120-training-days of using NVIDIA GPU 4090.
[0380] Phase 1 Training –237
[0381] Dataset: In Phase 1 of training, it involves the training of 237 only. In AI technical term, this process called “fine-tuning” of transformer model. Rather than train a model from scratch, the model is fine-tuned on top of existing pre-trained large language model. Three datasets are involved: social media dataset, online web dataset, police historical hiking case dataset.
[0382] Aim of fine-tuning: On top of the existing large language transformer model, the ‘fine-tuning’ process allows the model to learn more specific (in this case “hiking” ) knowledge and keywords for data extraction.
[0383] Key algorithm: There is no restriction in the fine-tuning algorithm such as LoRA or prompt tuning. In our case, utilize “Low-Rank Adaptation (LoRA) ” fine-tuning method. It adds low-rank matrices to existing model weights with no new layers.
[0384] Parameters: The parameters needs to tune based on different datasets. In particular, we use “Rank=8” , “lora_dropout=0.1” , ” bias=none” parameters. Learning rate is adaptive starting from 1e-4 and AdamW as optimizer.
[0385] Phase 2 Training –231
[0386] Dataset to 231: Data set includes the time series of [1] GPS Data –Longitude [2] GPS Data –Latitude [3] Elevation [4] Elevation Gain [5] Walking Speed [6] Amateur Level.
[0387] Training Methodology: 231 (exhaustion level estimator; Block 4) is trained separately in this phase before integrating back to the whole architecture. In order to train the block 4, a single RELU-fully-connected layer is appended with a single numeric output value (the estimated exhaustion level) . We called the training block as “Block4+RELU+FC” module.
[0388] During training, datasets are fed into the “Block4+RELU+FC” layer. The output exhaustion level will be predicted. Training is conducted via backward propagation mechanism by calculating the errors from ground-truth exhaustion level. After block-4 is trained, the appended “RELU+FC” layer is removed. The block-4 is integrated back to the original architecture.
[0389] Training Data Output:
[0390] The ground-truth exhaustion level is calculated using different method such as Pandolf Model, Metabolic Equivalents, Hiking Difficulty Score, Heart-Rate-Monitoring or rate of Perceived Exertion (RPE) . In our case, hiking difficulty score is used “hiking difficulty score= (Distance in miles x2) + (Elevation Gain in feet / 1000) ” in the early phase of training as the data is easier to collect.
[0391] Parameters: Learning rate is adaptive starting from 1e-4 and AdamW as optimizer.
[0392] Aim of Training: The objective of the training is to find the compressed essential feature embedding that imply the exhaustion level of hiker.
[0393] Phase 2 Training –232
[0394] Dataset to 232: Data set includes the time series of [1] GPS Data –Longitude [2] GPS Data –Latitude [3] Elevation [4] Walking Speed [5] Amateur Level.
[0395] Training Methodology: 232 (reasonable location estimator; Block 5) is trained separately in this phase before integrating back to the whole architecture. In order to train the block 5, a single RELU-fully-connected layer is appended with a two numeric output value (the estimated location latitude and longitude) . We called the training block as “Block5+RELU+FC” module.
[0396] During training, datasets are fed into the “Block5+RELU+FC” layer. The output location coordinates will be predicted. Training is conducted via backward propagation mechanism by calculating the errors from ground-truth coordinates. After block-5 is trained, the appended “RELU+FC” layer is removed. The block-5 is integrated back to the original architecture.
[0397] Training Data Output:
[0398] The ground-truth coordinates obtained directly by applying sliding time window in the historical data.
[0399] Parameters: Learning rate is adaptive starting from 1e-4 and AdamW as optimizer.
[0400] Aim of Training: The objective of the training is to find the compressed essential feature embedding that imply the potential whereabout of hiker.
[0401] Phase 2 Training –233
[0402] Dataset to 233: Data set includes the time series of [1] zone.
[0403] Training Methodology: 233 (zone estimator; Block 6) is trained separately in this phase before integrating back to the whole architecture. In order to train the block 6, a single RELU-fully-connected layer is appended with a two numeric output value (the estimated zone) . We called the training block as “Block6+RELU+FC” module.
[0404] During training, datasets are fed into the “Block6+RELU+FC” layer. The output zone will be predicted. Training is conducted via backward propagation mechanism by calculating the errors from ground-truth zone. After block-6 is trained, the appended “RELU+FC” layer is removed. The block-6 is integrated back to the original architecture.
[0405] Training Data Output:
[0406] The ground-truth zone obtained directly by applying sliding time window in the historical data.
[0407] Parameters: Learning rate is adaptive starting from 1e-4 and AdamW as optimizer.
[0408] Aim of Training: The objective of the training is to find the compressed essential feature embedding that imply the potential zone attributes of hiker location.
[0409] Phase 2 Training –234
[0410] Dataset to 234: Data set includes the time series of [1] Mobile Cell Signal Strength [2] Battery Level [3] At-Resting-Place [4] At-Dangerous-Zone [5] Type-of-Terrain.
[0411] Training Methodology: 234 (signal landscape estimator; Block 7) is trained separately in this phase before integrating back to the whole architecture. In order to train the 234, a single RELU-fully-connected layer is appended with a single numeric output value (the estimated landscape-difficulty) . We called the training block as “Block7+RELU+FC” module.
[0412] During training, datasets are fed into the “Block7+RELU+FC” layer. The output landscape-difficulty will be predicted. Training is conducted via backward propagation mechanism by calculating the errors from ground-truth. After 237 is trained, the appended “RELU+FC” layer is removed. The 237 is integrated back to the original architecture.
[0413] Training Data Output:
[0414] The ground-truth landscape-difficulty is annotated manually in each landscape region.
[0415] Parameters: Learning rate is adaptive starting from 1e-4 and AdamW as optimizer.
[0416] Aim of Training: The objective of the training is to find the compressed essential feature embedding that imply the landscape difficulty.
[0417] Phase 2 Training –233, 234, and 235
[0418] Dataset to 233: Data set includes the time series of [1] zone.
[0419] Dataset to 234: Data set includes the time series of [1] Mobile Cell Signal Strength [2] Battery Level [3] At-Resting-Place [4] At-Dangerous-Zone [5] Type-of-Terrain.
[0420] Training Methodology: 233, 234 and 235 (zone estimator, signal landscape estimator and signal abnormality estimator; Blocks 6, 7, 8) are trained separately in this phase before integrating back to the whole architecture. Parameters in 233 and 234 are frozen. Only parameters in block 8 is trainable. In order to train the structure, a single RELU-fully-connected layer is appended with a single numeric output value (the estimated signal abnormality) . We called the training block as “Block678+RELU+FC” module.
[0421] During training, datasets are fed into the “Block678+RELU+FC” layer. The output will be predicted. Training is conducted via backward propagation mechanism by calculating the errors from ground-truth. After 235 is trained, the appended “RELU+FC” layer is removed. The block-8 is integrated back to the original architecture.
[0422] Training Data Output:
[0423] The ground-truth signal abnormality is annotated manually.
[0424] Parameters: Learning rate is adaptive starting from 1e-4 and AdamW as optimizer.
[0425] Aim of Training: The objective of the training is to find the compressed essential feature embedding that imply the signal abnormality.
[0426] Phase 2 Training –236
[0427] Dataset to 236: Data set includes the time series of [1] Temperature [2] Humidity [3] Windspeed [4] Wind-direction [5] Fogginess.
[0428] Training Methodology: 236 (environmental impact estimator; Block 9) is trained separately in this phase before integrating back to the whole architecture. In order to train the 236, a single RELU-fully-connected layer is appended with a single numeric output value (the estimated environment-unfavourable-factor) . We called the training block as “Block9+RELU+FC” module.
[0429] During training, datasets are fed into the “Block9+RELU+FC” layer. The output zone will be predicted. Training is conducted via backward propagation mechanism by calculating the errors from ground-truth. After 236 is trained, the appended “RELU+FC” layer is removed. The 234 is integrated back to the original architecture.
[0430] Training Data Output:
[0431] The ground-truth environment-unfavourable-factor is annotated manually in the case data.
[0432] Parameters: Learning rate is adaptive starting from 1e-4 and AdamW as optimizer.
[0433] Aim of Training: The objective of the training is to find the compressed essential feature embedding that imply the environmental difficulty.
[0434] Phase 3 Training –2391
[0435] Dataset to the whole architecture: Datasets of all cases is fed into 231, 232, 233, 234, 236, and 237.
[0436] Training Methodology:
[0437] Only parameters of 2391 is trainable. Parameters from other blocks are frozen.
[0438] During training, datasets of all cases are fed into 231, 232, 233, 234, 236, and 237.
[0439] The output risk level will be predicted. Training is conducted via backward propagation mechanism by calculating the errors from ground-truth.
[0440] Training Data Output:
[0441] The ground-truth risk level is annotated manually in the case data.
[0442] Parameters: Learning rate is adaptive starting from 1e-4 and AdamW as optimizer.
[0443] Aim of Training: The objective of the training is to estimate the risk level of the hiker.
[0444] Phase 4 Training –2392
[0445] Dataset: In Phase 4 of training, it involves the training of 2392 only. Parameters from all other blocks are frozen. In AI technical term, this process called “fine-tuning” of transformer model. Rather than train a model from scratch, the model is fine-tuned on top of existing pre-trained large language model. All dataset are used.
[0446] Aim of fine-tuning: On top of the existing large language transformer model, the ‘fine-tuning’ process allows the model to learn more specific (in this case “recommendation” ) knowledge.
[0447] Key algorithm: There is no restriction in the fine-tuning algorithm such as LoRA or prompt tuning. In our case, we utilize “Low-Rank Adaptation (LoRA) ” fine-tuning method. It adds low-rank matrices to existing model weights with no new layers.
[0448] Parameters: The parameters needs to tune based on different datasets. In particular, we use “Rank=8” , “lora_dropout=0.1” , ” bias=none” parameters. Learning rate is adaptive starting from 1e-4 and AdamW as optimizer.
[0449] Model Output 2391
[0450] Risk Level of Hiker: Numeric, 32-bit floating point number
[0451] Numeric value from 0 to 100. The number represent the magnitude of risk. 0 indicates the lowest risk and 100 indicates the highest risk.
[0452] Model Output from 2392
[0453] Textual Recommendations &Action Insights
[0454] “Recommendation” output: Text data type. Typically range from 100 to 300 words. 10.3 TRAINING EFFECTS USING THE DESIRED DATA
[0455] In some examples, the designed AI structures in 231-236 utilize the tailored use of data with AI technical structures called cross-attention structures with LSTM. The power of cross-attention mechanism provides an important mathematical focus on the required data among all information. Fusion with the processing of time-series module such as LSTM and Fully-Connected layers, our validation test datasets increase the average retrieval precision to 90%from 70%baseline. With the assistance of historical big data records in 220, the personalization effect of hiker estimation improves by 5.3%. The human feedback loop with the actual outcome provide continuous improvements of the system, with roughly 1-3%for every 1000 feedbacks.
[0456] The above embodiments are described by way of example only. Many variations are possible without departing from the scope of the invention as defined in the appended claims.
[0457] For clarity of explanation, in some instances the present technology has been presented as including individual functional blocks including functional blocks comprising devices, device components, steps or routines in a method embodied in software, or combinations of hardware and software.
[0458] Methods according to the above-described examples can be implemented using computer-executable instructions that are stored or otherwise available from computer readable media. Such instructions can comprise, for example, instructions and data which cause or otherwise configure a general-purpose computer, special purpose computer, or special purpose processing device to perform a certain function or group of functions. Portions of computer resources used can be accessible over a network. The computer executable instructions may be, for example, binaries, intermediate format instructions such as assembly language, firmware, or source code. Examples of computer-readable media that may be used to store instructions, information used, and / or information created during methods according to described examples include magnetic or optical disks, flash memory, Universal Serial Bus (USB) devices provided with non-volatile memory, networked storage devices, and so on.
[0459] Devices implementing methods according to these disclosures can comprise hardware, firmware and / or software, and can take any of a variety of form factors. Typical examples of such form factors include laptops, smart phones, small form factor personal computers, personal digital assistants, and so on. Functionality described herein also can be embodied in peripherals or add-in cards. Such functionality can also be implemented on a circuit board among different chips or different processes executing in a single device, by way of further example.
[0460] The instructions, media for conveying such instructions, computing resources for executing them, and other structures for supporting such computing resources are means for providing the functions described in these disclosures.
[0461] Although a variety of examples and other information was used to explain aspects within the scope of the appended claims, no limitation of the claims should be implied based on particular features or arrangements in such examples, as one of ordinary skill would be able to use these examples to derive a wide variety of implementations. Further and although some subject matter may have been described in language specific to examples of structural features and / or method steps, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to these described features or acts. For example, such functionality can be distributed differently or performed in components other than those identified herein. Rather, the described features and steps are disclosed as examples of components of systems and methods within the scope of the appended claims.
[0462] It is to be understood that any feature described in relation to any one example may be used alone, or in combination with other features described, and may also be used in combination with any features of any other of the examples, or any combination of any other of the examples.
[0463] The exemplary embodiments of the present invention are thus fully described. Although the description referred to particular embodiments, it will be clear to one skilled in the art that the present invention may be practiced with variation of these specific details. Hence this invention should not be construed as limited to the embodiments set forth herein.
[0464] Methods discussed within different figures can be added to or exchanged with methods in other figures. Further, specific numerical data values (such as specific quantities, numbers, categories, etc. ) or other specific information should be interpreted as illustrative for discussing example embodiments. Such specific information is not provided to limit example embodiment.
Claims
1.A target person risk assessment and recommendation system, comprising:a plurality of data collecting modules;a big-data database module;a data analysis module; anda user interface module,wherein the plurality of data collecting modules, the big-data database module, the data analysis module and user interface module are operatively connected with each other,wherein the plurality of data collecting modules comprise:a mobile data collector that is configured to obtain at least one mobile data from a mobile device of the target person;an environmental data collector that is configured to obtain at least one environmental data from at least one environmental data source; andan open data collector that is configured to obtain at least one open data from internet,wherein the big-data database module is configured to store at least one historical data,wherein the data analysis module comprises:a plurality of estimators that are configured to analyze the at least one mobile data, the at least one environmental data, the at least one open data from the plurality of data collecting modules, and / or the at least one historical data from the big-data database to obtain a plurality of feature estimate data;at least one general embedding layer that is configured to analyze the at least one feature estimate data from the plurality of estimators to obtain at least one general feature data; anda classifier that is configured to analyze the at least one general feature data from the at least one general embedding layer to obtain the risk assessment of the target person, and at least one recommendation, and wherein the user interface module is configured to output the risk assessment, and the at least one recommendation from the classifier.2.The system of claim 1,wherein the open data collector is configured to collect at least one online text that is associated with at least one hot spot location,and wherein the plurality of estimators comprise:at least one artificial intelligence (AI) natural language processing (NLP) estimator that is configured to analyze the at least one online text from the open data collector based on the at least one hot spot location to obtain at least one hot-spot feature estimate data; and / oran environmental impact estimator that is configured to analyze the at least one environmental data from the environmental data collector to obtain at least one environmental feature estimate data.3.The system of claim 1, wherein the big-data database module is configured to receive and process the at least one mobile data from the mobile data collector, the at least one environmental data from the environmental data collector, the at least one open data from the open data collector, and / or the plurality of feature estimate data from the plurality of estimators to update the at least one historical data.4.The system of claim 1,wherein the mobile data collector is configured to collect at least one GNSS data, at least one coordinate data, at least one elevation change data, at least one signal characteristics data, and at least one amateur level data from the mobile device of the target hiker,wherein the big-data database module is configured to store at least one historical elevation change feature data, at least one historical walking speed feature data, at least one historical hiking location feature data, at least one historical hiking speed feature data, and / or at least one historical signal landscape feature data,wherein the plurality of estimators comprise:an exhaustion level estimator that is configured to analyze the GNSS data, the at least one elevation change data, at least one amateur level data, the at least one historical elevation change feature data, and the at least one historical walking speed data to obtain at least one exhaustion feature estimate data;a reasonable location estimator that is configured to analyze the at least one coordinate data, at least one amateur level data, at least one historical hiking location feature data, and / or at least one historical hiking speed feature data to obtain at least one reasonable location feature estimate data;a zone estimator that is configured to analyze the at least one signal characteristics data to obtain at least one zone feature estimate data;a signal landscape estimator that is configured to analyze the at least one signal characteristics data and / or the historical signal landscape data to obtain at least one signal landscape feature estimate data; and / or a signal abnormality estimator that is configured to analyze the at least one zone feature estimate data from the zone estimator, the at least one signal landscape feature estimate data from the signal landscape estimator, and the at least one amateur level data from the mobile data collector to obtain at least one signal abnormality feature estimate data.5.The system of claim 1, wherein the at least one general embedding layer comprises:a general embedding concatenation submodule that is configured to concatenate the plurality of feature estimate data from the plurality of estimators to obtain at least one pre-general abnormality feature data;a general embedding fully-connected layer submodule that is configured to analyze the at least one pre-general abnormality feature data from the general embedding concatenation submodule to obtain at least one analyzed pre-general abnormality feature data; anda general embedding normalization layer submodule that is configured to analyze the at least one analyzed pre-general abnormality feature data from the general embedding fully-connected layer submodule to obtain the at least one general feature data.6.The system of claim 1, wherein the classifier comprises:a classifier AI attention-weighted risk calculation submodule comprising:a classifier self-attention layer that is configured to analyze the at least one general feature data from the at least one general embedding layer to obtain at least one self-attention feature data; anda classifier two-layers-fully-connected layer that is configured to analyze the at least one self-attention feature data from the classifier self-attention layer to obtain the risk assessment; anda classifier transformer-based recommendation submodule comprising:a classifier transformer-based Large Language Model (LLM) that is configured to analyze the at least one general feature data from the at least one general embedding layer, at least one self-attention feature data from the classifier self-attention layer, and the risk assessment from the classifier two-layers-fully-connected layer to obtain the at least one recommendation.7.The system of claim 1, wherein the user interface module is configured to further receives at least one action outcome data of the hiker, and wherein the user interface is configured to send the at least one action outcome data to the big-data database to update the at least one historical data.8.A risk assessment of a target person and recommendation system, comprising:a mobile data collector configured to obtain at least one mobile data comprising at least one GNSS data, at least one coordinate data, at least one elevation change data, at least one signal characteristics data, and at least one amateur level data from the mobile device of the target hiker from the mobile device of the target person;an environmental data collector configured to obtain at least one environmental data from at least one environmental data source;an open data collector configured to obtain at least one open data comprising at least one online text that is associated with at least one hot spot location from internet;a big-data database configured to store at least one historical data comprising at least one historical elevation change feature data, at least one historical walking speed feature data, at least one historical hiking location feature data, at least one historical hiking speed feature data, and at least one historical signal landscape feature data;at least one artificial intelligence (AI) natural language processing (NLP) estimator that is configured to analyze the at least one online text from the open data collector based on the at least one hot spot location to obtain at least one hot-spot feature estimate data;an environmental impact estimator that is configured to analyze the at least one environmental data from the environmental data collector to obtain at least one environmental feature estimate data;an exhaustion level estimator that is configured to analyze the GNSS data, the at least one elevation change data, at least one amateur level data, the at least one historical elevation change feature data, and the at least one historical walking speed data to obtain at least one exhaustion feature estimate data;a reasonable location estimator that is configured to analyze the at least one coordinate data, at least one amateur level data, at least one historical hiking location feature data, and at least one historical hiking speed feature data to obtain at least one reasonable location feature estimate data;a zone estimator that is configured to analyze the at least one signal characteristics data to obtain at least one zone feature estimate data;a signal landscape estimator that is configured to analyze the at least one signal characteristics data and the historical signal landscape data to obtain at least one signal landscape feature estimate data;a signal abnormality estimator that is configured to analyze the at least one zone feature estimate data from the zone estimator, the at least one signal landscape feature estimate data from the signal landscape estimator, and the at least one amateur level data from the mobile data collector to obtain at least one signal abnormality feature estimate data;a general embedding concatenation submodule that is configured to concatenate the at least one hot-spot feature estimate data from the at least one artificial intelligence (AI) natural language processing (NLP) estimator, the at least one environmental feature estimate data from the environmental impact estimator, the at least one exhaustion feature estimate data from the exhaustion level estimator, the at least one reasonable location feature estimate data from the reasonable location estimator, the at least one zone feature estimate data from the zone estimator, the at least one signal landscape feature estimate data from the signal landscape estimator, and the at least one signal abnormality feature estimate data from the signal abnormality estimator to obtain at least one pre-general abnormality feature data;a general embedding fully-connected layer submodule that is configured to analyze the at least one pre-general abnormality feature data from the general embedding concatenation submodule to obtain at least one analyzed pre-general abnormality feature data;a general embedding normalization layer submodule that is configured to analyze the at least one analyzed pre-general abnormality feature data from the general embedding fully-connected layer submodule to obtain the at least one general feature data;a classifier self-attention layer that is configured to analyze the at least one general feature data from the general embedding normalization layer submodule to obtain at least one self-attention feature data;a classifier two-layers-fully-connected layer that is configured to analyze the at least one self-attention feature data from the self-attention layer submodule to obtain the risk assessment;a classifier transformer-based Large Language Model (LLM) submodule that is configured to analyze the at least one general feature data from the general embedding normalization layer submodule, at least one self-attention feature data from the classifier self-attention layer, and the risk assessment from the classifier two-layers-fully-connected layer to obtain the at least one recommendation;a user interface configured to output the risk assessment from the two-layers-fully-connected layers submodule, and the at least one recommendation from the classifier transformer-based Large Language Model (LLM) submodule,wherein the mobile data collector, the environmental data collector, the open data collector, the big-data database, the at least one artificial intelligence (AI) natural language processing (NLP) estimator, the environmental impact estimator, the exhaustion level estimator, the reasonable location estimator, the zone estimator, the signal landscape estimator, the signal abnormality estimator, the general embedding concatenation submodule, the general embedding fully-connected layer submodule, the general embedding normalization layer submodule, the classifier self-attention layer submodule, the classifier two-layers-fully-connected layer, the classifier transformer-based Large Language Model (LLM) submodule, and the user interface are operatively connected with each other,wherein the big-data database is configured to receive and process at least one mobile data from the mobile data collector, the at least one environmental data from the environmental data collector, the at least at least one open data from the open data collector, the at least one hot-spot feature estimate data from the at least one artificial intelligence (AI) natural language processing (NLP) estimator, the at least one environmental feature estimate data from the environmental impact estimator, the at least one exhaustion feature estimate data from the exhaustion level estimator, the at least one reasonable location feature estimate data from the reasonable location estimator, the at least one zone feature estimate data from the zone estimator, the at least one signal landscape feature estimate data from the signal landscape estimator, and the at least one signal abnormality feature estimate data from the signal abnormality estimator to update the at least one historical data,wherein the user interface module is configured to receive at least one action outcome data of the target person,and wherein the user interface is configured to send the at least one action outcome data to the big-data database to update the at least one historical data.9.A target person risk assessment and recommendation method, comprising the steps of:(i) providing a system as claimed in any one of the preceding claims;(ii) obtaining, by the mobile data collector, at least one mobile data;(iii) obtaining, by the environmental data collector, at least one environmental data;(iv) obtaining, by the open data collector, at least one open data;(v) storing, by the big-data database, at least one historical data;(vi) analyzing, by the plurality of estimators, the at least one mobile data, the at least one environmental data, the at least one open data and / or the at least one historical data to obtain at least one feature estimate data;(vii) analyzing, by the at least one general embedding layer, the at least one feature estimate data to obtain at least one general feature data;(viii) analyzing, by the classifer, at least one general feature data to obtain the risk assessment of the target person and at least one recommendation; and(ix) outputting, by the user interface module, the risk assessment and the at least one recommendation.10.The method of claim 9,wherein the step (iv) obtaining at least one open data from Internet comprises the step of collecting, by the open data collector, at least one online text that is associated with at least one hot spot location,wherein the step (vi) analyzing the at least one mobile data, the at least one environmental data, the at least one open data from the plurality of data collectors, and / or the at least one historical data comprises the steps of:analyzing, by the at least one artificial intelligence (AI) natural language processing (NLP) estimator, the at least one online text from the open data collector based on the hot spot location to obtain at least one hot-spot feature estimate data; and / oranalyzing, by the environmental impact estimator, the at least one environmental data from the environmental data collector to obtain at least one environmental feature estimate data.11.The method of claim 9, further comprising the steps of:receiving and processing, by the big-data database module, the at least one mobile data, the at least one environmental data, the at least one open data, and / or the at least feature estimate data; andupdating, by the big-data database module, the at least one historical data.12.The method of claim 9, further comprising the steps of:collecting, by the mobile data collector, at least one GNSS data, at least one coordinate data, at least one elevation change data, at least one signal characteristics data, and at least one amateur level data from the mobile device of the target hiker;storing, by the big-data database module, at least one historical elevation change feature data, at least one historical walking speed feature data, at least one historical hiking location feature data, at least one historical hiking speed feature data, and / or at least one historical signal landscape feature data,wherein the step (vi) analyzing the at least one mobile data, the at least one environmental data, the at least one open data from the plurality of data collectors, and / or the at least one historical data further comprises the steps of:analyzing, by the exhaustion level estimator, the GNSS data, the at least one elevation change data, at least one amateur level data, the at least one historical elevation change feature data, and the at least one historical walking speed data to obtain at least one exhaustion feature estimate data;analyzing, by the reasonable location estimator, at least one coordinate data, at least one amateur level data, at least one historical hiking location feature data, and / or at least one historical hiking speed feature data to obtain at least one reasonable location feature estimate data;analyzing, by the zone estimator, the at least one signal characteristics data to obtain at least one zone feature estimate data;analyzing, by the signal landscape estimator, the at least one signal characteristics data and / or the historical signal landscape data to obtain at least one signal landscape feature estimate data; and / oranalyzing, by the signal abnormality estimator, the at least one zone feature estimate data from the zone estimator, the at least one signal landscape feature estimate data from the signal landscape estimator, and the at least one amateur level data from the mobile data collector to obtain at least one signal abnormality feature estimate data.13.The method of claim 9, wherein the step (vii) analyzing the at least one feature estimate data comprises the steps of:concatenating, by the general embedding concatenation submodule, the plurality of feature estimate data from the plurality of estimators to obtain at least one pre-general abnormality feature data;analyzing, by the general embedding fully-connected layer submodule, the at least one pre-general abnormality feature data from the general embedding concatenation submodule to obtain at least one analyzed pre-general abnormality feature data; andanalyzing, by the general embedding normalization layer submodule, the at least one analyzed pre-general abnormality feature data from the general embedding fully-connected layer submodule to obtain the at least one general feature data.14.The method of claim 9, wherein the step (viii) analyzing at least one general feature data comprises the steps of:analyzing, by the classifier self-attention layer, the at least one general feature data to obtain at least one self-attention feature data;analyzing, by the classifier two-layers-fully-connected layer, the at least one self-attention feature data to obtain the risk assessment; andanalyzing, by the classifier transformer-based Large Language Model (LLM) , the at least one general feature data, at least one self-attention feature data, and the risk assessment to obtain the at least one recommendation.15.The method of claim 9, further comprising the steps of:receiving, by the user interface module, at least one action outcome data of the target person;sending, by the user interface module, the at least one action outcome data to the big-data database; andupdating, by the big-data database module, the at least one historical data.16.A method of assessing risk of a target person and providing recommendation, comprising the steps of:providing a system as claimed in claim 1;obtaining, by the mobile data collector, at least one mobile data comprising at least one GNSS data, at least one coordinate data, at least one elevation change data, at least one signal characteristics data, and at least one amateur level data from the mobile device of the target hiker from the mobile device of the target person;obtaining, by the environmental data collector, at least one environmental data from at least one environmental data source;obtaining, by the open data collector, at least one open data comprising at least one online text that is associated with at least one hot spot location from internet;storing, by the big-data database, at least one historical data comprising at least one historical elevation change feature data, at least one historical walking speed feature data, at least one historical hiking location feature data, at least one historical hiking speed feature data, and at least one historical signal landscape feature data;analyzing, by the at least one artificial intelligence (AI) natural language processing (NLP) estimator, the at least one online text based on the at least one hot spot location to obtain at least one hot-spot feature estimate data;analyzing, by the environmental impact estimator, the at least one environmental data to obtain at least one environmental feature estimate data;analyzing, by the exhaustion level estimator, the GNSS data, the at least one elevation change data, at least one amateur level data, the at least one historical elevation change feature data, and the at least one historical walking speed data to obtain at least one exhaustion feature estimate data;analyzing, by the reasonable location estimator, the at least one coordinate data, at least one amateur level data, at least one historical hiking location feature data, and at least one historical hiking speed feature data to obtain at least one reasonable location feature estimate data;analyzing, by the zone estimator, the at least one signal characteristics data to obtain at least one zone feature estimate data;analyzing, by the signal landscape estimator, the at least one signal characteristics data and the historical signal landscape data to obtain at least one signal landscape feature estimate data;analyzing, by the signal abnormality estimator, the at least one zone feature estimate data from the zone estimator, the at least one signal landscape feature estimate data from the signal landscape estimator, and the at least one amateur level data from the mobile data collector to obtain at least one signal abnormality feature estimate data.concatenating, by the general embedding concatenation submodule, the at least one hot-spot feature estimate data, the at least one environmental feature estimate data, the at least one exhaustion feature estimate data, the at least one reasonable location feature estimate data, the at least one zone feature estimate data, the at least one signal landscape feature estimate data, and the at least one signal abnormality feature estimate data to obtain at least one pre-general abnormality feature data;analyzing, by the general embedding fully-connected layer submodule, the at least one pre-general abnormality feature data to obtain at least one analyzed pre-general abnormality feature data;analyzing, by the a general embedding normalization layer submodule, the at least one analyzed pre-general abnormality feature data to obtain the at least one general feature data;analyzing, by the classifier self-attention layer, the at least one general feature data to obtain at least one self-attention feature data;analyzing, by the classifier two-layers-fully-connected layers submodule, the at least one self-attention feature data to obtain the risk assessment;analyzing, by the classifier transformer-based Large Language Model (LLM) submodule, the at least one general feature data, at least one self-attention feature data and the risk assessment to obtain the at least one recommendation;outputting, by the user interface, the risk assessment from the classifier two-layers-fully-connected layer, and the at least one recommendation from the classifier transformer-based Large Language Model (LLM) submodule;receiving and processing, by the big-data database, at least one mobile data, the at least one environmental data, the at least at least one open, the at least one hot-spot feature estimate data, the at least one environmental feature estimate data, the at least one exhaustion feature estimate data, the at least one reasonable location feature estimate data, the at least one zone feature estimate data, the at least one signal landscape feature estimate data, and the at least one signal abnormality feature estimate data to update the at least one historical data;receiving, by the user interface, at least one action outcome data of the target person;sending, by the user interface, the at least one action outcome data to the big-data database; andupdating, by the big-data database, the at least one historical data.
Citation Information
Patent Citations
A tourism data analysis system and method
CN109697214A
Plateau mountain tourism safety risk early warning method based on reinforcement learning
CN113361665A
Ecological tourism real-time navigation system based on holographic map
CN114817438A
Digital cultural tourism management system based on multi-source data analysis
CN116805183A
Tour operations predictive risk management
US20200019903A1