A risk assessment and early warning method and system based on big data analysis

Abstract

The present application relates to the technical field of risk early warning, in particular to a risk assessment and early warning method and system based on big data analysis; multi-source heterogeneous data is collected through a community internet of things sensing layer, a management service layer and a social network layer; a dynamic mapping relationship between a virtual social account and a physical real estate unit is established by using a natural language processing and address matching algorithm, the multi-source heterogeneous data is cleaned and anchored to a community geographic grid to form a feature data space containing space-time attributes; a community risk identification model is constructed to analyze the feature data space, identify risk abnormal points in the feature data space, and convert the risk abnormal points into a continuously distributed risk potential field; the potential value distribution, spatial gradient vector and type of the risk event of the risk potential field are identified; and the risk evolution trajectory is predicted, and when the predicted risk potential exceeds a safety threshold, an early warning signal containing the coordinates, type and diffusion range of the risk source is generated.

Description

Technical Field

[0001] This invention relates to the field of risk warning technology, specifically to a risk assessment and warning method and system based on big data analysis. Background Technology

[0002] Community governance is becoming increasingly complex, involving multiple dimensions such as facility maintenance, environmental order, and neighborly relations. In order to improve management efficiency, many communities have introduced information-based smart community management systems, using sensors to monitor environmental conditions or processing residents' requests through electronic work orders.

[0003] However, significant technical bottlenecks still exist in existing community risk management and decision support practices.

[0004] Community data is highly fragmented. Physical sensing data from environmental monitoring equipment, service records from property management systems, and social media sentiment data from resident communication groups are often stored in independent systems, forming data silos and making it difficult for managers to conduct comprehensive cross-dimensional analysis. Secondly, with the widespread adoption of mobile internet, a large number of potential risk signals are hidden in the unstructured communication text of virtual social networks. This information usually lacks precise geolocation identifiers, making it impossible for management systems to accurately link online virtual identities with offline physical property units, resulting in a disconnect between risk detection and on-site response.

[0005] Therefore, how to break down the barriers between multi-source data, solve the problem of accurately mapping virtual social accounts to physical real estate units, and establish an analysis mechanism that can quantify the spatial aggregation and dynamic diffusion trends of risks in order to achieve a shift from passive response to proactive early warning are technical problems that urgently need to be solved in the field of community governance and management decision-making.

[0006] To address this, a risk assessment and early warning method and system based on big data analysis is proposed. Summary of the Invention

[0007] The purpose of this invention is to provide a risk assessment and early warning method and system based on big data analysis. By identifying the potential energy distribution, spatial gradient vector, and type of risk event of the risk potential energy field, and predicting the risk evolution trajectory, an early warning signal containing the risk source coordinates, type, and diffusion range is generated when the predicted risk potential energy exceeds the safety threshold.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A risk assessment and early warning method based on big data analysis:

[0010] Collect multi-source heterogeneous data through the community IoT sensing layer, management service layer, and social network layer;

[0011] By using natural language processing and address matching algorithms, a dynamic mapping relationship between virtual social accounts and physical real estate units is established. Multi-source heterogeneous data is cleaned and anchored to the community geographic grid to form a feature data space containing spatiotemporal attributes.

[0012] A community risk identification model is constructed to analyze the feature data space, identify risk anomalies in the feature data space, and transform the risk anomalies into a continuously distributed risk potential energy field; the potential energy value distribution of the risk potential energy field is calculated to quantify the degree of risk aggregation in physical space; the spatial gradient vector of the potential energy field is calculated to characterize the diffusion direction and rate of risk in spatial dimension; and the type of risk event is identified by combining the semantic labels of the feature data space with the physical space neighborhood relationship.

[0013] Based on the type of risk event, the distribution of potential energy value, and the spatial gradient vector, the trajectory of risk evolution is predicted. When the predicted risk potential energy exceeds the safety threshold, an early warning signal containing the coordinates, type, and diffusion range of the risk source is generated.

[0014] The process of acquiring the multi-source heterogeneous data includes:

[0015] The community IoT sensing layer connects to IoT sensors to collect physical sensing data including environmental noise decibel values, overflowing trash cans, images of objects thrown from high-rise buildings, and elevator vibration frequency.

[0016] The management service layer connects to the property management system and public service hotline interface, and collects service management data including inspection logs of public facilities, processing time of repair work orders, and written records of resident complaints.

[0017] The social network layer accesses homeowner communication groups and community forums, collecting unstructured social data including publicly posted communication texts, image information, and likes, comments, and interaction data from residents;

[0018] The collected heterogeneous data from multiple sources are tagged with timestamps to construct a multi-source heterogeneous data system covering physical environment status, service response efficiency, and community sentiment.

[0019] The process of establishing the dynamic mapping relationship includes:

[0020] Establish a standard physical address database for the community and assign a unique spatial code to each independent physical property unit.

[0021] For unstructured social data collected from the social network layer, the building number, unit number and floor location features implicit in the text are extracted; the semantic dependency relationship between the extracted address features and the user's first-person pronoun is parsed using a dependency parsing algorithm to determine whether the user has the residential attribute of the address.

[0022] Simultaneously, regular expressions are used to perform pattern matching on the numerical sequences in users' social nicknames, and cross-validation is performed with the owner information that has been confirmed in the property management system. Based on the validation results, the virtual social accounts are bound to the corresponding physical property unit spatial codes to generate a dynamic mapping relationship table.

[0023] The process of constructing the feature data space includes:

[0024] Based on the identification and processing of multi-source heterogeneous data, physical stress index, emotional stress index and service stress index are obtained;

[0025] Using physical property units as the smallest geographic grid unit, and utilizing the established dynamic mapping relationship, the physical stress index, emotional stress index, and service stress index are accurately projected onto the corresponding geographic grid coordinates.

[0026] Using a preset time window as the granularity, multi-source indices within the same geographic grid are fused to generate a comprehensive state vector for the grid. The comprehensive state vectors of all geographic grids in continuous time series are combined to construct a multi-dimensional feature data space that dynamically reflects the overall operation of the community.

[0027] The process of obtaining the physical stress index, emotional stress index, and service stress index includes:

[0028] Feature extraction and normalization are performed on physical sensing data, continuous environmental monitoring values ​​are mapped into interference indexes, discrete abnormal events are converted into occurrence frequencies, and a weighted fusion algorithm is used to generate a physical stress index that characterizes the state of environmental facilities.

[0029] The negative sentiment confidence level of unstructured social data is calculated using a sentiment semantic analysis model, and the social dissemination weight is calculated by combining the interaction popularity and dissemination breadth of the data. The negative sentiment confidence level is then weighted and corrected using the social dissemination weight to generate an emotional stress index that represents the intensity of group emotions.

[0030] Extract work order response time and maintenance handling records from service management data, calculate the deviation of actual processing time from standard time, combine with missing maintenance records to conduct performance evaluation, and generate a service pressure index that characterizes the degree of lag in property services.

[0031] The process of obtaining the risk potential energy field includes:

[0032] Traverse each geographic grid in the feature data space, use the Euclidean norm algorithm to obtain the magnitude of the comprehensive state vector, and define it as the comprehensive state value; monitor the fluctuation of the comprehensive state value, and mark it as a discrete risk anomaly when the state value exceeds the threshold set based on the historical baseline.

[0033] By introducing a physical field theory model and using a kernel density estimation algorithm, the radiation influence range and attenuation coefficient of each risk anomaly point are calculated by combining the Gaussian kernel function.

[0034] The radiation impact values ​​of all risk anomalies are spatially superimposed, and a time decay factor is introduced in the time dimension to accumulate the residual impact of historical risks. This smoothly maps the discretely distributed risk anomalies into a continuous risk potential energy field covering the entire geographical area of ​​the community. The value of each point in the risk potential energy field represents the risk potential energy value of the current location.

[0035] For the potential energy distribution of the risk potential energy field, the peak center of the potential energy distribution is determined by calculating the local maxima of the risk potential energy field, thereby quantifying the high-density clustering area of ​​risk in physical space.

[0036] Given the spatial gradient vector of the potential energy field, the vector field of the spatial gradient of the risk potential energy field is obtained, and the spatial gradient vector of the potential energy field is obtained. The magnitude of the spatial gradient vector represents the rate of risk diffusion, and the vector direction represents the direction of risk diffusion in physical space.

[0037] Based on the type of risk event, extract the original text semantic tags corresponding to the risk anomaly points, and distinguish between physically transmitted events and emotionally contagious events according to the tag type and the spatial gradient vector direction.

[0038] The current risk event type, potential energy distribution, and spatial gradient vector are input into an evolutionary prediction model based on a long short-term memory network to simulate the changes in the risk potential energy field within future time steps and generate a risk evolution trajectory.

[0039] The system monitors the peak potential energy in the evolution trajectory in real time. Once the risk potential energy value at a future moment exceeds the preset safety threshold, an early warning mechanism is immediately triggered. The system traces back to the starting grid coordinates that caused the potential energy to exceed the limit as the risk source coordinates. The risk type is determined based on the identified event semantics, and the risk diffusion range is determined based on the predicted potential energy field coverage contour lines, thus generating an early warning signal.

[0040] A risk assessment and early warning system based on big data analysis includes:

[0041] The data acquisition module collects multi-source heterogeneous data through the community IoT sensing layer, management service layer, and social network layer;

[0042] The data processing module uses natural language processing and address matching algorithms to establish a dynamic mapping relationship between virtual social accounts and physical real estate units. After cleaning multi-source heterogeneous data, it anchors it to the community geographic grid to form a feature data space containing spatiotemporal attributes.

[0043] The risk identification module constructs a community risk identification model to analyze the feature data space, identify risk anomalies in the feature data space, and transforms the risk anomalies into a continuously distributed risk potential energy field; calculates the potential energy value distribution of the risk potential energy field to quantify the degree of risk aggregation in physical space; calculates the spatial gradient vector of the potential energy field to characterize the diffusion direction and rate of risk in spatial dimensions; and identifies the type of risk event by combining the semantic labels of the feature data space with the physical space neighborhood relationship.

[0044] The risk warning module predicts the trajectory of risk evolution based on the type of risk event, the distribution of potential energy value, and the spatial gradient vector. When the predicted risk potential energy exceeds the safety threshold, it generates a warning signal containing the coordinates, type, and diffusion range of the risk source.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] 1. This invention effectively solves the problems of data silos and the difficulty in implementing virtual risks in traditional community governance by constructing a three-dimensional perception network covering the physical environment, service management, and social networks. Utilizing natural language processing and address matching algorithms, a dynamic mapping relationship between virtual social accounts and physical property units is established, accurately anchoring unstructured discussion texts lacking geographical identifiers hidden in online groups to specific physical apartment numbers. This provides a complete and solid data foundation for subsequent exploration of cross-dimensional risk transmission mechanisms, improving the precision of community governance.

[0047] 2. This invention introduces a physical field theory model to transform discretely distributed community anomaly data into a continuously distributed risk potential energy field, enabling in-depth mathematical quantification and visualization analysis of community risk status. By calculating the potential energy distribution of the risk potential energy field, the degree of risk aggregation in physical space can be accurately quantified; spatial gradient vectors are used to characterize the diffusion characteristics of risk, with the magnitude of the gradient vector reflecting the potential rate of risk diffusion and the direction revealing the path of risk flow; combined with semantic tags and physical spatial neighborhood relationships, it is possible to distinguish whether the risk is due to physical facility failure transmission or social emotional contagion; it can accurately identify where the risk will flow and how it will spread, providing a basis for formulating measures to prevent the spread of risk.

[0048] 3. This invention, based on an evolutionary prediction model using Long Short-Term Memory (LSTM) networks, enables forward-looking prediction of community risk evolution trajectories. By using a risk potential energy field sequence generated from historical periods as input, it can simulate changes in risk potential energy within future time steps and generate dynamic evolutionary trajectories. This allows community managers to perceive potential crisis signals in advance and obtain detailed early warning work orders containing the coordinates, type, and expected spread of risk sources. This not only significantly reduces the cost and social impact of risk governance but also promotes a fundamental shift in community governance models from traditional post-event remediation to pre-event prevention and proactive intervention, effectively enhancing the community's risk early warning capabilities. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating a risk assessment and early warning method based on big data analysis according to the present invention.

[0050] Figure 2 This is a logical diagram of a risk assessment and early warning method based on big data analysis according to the present invention;

[0051] Figure 3 This is a schematic diagram of the structure of a risk assessment and early warning system based on big data analysis according to the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1:

[0054] This invention proposes a risk assessment and early warning method based on big data analysis. The process of the method is as follows: Figure 1 As shown, the logic of the method is as follows: Figure 2 As shown; including:

[0055] Collect multi-source heterogeneous data through the community IoT sensing layer, management service layer, and social network layer;

[0056] By using natural language processing and address matching algorithms, a dynamic mapping relationship between virtual social accounts and physical real estate units is established. Multi-source heterogeneous data is cleaned and anchored to the community geographic grid to form a feature data space containing spatiotemporal attributes.

[0057] A community risk identification model is constructed to analyze the feature data space, identify risk anomalies in the feature data space, and transform the risk anomalies into a continuously distributed risk potential energy field; the potential energy value distribution of the risk potential energy field is calculated to quantify the degree of risk aggregation in physical space; the spatial gradient vector of the potential energy field is calculated to characterize the diffusion direction and rate of risk in spatial dimension; and the type of risk event is identified by combining the semantic labels of the feature data space with the physical space neighborhood relationship.

[0058] Based on the type of risk event, the distribution of potential energy value, and the spatial gradient vector, the trajectory of risk evolution is predicted. When the predicted risk potential energy exceeds the safety threshold, an early warning signal containing the coordinates, type, and diffusion range of the risk source is generated.

[0059] The process of acquiring the multi-source heterogeneous data includes:

[0060] The community IoT sensing layer connects to IoT sensors to collect physical sensing data including environmental noise decibel values, overflowing trash cans, images of objects thrown from high-rise buildings, and elevator vibration frequency.

[0061] The management service layer connects to the property management system and public service hotline interface, and collects service management data including inspection logs of public facilities, processing time of repair work orders, and written records of resident complaints.

[0062] The social network layer accesses homeowner communication groups and community forums, collecting unstructured social data including publicly posted communication texts, image information, and likes, comments, and interaction data from residents;

[0063] The collected heterogeneous data from multiple sources are tagged with timestamps to construct a multi-source heterogeneous data system covering physical environment status, service response efficiency, and community sentiment.

[0064] Furthermore, the process of collecting and cleaning physical sensing data includes:

[0065] Sensor devices within the community are accessed via an IoT gateway; for environmental noise, a decibel meter is used to collect sound intensity values ​​at preset time thresholds, such as every five minutes; for high-altitude littering and overflowing garbage, the edge computing module of a smart camera is used to generate a structured log containing event type, occurrence time, and device number when a littering action is detected or the pixel ratio of a garbage can exceeds the warning line; for facility operation, vibration frequency data in the vertical direction from the elevator car acceleration sensor is collected;

[0066] For abnormal extreme values generated by sensors due to network jitter, such as instantaneous zeroing or values exceeding physical limits, a median filtering algorithm is used for smoothing and denoising. For missing data points in the time series, a linear interpolation method is adopted to fill them with the average value of the two valid data before and after the missing moment, ensuring data continuity.

[0067] The collection and structuring of service management data include:

[0068] Collection mechanism: Read the database of the property management system through the application programming interface, extract the creation timestamp, completion timestamp, processing status of the repair work order, and the text records of complaints submitted by residents through phone or software.

[0069] Data structuring: Perform word segmentation on the unstructured complaint text, remove stop words such as "de", "le" and other meaningless auxiliary words, and retain nouns and verbs with actual meanings.

[0070] The collection and desensitization process of social network data include:

[0071] Collection mechanism: On the premise of obtaining user authorization, access the owner's WeChat group and community forum, capture the text content and pictures posted by residents, and synchronously count the number of likes and comments on this content;

[0072] Privacy desensitization: Before data is stored in the database, use the hash algorithm to encrypt the user's original identity information, and only retain the virtual account identifier for correlation analysis;

[0073] Due to the different data sources and various time granularities above, all data is uniformly converted into the standard format of year, month, day, hour, minute and second, and based on the community geographic information system, geographic coordinate labels are added to the data of all fixed points, such as the sensor location.

[0074] The present invention constructs a three-dimensional stereo perception network covering the physical environment, service management and social network, breaking the data island phenomenon in traditional community governance and achieving full coverage of perception; it can not only perceive visible equipment failures, but also perceive residents' emotions and service efficiency, significantly improving the breadth of risk discovery and enhancing the associated value of data; through a unified timestamp label, the chain reaction process of a certain risk event at the physical, service and social levels can be traced back, providing a complete data chain for subsequent causal analysis; enhancing the robustness of the system, the cross-validation mechanism of multi-source data effectively reduces the false alarm rate that may be brought by a single data source, laying a solid data foundation for accurate early warning.

[0075] The establishment process of the dynamic mapping relationship includes:

[0076] Establish a standard physical address library for the community, and assign a unique spatial code to each independent physical property unit;

[0077] For unstructured social data collected from the social network layer, the building number, unit number and floor location features implicit in the text are extracted; the semantic dependency relationship between the extracted address features and the user's first-person pronoun is parsed using a dependency parsing algorithm to determine whether the user has the residential attribute of the address.

[0078] Simultaneously, regular expressions are used to perform pattern matching on the numerical sequences in users' social nicknames, and cross-validation is performed with the owner information that has been confirmed in the property management system. Based on the validation results, the virtual social accounts are bound to the corresponding physical property unit spatial codes to generate a dynamic mapping relationship table.

[0079] Based on the community building drawings, a tree-structured physical address database was established, and a unique Chinese spatial code was assigned to each independent physical property unit, such as Room 301, Unit 1, Building 5, Phase 1.

[0080] Address entity recognition based on deep learning includes:

[0081] A bidirectional long short-term memory network combined with a conditional random field model is used. This model has a three-layer structure:

[0082] Input layer: Transforms unstructured social data into sequences of word vectors;

[0083] Bidirectional Long Short-Term Memory Network Layer: Extracts contextual features of the text from front to back and from back to front respectively, capturing the distance dependency between "Five Buildings" and "Three Zero One";

[0084] Conditional Random Field Layer: Constrains the output label sequence to ensure that the identified address conforms to the logical rule of "building number + unit number + room number";

[0085] Recognition process: The model automatically scans the chat text and identifies the hidden "building number", "unit number" and "floor location words", such as "upstairs" and "my door".

[0086] The semantic dependency judgment principle of residential attributes includes:

[0087] Using dependency parsing algorithms, construct a dependency relationship tree between words in a sentence;

[0088] Logical judgment: The focus is on analyzing whether there is a "possession relationship" or "attribute-head relationship" between the first-person pronouns in the sentence, such as "I" and "we," and the extracted address entities. For example, in the sentence "My home is in Building 5," there is a possession relationship between "I" and "Building 5," indicating that the user has the residential attribute of that address; however, in the sentence "Building 5 is on fire," there is no possession relationship, indicating that it is an eyewitness description rather than a residential attribute.

[0089] Multimodal cross-validation and dynamic binding include:

[0090] Regular expression matching: Use the preset "number plus hyphen plus number" rule to scan user nicknames and extract suspected room numbers;

[0091] Property rights verification: The extracted room number is compared with the registered owner's mobile phone number in the property system. When the verification is successful, the virtual social account is bound to the physical property space code, stored in the dynamic mapping relationship table, and updated regularly.

[0092] This invention introduces NLP and address matching algorithms to achieve the physical realization of virtual risks, accurately anchoring the fluctuating online complaints to specific geographical coordinates; it significantly improves the accuracy of identity recognition by effectively eliminating false alarms through multimodal cross-validation of semantic analysis, regular expression matching, and system rights confirmation, ensuring the accuracy of subsequent risk attribution; it supports dynamic updates to adapt to the characteristics of rapid population mobility and frequent tenant changes in communities, maintaining the real-time and effective "person-room-number" association, and providing technical support for accurate door-to-door investigation.

[0093] The process of constructing the feature data space includes:

[0094] Based on the identification and processing of multi-source heterogeneous data, physical stress index, emotional stress index and service stress index are obtained;

[0095] Using physical property units as the smallest geographic grid unit, and utilizing the established dynamic mapping relationship, the physical stress index, emotional stress index, and service stress index are accurately projected onto the corresponding geographic grid coordinates.

[0096] Using a preset time window as the granularity, multi-source indices within the same geographic grid are fused to generate a comprehensive state vector for the grid. The comprehensive state vectors of all geographic grids in continuous time series are combined to construct a multi-dimensional feature data space that dynamically reflects the overall operation of the community.

[0097] The process of obtaining the physical stress index, emotional stress index, and service stress index includes:

[0098] Feature extraction and normalization are performed on physical sensing data, continuous environmental monitoring values ​​are mapped into interference indexes, discrete abnormal events are converted into occurrence frequencies, and a weighted fusion algorithm is used to generate a physical stress index that characterizes the state of environmental facilities.

[0099] The negative sentiment confidence level of unstructured social data is calculated using a sentiment semantic analysis model, and the social dissemination weight is calculated by combining the interaction popularity and dissemination breadth of the data. The negative sentiment confidence level is then weighted and corrected using the social dissemination weight to generate an emotional stress index that represents the intensity of group emotions.

[0100] Extract work order response time and maintenance handling records from service management data, calculate the deviation of actual processing time from standard time, combine with missing maintenance records to conduct performance evaluation, and generate a service pressure index that characterizes the degree of lag in property services.

[0101] The calculation of the physical pressure index includes:

[0102] Noise interference mapping: According to the Weber-Fechner law, the human ear's perception of sound changes is logarithmic; the collected decibel values ​​are mapped to an interference index between zero and one using a logarithmic function; specifically, 45 dB is set as the zero interference threshold, 85 dB as the full interference threshold, and the intermediate values ​​are logarithmically interpolated.

[0103] Anomaly frequency: Statistically count the number of discrete alarm events within a unit time window, such as one hour, and convert them into a value between zero and one using the max-min normalization method;

[0104] Weighted fusion: The noise interference level, abnormal frequency index, and normalized elevator vibration deviation value are added together according to preset weights to generate a physical pressure index. Among them, elevator vibration deviation normalization is to normalize the range of the collected elevator operating vibration frequency with the historical normal operating frequency.

[0105] Furthermore, the objective weights of each indicator are calculated using the entropy weight method to reflect the differences in the amount of information carried by the indicator data. The specific steps include: First, selecting historical data from past time periods to construct a time series matrix containing noise interference, abnormal frequency indicators, and vibration deviation values; second, calculating the proportion of each indicator value at each moment to the total sum of that indicator over the entire time window; third, using this proportion to calculate the information entropy value of each indicator; the smaller the information entropy, the greater the dispersion of the indicator data and the more risk information it carries; finally, calculating the difference coefficient of each indicator (i.e., 1 minus the information entropy) and normalizing the difference coefficient to obtain the objective weight of each indicator. This dynamic weight is then used to weight and sum the three component indicators to generate the final physical stress index. The weight configuration is updated every 24 hours to adapt to changes in risk characteristics in different community environments.

[0106] The calculation of the emotional stress index includes:

[0107] Negative sentiment confidence calculation: Sentiment classification is performed using a transformer-based bidirectional encoding representation model. The model is pre-trained on a massive social corpus and fine-tuned on community-specific corpus, and can output the probability value of the text belonging to different sentiment types.

[0108] Social media dissemination weight adjustment: Interaction popularity is introduced as an amplification factor; the calculation method is that the social media dissemination weight is equal to 1 plus (the natural logarithm of the sum of likes and comments); this weight is multiplied by the probability value of negative emotions, and an S-shaped function, such as the Sigmoid function, is used to limit the result to between 0 and 1; the principle on which this design is based is: the more interaction there is with negative information, the stronger the group resonance it triggers, and the higher the emotional pressure index should be.

[0109] The calculation of the service stress index includes:

[0110] Performance deviation assessment: Calculate the difference between the actual processing time of a work order and the standard promised processing time;

[0111] Exponential penalty calculation: If the difference is positive, it means timeout; the penalty value is calculated using an exponential function, which has the characteristics of being flat in the early stage and steep in the later stage, simulating the psychological process of residents' patience with overtime service being consumed at an accelerated rate over time.

[0112] Feature data space construction: Using the established mapping relationship, the three stress indices mentioned above are accurately filled into the corresponding physical property unit grid; with a preset time window, such as 1 hour as the granularity, a multi-dimensional feature data space containing geographic coordinates and stress index dimensions is constructed.

[0113] This invention establishes a unified mathematical expression model for complex community data by constructing a feature data space containing spatiotemporal attributes; it achieves standardized fusion of heterogeneous data, transforming data from sensors or WeChat groups into vectors of a unified dimension, enabling data from different sources to be calculated and compared within the same model; it preserves fine-grained risk features, using households as the smallest unit and retaining the vector structure, avoiding the loss of micro-details in macro-statistics and facilitating the discovery of hidden risks; and it constructs a risk foundation from a full spatiotemporal perspective, where the multi-dimensional feature space not only records the current risk state but also preserves the historical evolution sequence, providing a standardized input data format for subsequent spatiotemporal sequence prediction using deep learning models.

[0114] The process of obtaining the risk potential energy field includes:

[0115] Traverse each geographic grid in the feature data space, and use the Euclidean norm algorithm to obtain the magnitude of the comprehensive state vector, which is defined as the comprehensive state value; monitor the fluctuation of the comprehensive state value, and when the state value exceeds the threshold set based on the historical baseline, it is marked as a discrete risk anomaly point; wherein, the geographic grid is based on each household's real estate unit as an individual, that is, each geographic network represents one household, the geographic grid is a three-dimensional spatial attribute, the horizontal and vertical axes represent geographic latitude and longitude projection, and the z-axis represents the floor height or relative altitude.

[0116] By introducing a physical field theory model and using a kernel density estimation algorithm, the radiation influence range and attenuation coefficient of each risk anomaly point are calculated by combining the Gaussian kernel function.

[0117] The radiation impact values ​​of all risk anomalies are spatially superimposed, and a time decay factor is introduced in the time dimension to accumulate the residual impact of historical risks. This smoothly maps the discretely distributed risk anomalies into a continuous risk potential energy field covering the entire geographical area of ​​the community. The value of each point in the risk potential energy field represents the risk potential energy value of the current location.

[0118] The core logic of the time decay factor is that the influence of a risk event decreases rapidly at first and then slowly over time, until it approaches zero, but it does not disappear immediately.

[0119] Construct a decay function based on time intervals. Assuming the time difference between the current moment and the previous moment is the time step, the decay factor can be defined as follows:

[0120] Basic definition: The attenuation factor is a value between 0 and 1.

[0121] The formula describes the attenuation factor as equal to the negative power of the natural constant, where the power is the attenuation rate coefficient multiplied by the elapsed time interval.

[0122] Attenuation rate coefficient: This is a configurable parameter. A larger coefficient means the risk is forgotten more quickly; for example, noise interference disappears quickly once it stops. A smaller coefficient means the risk lingers longer; for example, water leakage causing damp walls has a longer duration of impact. It is obtained by normalizing the average duration of different risk types from historical data.

[0123] For the potential energy distribution of the risk potential energy field, the peak center of the potential energy distribution is determined by calculating the local maxima of the risk potential energy field, thereby quantifying the high-density clustering area of ​​risk in physical space.

[0124] Given the spatial gradient vector of the potential energy field, the vector field of the spatial gradient of the risk potential energy field is obtained, and the spatial gradient vector of the potential energy field is obtained. The magnitude of the spatial gradient vector represents the rate of risk diffusion, and the vector direction represents the direction of risk diffusion in physical space.

[0125] Based on the type of risk event, extract the original text semantic tags corresponding to the risk anomaly points, and distinguish between physically transmitted events and emotionally contagious events according to the tag type and the spatial gradient vector direction.

[0126] Using field theory to transform discrete data into a continuous risk field:

[0127] Comprehensive state numerical dimensionality reduction: For each grid, its three stress indices of "physical, emotional, and service" are regarded as a three-dimensional vector. The geometric length of the vector is calculated using the Euclidean norm algorithm. The formula is the arithmetic square root of the sum of the squares of the three components. The magnitude of the vector is calculated and defined as the comprehensive state value of the grid.

[0128] Outlier identification: The statistical process control algorithm is used to monitor the value, maintain a sliding time window, and calculate the historical average and standard deviation of the comprehensive state value within the window. If the current value exceeds the historical average plus three times the standard deviation, according to statistical principles, the probability of this event is extremely low, and it is marked as a discrete risk outlier.

[0129] Field superposition effect: The radiation impact values ​​generated by all anomalies are superimposed on the feature space; if multiple anomalies are spatially adjacent or temporally close, their radiation ranges will overlap, resulting in a significant increase in the values ​​of the superimposed region. Ultimately, a risk potential energy field covering the entire community is formed.

[0130] Furthermore, vector analysis is used to reveal the fluid characteristics of the risk:

[0131] Aggregation quantification: Scan the risk potential energy field and use the local extremum search algorithm to find the peak point; this point is the potential energy center, representing the area where the risk is most densely concentrated in the physical space, such as the core building.

[0132] Diffusion trend characterization: Taking the partial derivative of the risk potential energy field in the spatial dimension yields the spatial gradient vector. The negative gradient vector of the potential energy field points in the direction of the fastest decrease in potential energy; therefore, the direction of the gradient vector based on the potential energy field can characterize the direction of risk diffusion. In a community scenario, this represents the flow of risk from high-risk areas to low-risk areas; for example, from upstairs to downstairs, or from one building to another. The magnitude of the gradient vector represents the steepness of the potential energy change; the steeper the slope, the stronger the potential driving force and the faster the rate of risk diffusion.

[0133] Risk mechanisms include physical transmission and emotional contagion identification:

[0134] Based on the type of risk event, extract the original text semantic tags corresponding to the risk anomaly points; including physical attribute tags and social attribute tags; among them, physical attribute tags include "water leakage", "noise" or "vibration", etc.; social attribute tags include "service problem" and "unreasonable charges", etc.

[0135] Physical conduction type judgment (orderly diffusion): If the gradient vector around the risk anomaly point shows a continuous radial or linear flow direction when monitoring physical attribute tags, that is, the angle between the gradient vectors of adjacent grids is less than the preset threshold, and the potential energy value shows a smooth decay characteristic with the gradient direction, then it is judged as a physical conduction type event; for example, the gradient vector of a water leakage event should point to the direction of gravity, that is, the negative Z axis, and be distributed vertically along the pipe well.

[0136] Emotional contagion type determination (discrete jump): If social attribute tags are detected, the gradient vector around the risk anomaly point presents a discrete island shape, that is, the potential energy of the point is extremely high, but the potential energy of its physical neighborhood is extremely low, resulting in the gradient vector magnitude being extremely large and quickly returning to zero; or multiple non-adjacent high potential energy centers with randomly diverging gradient directions appear simultaneously within the community, showing multi-point resonance characteristics, then it is determined to be an emotional contagion type event.

[0137] This invention achieves precise characterization of risk evolution characteristics through in-depth mathematical analysis of the risk potential energy field; accurately locates the core of the risk by calculating the local maxima of the potential energy distribution, enabling rapid identification of the most pressing risk points from complex community data and guiding priority resource allocation; predicts risk diffusion paths by using spatial gradient vectors to scientifically calculate the direction and velocity of risk, such as predicting which buildings will see the spread of discontent; and intelligently identifies risk mechanisms by combining semantic tags with physical spatial neighborhood relationships, automatically distinguishing whether the risk is due to physical facility malfunctions or social emotional contagion, providing a basis for subsequent generation of targeted governance strategies.

[0138] The community risk identification model adopts a statistical analysis framework based on physical field theory, and its specific identification process includes:

[0139] The first step is to calculate the comprehensive state value: Traverse each geographic grid in the feature data space and extract the physical stress index, emotional stress index, and service stress index within that grid. First, confirm that each index has been normalized (values ​​are all between 0 and 1). Then, using the weighted Euclidean norm algorithm, calculate the magnitude of the state vector composed of the three indices according to preset business weights. This magnitude is defined as the comprehensive state value of that grid. Monitor the fluctuation of this value; when it exceeds three times the standard deviation threshold calculated based on historical moving averages, mark the grid as a discrete risk anomaly.

[0140] The second step is the construction of a continuous risk potential field: Physical field theory methods are introduced, and a kernel density estimation algorithm is used to map discrete data into a continuous representation. Centered on each risk anomaly point, a Gaussian kernel function is used to calculate its radiative attenuation value to the surrounding geographic grid. The radiative attenuation values ​​of all anomaly points are spatially accumulated, and an exponential decay factor is introduced in the time dimension to retain the residual impact of historical risks, thereby generating a continuous risk potential field covering the entire community. The value at any point in the field represents the current degree of risk accumulation at that physical location. The kernel density estimation algorithm can employ a three-dimensional anisotropic kernel density estimation algorithm.

[0141] The third step is vectorized feature analysis: For the risk potential energy field, a local extremum search algorithm is used to determine the peak center of the potential energy distribution, quantifying the high-density clustering area of ​​risk. Simultaneously, the risk potential energy field is differentiated to obtain the spatial gradient vector field. The magnitude of the gradient vector represents the drastic change in risk potential energy (i.e., the potential diffusion rate), and the reverse direction of the gradient vector represents the diffusion path of risk from the high-potential energy region to the low-potential energy region.

[0142] The fourth step is risk type and mechanism identification: using natural language processing technology to extract semantic tags from the original text associated with risk anomalies, such as "facility failure" or "neighborhood atmosphere". If the semantic tag is physical and the surrounding gradient vector shows a continuous and smooth radial distribution, it is determined to be a physical transmission event; if the semantic tag is social and the surrounding gradient vector shows a multi-point discrete or jump distribution, it is determined to be an emotion contagion event.

[0143] The current risk event type, potential energy distribution, and spatial gradient vector are input into an evolutionary prediction model based on a long short-term memory network to simulate the changes in the risk potential energy field within future time steps and generate a risk evolution trajectory.

[0144] The system monitors the peak potential energy in the evolution trajectory in real time. Once the risk potential energy value at a future moment exceeds the preset safety threshold, an early warning mechanism is immediately triggered. The system traces back to the starting grid coordinates that caused the potential energy to exceed the limit as the risk source coordinates. The risk type is determined based on the identified event semantics, and the risk diffusion range is determined based on the predicted potential energy field coverage contour lines, thus generating an early warning signal.

[0145] Evolutionary trajectory prediction employs a Long Short-Term Memory (LSTM) network. This model, by introducing forgetting, input, and output gates, effectively captures long-term dependencies in time-series data. Specifically, it uses a two-layer stacked LTM network structure. The first layer is an LTM layer with 128 neurons, designed to return the complete sequence, followed by a random deactivation layer with a dropout rate of 0.3 to prevent overfitting. The second layer is an LTM layer with 64 neurons, returning only the final state, followed by a random deactivation layer with a dropout rate of 0.2. The model's backend connects to two fully connected layers. The first fully connected layer contains 32 neurons and uses a linear rectified function as the activation function. The second fully connected layer contains 3 neurons and uses a linear activation function, outputting the x-coordinate, y-coordinate, and predicted potential energy value of the risk center at future time steps.

[0146] During training, mean squared error was used as the loss function, and an adaptive moment estimator (AME) was used for parameter updates. The initial learning rate was set to 0.1%, and the batch size was set to 32. The total number of training epochs was set to 200, and an early stopping mechanism was implemented: training automatically stopped when the loss on the validation set did not decrease within 15 consecutive epochs. The evaluation metrics required a mean absolute error of less than 0.1 and a prediction accuracy of greater than 80% for risky events.

[0147] Input / Output: The risk potential energy field sequence generated in the past 24 hours is used as the time step sequence input to the model. The model outputs a potential energy field distribution prediction map for the next time step, such as one hour. Specifically, for each time step, a 12-dimensional feature vector of the entire community is extracted as input: including the global average of four dimensions, namely the average of physical pressure index, emotional pressure index, service pressure index and comprehensive potential energy value; the global maximum of four dimensions, namely the maximum value corresponding to the above four indices; and the peak feature containing four dimensions, namely the x-coordinate, y-coordinate, height coordinate and potential energy value of the highest potential energy point in the risk potential energy field at the current moment.

[0148] Threshold trigger: Real-time monitoring of the potential energy peak value in the prediction map. Once it is found that the potential energy value of a certain area is expected to exceed the safety threshold at a future time, an early warning mechanism is immediately triggered.

[0149] Reverse tracing: Tracing back along the gradient vector to pinpoint the starting grid coordinates that caused the potential energy to exceed the limit, as the source of the risk.

[0150] Signal output: Based on the identified risk type, a structured early warning work order containing the risk source coordinates, risk type and expected spread range is generated and sent to the community management terminal.

[0151] Example 2:

[0152] This invention proposes a risk assessment and early warning system based on big data analysis, used to implement a risk assessment and early warning method based on big data analysis; the structure of the system is as follows: Figure 3 As shown, it includes a data acquisition module, a data processing module, a risk identification module, and a risk warning module.

[0153] The data acquisition module collects multi-source heterogeneous data through the community IoT sensing layer, management service layer, and social network layer.

[0154] The data processing module uses natural language processing and address matching algorithms to establish a dynamic mapping relationship between virtual social accounts and physical property units. After cleaning multi-source heterogeneous data, it anchors it to the community geographic grid to form a feature data space containing spatiotemporal attributes.

[0155] The risk identification module constructs a community risk identification model to analyze the feature data space, identify risk anomalies in the feature data space, and transforms the risk anomalies into a continuously distributed risk potential energy field; calculates the potential energy value distribution of the risk potential energy field to quantify the degree of risk aggregation in physical space; calculates the spatial gradient vector of the potential energy field to characterize the diffusion direction and rate of risk in spatial dimensions; and identifies the type of risk event by combining the semantic labels of the feature data space with the neighborhood relationship in physical space.

[0156] The process of identifying the feature data space by the community risk identification model includes:

[0157] Traverse each geographic grid in the feature data space, use the Euclidean norm algorithm to obtain the magnitude of the comprehensive state vector, and define it as the comprehensive state value; monitor the fluctuation of the comprehensive state value, and mark it as a discrete risk anomaly when the state value exceeds the threshold set based on the historical baseline.

[0158] By introducing a physical field theory model and using a kernel density estimation algorithm, the radiation influence range and attenuation coefficient of each risk anomaly point are calculated by combining the Gaussian kernel function.

[0159] The radiation impact values ​​of all risk anomalies are spatially superimposed, and a time decay factor is introduced in the time dimension to accumulate the residual impact of historical risks. This smoothly maps the discretely distributed risk anomalies into a continuous risk potential energy field covering the entire geographical area of ​​the community. The value of each point in the risk potential energy field represents the risk potential energy value of the current real estate unit.

[0160] For the potential energy distribution of the risk potential energy field, the peak center of the potential energy distribution is determined by calculating the local maxima of the risk potential energy field, thereby quantifying the high-density clustering area of ​​risk in physical space.

[0161] Given the spatial gradient vector of the potential energy field, the vector field of the spatial gradient of the risk potential energy field is obtained, and the spatial gradient vector of the potential energy field is obtained. The magnitude of the spatial gradient vector represents the rate of risk diffusion, and the vector direction represents the direction of risk diffusion in physical space.

[0162] Based on the type of risk event, extract the original text semantic tags corresponding to the risk anomaly points, and distinguish between physically transmitted events and emotionally contagious events according to the tag type and the spatial gradient vector direction.

[0163] Furthermore, a spatiotemporal graph convolutional network architecture based on an attention mechanism can also be used to construct a community risk identification model:

[0164] The community geographic grid is treated as graph nodes, and the physical adjacency and social relationships between grids are treated as graph edges to construct a spatiotemporal heterogeneous graph of the community. The feature data of each node and its neighboring nodes are aggregated by graph convolutional layers, and the contribution weight of different neighboring nodes to the risk state of the central node is dynamically calculated through a spatial attention mechanism. The evolution of node features in the time dimension is captured by temporal convolutional layers, and the feature influence of key time steps is weighted through a temporal attention mechanism. Finally, a risk potential field representation that integrates high-order spatiotemporal features is output.

[0165] The traditional linear weighted method is replaced by ST-GCN (Spatiotemporal Graph Convolutional Network), which is at the forefront of deep learning.

[0166] First, the graph construction phase: Unlike a simple grid, a heterogeneous graph is constructed; nodes represent housing units, and edges include not only physical edges (up, down, left, and right neighbors), but also functional edges (sharing the same elevator) and social edges (being in the same social group).

[0167] Secondly, in the spatial convolution stage: data enters the graph convolutional layer; for node A, not only is A's data read, but the data of all its neighbors (physical, functional, and social neighbors) are also automatically captured and aggregated; a multi-head attention mechanism is introduced to automatically learn weights; for example, when analyzing the risk of infectious diseases, neighbors sharing elevators are automatically assigned high weights; when analyzing the risk of noise, upstairs neighbors are assigned high weights.

[0168] Finally, in the temporal convolution stage: a TCN (Temporal Convolutional Network) is used to process the time series; temporal attention is introduced to focus the model on key time nodes and ignore noisy data during the quiet period at night. The final output feature vector is not just a sum of several exponents, but a high-dimensional tensor containing complex spatial topology and temporal rhythms. The output high-dimensional feature tensor is input into a fully connected layer, and the high-dimensional features are mapped to a one-dimensional risk potential scalar value through linear transformation and activation function; this scalar value is used as the risk potential intensity of the corresponding geographic grid, thereby constructing a continuous risk potential field; subsequently, based on this risk potential field, the steps of potential value distribution, spatial gradient vector, and identification of risk evolution trajectory are performed.

[0169] This approach captures the complex relationships in non-Euclidean spaces, where community risks often do not follow simple linear distance decay. For example, an elevator malfunction affects a vertical string of households rather than horizontal neighbors. The ST-GCN model can perfectly handle such irregular topologies, better reflecting the physical reality of communities than traditional grid methods, and achieving adaptive feature fusion. Through an attention mechanism, the model can automatically adjust its focus for different risk scenarios without manually pre-setting fixed weight formulas, greatly improving the model's generalization ability in variable scenarios. It also improves the detection rate of weak signals. By aggregating neighbor information, the model can use the features of surrounding nodes to fill in or enhance the information of the central node, amplifying weak anomalous signals of a single household with the support of the neighborhood, effectively avoiding the underreporting of early hidden risks.

[0170] Furthermore, identifying the type of risk event may also include constructing a social network propagation graph for auxiliary determination:

[0171] Based on cleaned social network interaction data, a social network propagation graph is constructed with virtual social accounts as nodes and interaction frequency and emotional resonance as edge weights, where edges are constructed according to the social interactions between nodes. The betweenness centrality and community clustering coefficient of the virtual accounts corresponding to risk anomalies in the graph are calculated to quantify their propagation influence in the public opinion network. The label propagation algorithm is used to simulate the flow path of negative emotions in the graph, and the cosine similarity between the emotion propagation flow and the gradient vector of the physical space is calculated. If the similarity is lower than a preset threshold and the propagation influence is higher than the average level of the community, the risk type is confirmed as a non-geographical emotional contagion event, and key propagation nodes in the graph are identified as secondary risk sources.

[0172] First, the system extracts four types of behaviors from historical interaction data: "reply", "quote", "like" and "jointly participate in topics", and constructs a weighted directed graph; nodes represent virtual accounts, edges represent interaction relationships, and the weights are determined by the frequency of interaction and the similarity of sentiment semantics.

[0173] Secondly, the Louvain algorithm is used to divide the graph into communities and identify implicit communities with different themes, such as mutual aid communities; the betweenness centrality of abnormal nodes is calculated to determine whether they are in the "throat" position of information flow; if the betweenness is high, it means that the node is a key bridge for the spread of emotions across communities.

[0174] Finally, the emotional flow paths (virtual space) in the graph are mapped back to the physical grid to form a virtual propagation vector. This vector is then compared with the spatial gradient vector of the physical potential field using cosine similarity. If the two directions are significantly divergent—for example, the physical gradient points downstairs, but the virtual propagation vector points to a resident in Phase II one kilometer away—it indicates that the risk propagation does not follow physical laws. Based on this, it is identified as an emotionally contagious event, and super-spreader accounts that connect different communities in the graph are marked as key targets for early warning and intervention.

[0175] This solution overcomes the limitations of physical space analysis by constructing a social graph, visualizing the interpersonal networks hidden behind physical buildings, and accurately identifying the risk of emotional resonance that transcends physical isolation. It improves qualitative accuracy by calculating the similarity between virtual propagation vectors and physical gradient vectors, using proof by contradiction (i.e., if the propagation path does not conform to physical laws, it indicates emotional contagion) for double verification. This effectively avoids misjudging accidental physical proximity events as mass public opinion events or misjudging hidden public opinion spreads as isolated events. Furthermore, it enables targeted governance by identifying key propagation nodes and hidden communities in the graph. Early warning signals no longer only point to a single physical room number but can further pinpoint opinion leaders or the origin of emotions, providing community managers with a scientific basis for cutting off the transmission chain and conducting precise psychological counseling.

[0176] The risk warning module predicts the trajectory of risk evolution based on the type of risk event, the distribution of potential energy value, and the spatial gradient vector. When the predicted risk potential energy exceeds the safety threshold, it generates a warning signal containing the coordinates, type, and diffusion range of the risk source.

[0177] Before the step of predicting the risk evolution trajectory, the method also includes a step of risk root cause analysis using a causal knowledge graph: constructing a causal knowledge graph for the community governance domain containing triples of "risk entity-risk event-risk consequence"; using entity linking technology to map semantic labels in the feature data space to entity nodes in the knowledge graph, and mining deep-seated causal paths leading to the current risk anomaly through graph path search algorithms; verifying the effectiveness of the causal paths by analyzing the time lag correlation between different dimensional stress indices based on Granger causality tests, and assigning high-weight evolutionary driving factors to the verified root cause nodes, which are then input into the evolution prediction model.

[0178] This embodiment achieves the tracing of origins from phenomena to essence by constructing a causal graph.

[0179] First, pre-build an ontology library containing common sense about community governance, for example, defining the rule: "Long-term noise (entity) -> leads to -> decreased sleep quality (consequence) -> triggers -> emotional change (event) -> induces -> risk event (consequence)".

[0180] By using natural language processing technology to extract causal triples from historical work orders and legal documents, a causal knowledge graph for community governance is constructed. In real-time monitoring, when an anomaly such as a risk event is identified, it is anchored to a node in the graph through entity links, and a multi-hop path search is initiated.

[0181] The system may detect multiple paths; for example, path A leads to noise, while path B leads to parking space grabbing.

[0182] To identify the true root cause, historical time-series data was used to perform a Granger causality test. The analysis revealed that the change in the parking space service pressure index significantly preceded the change in the noise interference pressure index on the time axis, and this was statistically significant. Based on this, parking space hogging was determined to be the root cause, and noise-related interference paths were pruned.

[0183] Traditional early warning systems often target surface symptoms, while this solution uses graph reasoning to uncover deep-seated structural problems, enabling governance measures to directly address pain points and prevent the recurrence of similar risks. Secondly, it improves the robustness of the predictive model by eliminating spurious variables through Granger causality tests, selecting driving factors with genuine causal force to input into the model. This avoids overfitting or incorrect predictions caused by spurious correlations in machine learning models, significantly improving the accuracy of evolutionary predictions in complex scenarios. Finally, it enhances the interpretability of early warnings. The generated warning reports not only include red risk areas but also generate clear causal evidence chains, i.e., graph reasoning paths, allowing community managers to intuitively understand the logical chain of risk formation.

[0184] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A risk assessment and early warning method based on big data analysis, characterized in that, include: Collect multi-source heterogeneous data through the community IoT sensing layer, management service layer, and social network layer; By using natural language processing and address matching algorithms, a dynamic mapping relationship between virtual social accounts and physical real estate units is established. Multi-source heterogeneous data is cleaned and anchored to the community geographic grid to form a feature data space containing spatiotemporal attributes. The process of constructing the feature data space includes: Based on the identification and processing of multi-source heterogeneous data, physical stress index, emotional stress index and service stress index are obtained; Using physical property units as the smallest geographic grid unit, and utilizing the established dynamic mapping relationship, the physical stress index, emotional stress index, and service stress index are accurately projected onto the corresponding geographic grid coordinates. Using a preset time window as the granularity, multi-source indices within the same geographic grid are fused to generate a comprehensive state vector for the grid; the comprehensive state vectors of all geographic grids in continuous time series are combined to construct a multi-dimensional feature data space that dynamically reflects the overall operation of the community. A community risk identification model is constructed to analyze the feature data space, identify risk anomalies in the feature data space, and transform the risk anomalies into a continuously distributed risk potential energy field; the potential energy value distribution of the risk potential energy field is calculated to quantify the degree of risk aggregation in physical space; the spatial gradient vector of the potential energy field is calculated to characterize the diffusion direction and rate of risk in spatial dimension; and the type of risk event is identified by combining the semantic labels of the feature data space with the physical space neighborhood relationship. The process of obtaining the risk potential energy field includes: Traverse each geographic grid in the feature data space, use the Euclidean norm algorithm to obtain the magnitude of the comprehensive state vector, and define it as the comprehensive state value; monitor the fluctuation of the comprehensive state value, and mark it as a discrete risk anomaly when the state value exceeds the threshold set based on the historical baseline. By introducing a physical field theory model and using a kernel density estimation algorithm, the radiation influence range and attenuation coefficient of each risk anomaly point are calculated by combining the Gaussian kernel function. The radiation impact values ​​of all risk anomalies are spatially superimposed, and a time decay factor is introduced in the time dimension to accumulate the residual impact of historical risks. The discretely distributed risk anomalies are smoothly mapped into a continuous risk potential energy field covering the entire geographical area of ​​the community. The value of each point in the risk potential energy field represents the risk potential energy value of the current location. Based on the type of risk event, the distribution of potential energy value, and the spatial gradient vector, the trajectory of risk evolution is predicted. When the predicted risk potential energy exceeds the safety threshold, an early warning signal containing the coordinates, type, and diffusion range of the risk source is generated.

2. The risk assessment and early warning method based on big data analysis according to claim 1, characterized in that: The process of acquiring the multi-source heterogeneous data includes: The community IoT sensing layer connects to IoT sensors to collect physical sensing data including environmental noise decibel values, overflowing trash cans, images of objects thrown from high-rise buildings, and elevator vibration frequency. The management service layer connects to the property management system and public service hotline interface, and collects service management data including inspection logs of public facilities, processing time of repair work orders, and written records of resident complaints. The social network layer accesses homeowner communication groups and community forums, collecting unstructured social data including publicly posted communication texts, image information, and likes, comments, and interaction data from residents; The collected heterogeneous data from multiple sources are tagged with timestamps to construct a multi-source heterogeneous data system covering physical environment status, service response efficiency, and community sentiment.

3. The risk assessment and early warning method based on big data analysis according to claim 1, characterized in that: The process of establishing the dynamic mapping relationship includes: Establish a standard physical address database for the community and assign a unique spatial code to each independent physical property unit. For unstructured social data collected from the social network layer, the building number, unit number and floor location features implicit in the text are extracted; the semantic dependency relationship between the extracted address features and the user's first-person pronoun is parsed using a dependency parsing algorithm to determine whether the user has the residential attribute of the address. Simultaneously, regular expressions are used to perform pattern matching on the numerical sequences in users' social nicknames, and cross-validation is performed with the owner information that has been confirmed in the property management system. Based on the validation results, the virtual social accounts are bound to the corresponding physical property unit spatial codes to generate a dynamic mapping relationship table.

4. The risk assessment and early warning method based on big data analysis according to claim 1, characterized in that: The process of obtaining the physical stress index, emotional stress index, and service stress index includes: Feature extraction and normalization are performed on physical sensing data, continuous environmental monitoring values ​​are mapped into interference indexes, discrete abnormal events are converted into occurrence frequencies, and a weighted fusion algorithm is used to generate a physical stress index that characterizes the state of environmental facilities. The negative sentiment confidence level of unstructured social data is calculated using a sentiment semantic analysis model, and the social dissemination weight is calculated by combining the interaction popularity and dissemination breadth of the data. The negative sentiment confidence level is then weighted and corrected using the social dissemination weight to generate an emotional stress index that represents the intensity of group emotions. Extract work order response time and maintenance handling records from service management data, calculate the deviation of actual processing time from standard time, combine with missing maintenance records to conduct performance evaluation, and generate a service pressure index that characterizes the degree of lag in property services.

5. The risk assessment and early warning method based on big data analysis according to claim 1, characterized in that: For the potential energy distribution of the risk potential energy field, the peak center of the potential energy distribution is determined by calculating the local maxima of the risk potential energy field, thereby quantifying the high-density clustering area of ​​risk in physical space. Given the spatial gradient vector of the potential energy field, the vector field of the spatial gradient of the risk potential energy field is obtained, and the spatial gradient vector of the potential energy field is obtained. The magnitude of the spatial gradient vector represents the rate of risk diffusion, and the vector direction represents the direction of risk diffusion in physical space. Based on the type of risk event, extract the original text semantic tags corresponding to the risk anomaly points, and distinguish between physically transmitted events and emotionally contagious events according to the tag type and the spatial gradient vector direction.

6. The risk assessment and early warning method based on big data analysis according to claim 1, characterized in that: The current risk event type, potential energy distribution, and spatial gradient vector are input into an evolutionary prediction model based on a long short-term memory network to simulate the changes in the risk potential energy field within future time steps and generate a risk evolution trajectory. The system monitors the peak potential energy in the evolution trajectory in real time. Once the risk potential energy value at a future moment exceeds the preset safety threshold, an early warning mechanism is immediately triggered. The system traces back to the starting grid coordinates that caused the potential energy to exceed the limit as the risk source coordinates. The risk type is determined based on the identified event semantics, and the risk diffusion range is determined based on the predicted potential energy field coverage contour lines, thus generating an early warning signal.

7. A risk assessment and early warning system based on big data analysis, characterized in that: The data acquisition module collects multi-source heterogeneous data through the community IoT sensing layer, management service layer, and social network layer; The data processing module uses natural language processing and address matching algorithms to establish a dynamic mapping relationship between virtual social accounts and physical real estate units. After cleaning multi-source heterogeneous data, it anchors it to the community geographic grid to form a feature data space containing spatiotemporal attributes. The process of constructing the feature data space includes: Based on the identification and processing of multi-source heterogeneous data, physical stress index, emotional stress index and service stress index are obtained; Using physical property units as the smallest geographic grid unit, and utilizing the established dynamic mapping relationship, the physical stress index, emotional stress index, and service stress index are accurately projected onto the corresponding geographic grid coordinates. Using a preset time window as the granularity, multi-source indices within the same geographic grid are fused to generate a comprehensive state vector for the grid; the comprehensive state vectors of all geographic grids in continuous time series are combined to construct a multi-dimensional feature data space that dynamically reflects the overall operation of the community. The risk identification module constructs a community risk identification model to analyze the feature data space, identify risk anomalies in the feature data space, and transforms the risk anomalies into a continuously distributed risk potential energy field; calculates the potential energy value distribution of the risk potential energy field to quantify the degree of risk aggregation in physical space; calculates the spatial gradient vector of the potential energy field to characterize the diffusion direction and rate of risk in spatial dimensions; and identifies the type of risk event by combining the semantic labels of the feature data space with the physical space neighborhood relationship. The process of obtaining the risk potential energy field includes: Traverse each geographic grid in the feature data space, use the Euclidean norm algorithm to obtain the magnitude of the comprehensive state vector, and define it as the comprehensive state value; monitor the fluctuation of the comprehensive state value, and mark it as a discrete risk anomaly when the state value exceeds the threshold set based on the historical baseline. By introducing a physical field theory model and using a kernel density estimation algorithm, the radiation influence range and attenuation coefficient of each risk anomaly point are calculated by combining the Gaussian kernel function. The radiation impact values ​​of all risk anomalies are spatially superimposed, and a time decay factor is introduced in the time dimension to accumulate the residual impact of historical risks. The discretely distributed risk anomalies are smoothly mapped into a continuous risk potential energy field covering the entire geographical area of ​​the community. The value of each point in the risk potential energy field represents the risk potential energy value of the current location. The risk warning module predicts the trajectory of risk evolution based on the type of risk event, the distribution of potential energy value, and the spatial gradient vector. When the predicted risk potential energy exceeds the safety threshold, it generates a warning signal containing the coordinates, type, and diffusion range of the risk source.

Images