Slope early warning method and system based on knowledge graph

Abstract

The invention discloses a slope early warning method and system based on a knowledge graph, and relates to the technical field of multi-source monitoring slope early warning, and the method comprises the steps: 1, unifying laser, radar, satellite remote sensing and underground water level monitoring flow into an index-structure-disaster situation graph containing traceability metadata; 2, generating an instability trigger coefficient by using a gradient lifting model driven by a coupling trend and credibility; step 3, automatically outputting short messages, broadcasts and social contact templates, and implanting abstract codes for publishing; 4, comparing the returned text in real time, marking deviation nodes according to graph consistency, and tracing a source link; and step 5, performing Hash solidification on the core early warning segment and spreading the performance to the alliance chain, and writing back the knowledge graph optimization template and channel. According to the scheme, multi-source data alignment is achieved, early warning is accurate and reliable, cross-medium semantic consistency and evidence chain traceability are achieved, and the requirement of the whole mountain disaster prevention process is met.

Description

Technical Field

[0001] This invention relates to the field of multi-source monitoring slope early warning technology, specifically to a slope early warning method and system based on knowledge graphs. Background Technology

[0002] Against the backdrop of rapid expansion in mountainous transportation and hydropower projects, numerous steep slopes have been excavated along major transportation routes, power transmission corridors, and reservoir banks. These slopes are constantly exposed to a complex load environment of heavy rainfall, frequent vibrations, and human disturbance. Instability in these slopes could potentially break roads, block waterways, or impact factories, causing a chain reaction of disasters from points to lines to surfaces. To mitigate these risks, the industry has gradually deployed multi-source monitoring networks, including laser point cloud scanning, ground-based synthetic aperture radar, low-orbit satellite interferometry, groundwater level wells, and ground acoustic microseismic monitoring, supplemented by multi-channel information dissemination systems such as SMS, emergency broadcasts, and social media. However, current technologies largely rely on closed data formats and algorithms from individual manufacturers: laser systems output point clouds with local coordinates, radar provides polar coordinate phase gratings, satellite products use international geodetic coordinates, and groundwater level wells record scattered text; at the dissemination end, SMS gateways, radio stations, and microblog interfaces are incompatible. Engineers often have to manually align coordinates, manually cut and paste key sentences, and manually verify receipts. With the surge in monitoring frequency and the number of devices, manual operations are increasingly unable to support minute-level early warnings, leading to a situation where "fragmented indicators, simplified mechanisms, and delayed decision-making" have become the norm. In addition, regulations increasingly emphasize the traceability of disaster information and joint confirmation by multiple agencies, making it difficult for existing point-based solutions to provide a complete technical loop from data collection to legal evidence preservation.

[0003] Regarding the need for rapid early warning of instability under combined loads on steep slopes, existing solutions reveal core flaws in the "monitoring-reasoning-issuance-verification-evidence" chain: There is a lack of a technical framework that can uniformly map heterogeneous monitoring streams, prior geological structures, and historical disaster records into a computable semantic space, and integrate data lineage, credibility assessment, cross-media consistent propagation, and tamper-proof evidence storage. Specifically, when lasers, radars, satellites, and water level wells exhibit high heterogeneity in sampling frequency, coordinate references, and noise characteristics, traditional systems can only align using the least common denominator, resulting in a loss of geometric accuracy; multi-source sequences cannot be assigned credibility weights before coupled inference, causing occasional noise to be amplified into false alarms; the same warning text, after being truncated or edited by users on SMS, broadcasts, and social media platforms, lacks automated verification methods, leading the public to mistakenly believe that information discrepancies are due to "official policy changes"; finally, even if engineers manually organize key segments, there is a lack of on-chain evidence storage paths for cross-institutional multi-signature consensus, which cannot meet the evidentiary standards for post-event accountability and insurance claims.

[0004] The aforementioned defects are particularly prominent in high-interference scenarios such as rainstorms at night or construction blasting: monitoring link latency and packet loss surge, manual alignment becomes even more delayed, false alarms or missing reports directly hinder on-site evacuation, and may even intensify public opinion on social media platforms, causing a secondary impact on the credibility of local governments and operating units.

[0005] Therefore, this invention provides a slope early warning method and system based on knowledge graphs. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a slope early warning method and system based on knowledge graphs. It integrates laser, radar, satellite remote sensing, and groundwater level monitoring data into an "indicator-structure-disaster" graph containing source metadata; it uses a gradient boosting model driven by coupling trends and credibility to generate instability triggering coefficients; it automatically outputs SMS, broadcast, and social media templates with embedded digest codes for publication; and it compares the returned text in real time, based on... Figure 1 The solution identifies and traces source links for deviation nodes, hashes and solidifies core early warning segments and propagation performance data to the consortium blockchain, and writes back to the knowledge graph to optimize templates and channels. This approach achieves multi-source data alignment, accurate and reliable early warning, cross-media semantic consistency, and traceable evidence chains, meeting the needs of the entire disaster prevention process in mountainous areas, thus solving the technical problems described in the background section.

[0007] To achieve the above objectives, the present invention provides the following technical solution: Knowledge graph-based slope early warning methods and systems include, based on lidar, satellite remote sensing and groundwater level monitoring, constructing a spatiotemporal network system of observation indicators, geological structures and historical disasters containing source-tracing metadata in the knowledge graph; The coupling trend between displacement and pore water pressure and the reliability of the data are calculated. The instability triggering coefficient is output using a gradient boosting model and written into the warning information entity along with the triggering factor. Based on the semantic relationship between the entities of the warning information, the transmission carrier, and the receiving subject, SMS, broadcast, and social media templates are automatically generated, a unique digest code is injected, and they are released in parallel to maintain semantic consistency across media. Real-time acquisition of text returned from various channels, categorized by digest code, timestamp, and content differences. Figure 1 Consistency rule comparison; if the threshold is exceeded, the node with information deviation is marked and the propagation path is traced back. The verified core early warning segments and propagation performance are used to generate hash fingerprints, which are then written into the consortium blockchain through multi-node consensus. Finally, the on-chain fingerprints and performance data are fully synchronized and written back to the knowledge graph.

[0008] Furthermore, a link reliability index is used to dynamically screen the laser, radar, satellite remote sensing, and groundwater level acquisition channels. When the link reliability index is lower than 0.85, the corresponding data packets are automatically removed, and the removal reason and timestamp are registered in the knowledge graph to ensure the accuracy of time sequence alignment and the integrity of data traceability.

[0009] Furthermore, the relationship between the observation index nodes and the geological structure nodes is established by weighting three factors: geological confidence, observation quality, and frequency of historical disasters. The edge weights are limited to the interval of 0 to 1 by logical normalization mapping, which dynamically weakens weak correlation paths and improves the accuracy of subsequent inference.

[0010] Furthermore, the displacement sequence is subjected to second-order difference accumulation, the pore water pressure sequence is subjected to multi-scale gradient calculation, and then continuous wavelet cross-correlation is performed and normalized according to energy to generate a coupling trend quantity in the range of 0 to 1. The data credibility index is calculated using residual entropy and sensor health mutual information, and both are used as inputs to the gradient boosting model.

[0011] Furthermore, the coupling trend quantity, real-time confidence index, and first-lag confidence index are scaled by quantile distance to form a three-dimensional feature tensor. The gradient boosting model achieves high recall and controlled false positive rate in class imbalance scenarios by combining the search learning rate, maximum tree depth, and subsample ratio through an adaptive differential evolution algorithm.

[0012] Furthermore, the semantic layer-representation layer mapping engine is used to convert the warning information entity into three templates: SMS, broadcast, and social media. Each template is embedded with a unique digest code and adaptive typesetting based on media features is performed. The templates are only released after the semantic restoration error is verified by reverse parsing to be no more than one percent.

[0013] Furthermore, priority scores are calculated for all pending messages based on coverage, real-time index, and interaction coefficient, and messages are pushed in batches in parallel based on these priority scores. After each message is published, the first receipt time is recorded and the channel sorting for the next batch is updated in real time to achieve dynamic load balancing under bandwidth-constrained conditions.

[0014] Furthermore, the digest code is first verified on the returned text, and then the content difference index is generated by fusing semantic vector cosine similarity and weighted edit distance. If the content difference index is higher than a preset threshold, the corresponding propagation carrier node is marked as an information deviation node and written into the deviation level attribute.

[0015] Furthermore, based on the time weight, channel weight, and semantic similarity weight, the edge cost is synthesized, and the minimum cost path search is performed upward from the deviation node to determine the difference source node. The impact score is calculated by combining the number of covered users, interaction activity, and re-propagation index, which is used to dynamically adjust the risk level.

[0016] Furthermore, the trigger coefficient, trigger factor, and three propagation performance indicators are concatenated into text, and a template version number and chain writing time description are added. A 384-bit on-chain fingerprint is generated using an anti-collision hash algorithm. Subsequently, the fingerprint is written to the consortium blockchain after multi-signature confirmation by the supervisor, emergency response, insurance, and scientific research nodes.

[0017] Furthermore, after hearing the on-chain transaction confirmation event, the block height, confirmation time description, and the on-chain fingerprint are written back to the knowledge graph. The template real-time score is updated based on the initial reach time and difference correction time. At the same time, the channel trust score is re-estimated based on the digest code verification success rate and chain writing delay, in order to adaptively optimize the subsequent template selection and release priority.

[0018] A knowledge graph-based slope early warning system includes, The data acquisition unit, based on lidar, satellite remote sensing and groundwater level monitoring flow, constructs a spatiotemporal network system of observation indicators, geological structures and historical disasters containing source traceability metadata in the knowledge graph; The data processing unit calculates the coupling trend between displacement and pore water pressure and the reliability of the data, uses a gradient boosting model to output the instability triggering coefficient, and writes it along with the triggering factor into the early warning information entity; The generation unit automatically generates SMS, broadcast, and social media templates based on the semantic relationship between the warning information entity, the transmission carrier, and the receiving subject, injects a unique digest code, and publishes them in parallel to maintain semantic consistency across media. The backtracking unit acquires text returned from various channels in real time, categorizing it by digest code, timestamp, and content differences. Figure 1 Consistency rule comparison; if the threshold is exceeded, the node with information deviation is marked and the propagation path is traced back. The feedback unit generates hash fingerprints from the verified core early warning segments and propagation performance, writes them into the consortium blockchain through multi-node consensus, and fully synchronizes the on-chain fingerprints and performance back to the knowledge graph.

[0019] This invention provides a slope early warning method and system based on knowledge graphs, which has the following beneficial effects: By integrating four types of real-time monitoring streams—laser, radar, satellite remote sensing, and groundwater level—into a single knowledge graph through unified coordinate regularization, link reliability screening, and source metadata binding, the originally discrete and easily conflicting data can achieve coordinate and time alignment in milliseconds. As a result, subsequent inference only requires a single call to obtain multi-source features of the same dimension, completely eliminating the redundant correction costs of the traditional monitoring system where "there are multiple sets of benchmarks for the same location."

[0020] By using dual quantification of displacement and pore water pressure coupling trend and data credibility index, the physical mechanism and statistical noise are complemented at the feature level. Combined with the adaptive parameter search of the gradient boosting model, the true signs of instability are amplified while the occasional noise is automatically suppressed. Compared with commonly used single threshold or static models, the early warning accuracy is significantly improved without relying on any additional hardware, while maintaining a low false alarm rate.

[0021] The semantic layer-presentation layer mapping engine, media feature adaptive typesetting, and unique digest code work together to automatically generate multimodal messages that adapt to SMS, broadcast, and social media for the same warning information, and ensure cross-channel semantic reversibility. This not only upgrades template maintenance from manual cutting and pasting to second-level automation, but also provides a natural fingerprint for the subsequent difference verification module through digest code, realizing seamless connection between the generation end and the verification end.

[0022] The return differential quantization engine employs a fusion strategy of semantic vectors and weighted edit distance, combined with... Figure 1 Consistency rules identify deviation nodes and weighted backtracking paths, enabling the identification of the source node that first introduced the deviation within seconds of message publication, and immediately quantifying its potential impact based on the number of people reached and the level of interaction activity.

[0023] The dual-domain summary solidification and multi-signature consensus chain writing mechanism solidifies the core early warning semantics and propagation performance into the consortium chain in one go. It also writes the chain writing time, verification success rate and template real-time score back to the knowledge graph by listening to on-chain receipts, driving the survival of the fittest templates and the dynamic rebalancing of channel trust. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the slope early warning method of the present invention; Figure 2 This is a schematic diagram of the slope early warning method system of the present invention. Detailed Implementation

[0025] 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.

[0026] Please see Figure 1 This invention provides a slope early warning method based on knowledge graphs, including: In real-world engineering projects, slope monitoring data often comes from multiple technical sources, including laser point cloud scanning, ground-based synthetic aperture radar, low-orbit satellite imagery, and groundwater wells. Each source naturally differs in spatial coordinates, time references, sampling resolution, and signal-to-noise ratio, and may even contain random interruptions and incomplete information. If these observation streams are stored side-by-side, cascading errors such as coordinate inaccuracies, time drift, and packet loss distortion will occur, ultimately leading to inaccurate location of hazard signs, unbalanced threshold settings, and difficulty in tracing the responsible parties.

[0027] By utilizing a ternary knowledge graph framework of observational indicators, geological structures, and historical disasters, scattered observations are translated into a spatiotemporal entity network that is reasonable, traceable, and version-manageable. Metadata conforming to W3C traceability standards is embedded at the network's underlying layer. Step one focuses on building a one-to-one mapping relationship between scattered monitoring streams and a unified semantic web, and meticulously records the entire mapping process to ensure that subsequent instability mechanism reasoning and blockchain fingerprint writing can use the same semantics and source benchmark.

[0028] Step 1: Establish a holographic, traceable, and queryable STEN to ensure that subsequent mechanism reasoning and on-chain fingerprint writing have a unified and reliable data semantic benchmark.

[0029] Step 101: Unified Spatiotemporal Semantics of Multi-Source Observation Streams Data acquisition link topology mapping: The multi-source monitoring system covers at least four types of terminals: laser scanners, radar antenna arrays, satellite ground receiving stations, and water level monitoring wells. For each type of terminal, a directed acyclic topology of node channels is first abstracted, defining the terminal as a node and the fiber optic, wireless microwave, or wired Ethernet connection as a channel. Each channel is dynamically assigned three sets of real-time characteristics within the same monitoring period: media transmission delay, node clock drift, and packet loss ratio. The system continuously collects these three sets of characteristics through distributed monitoring scripts, multiplies and superimposes the characteristics, and then introduces an exponential decay factor to obtain the link reliability index. The closer the link reliability index is to one, the more stable the transmission, the higher the synchronization accuracy, and the less data loss there is in the current monitoring period. Conversely, if the reliability index is lower than the set threshold of 0.85, the system immediately marks the corresponding channel as a weak link and outputs a maintenance alarm to prevent the weak link from affecting subsequent map timing alignment operations.

[0030] In this scheme, link reliability metrics act as a screening gate. Imagine that severe convective weather at night causes a brief attenuation of the microwave link. If weak links are not eliminated in advance, the radar phase information will carry random delays. This erroneous delay will manifest as irregular deformation pulses during spatial registration, directly masking the true displacement trend.

[0031] Coordinate regularization and unified survey area benchmark: Laser point clouds and radar measurements reside in the local engineering coordinate system and the equipment's built-in polar coordinate system, respectively, while satellite imagery uses the International Geodetic Coordinate System (IGCRS). To achieve spatial consistency, a local independent survey area coordinate system is first established based on the three-dimensional coordinates of the permanent survey markers in the survey area. Then, a seven-parameter rigid transformation model is used to transform the satellite image coordinates to the survey area coordinate system; the radar polar coordinates are transformed to the same coordinate system using the installation attitude calculation matrix; finally, the laser point clouds are superimposed onto a unified model through coplanar bundle adjustment. During the transformation process, translation vectors, rotation matrices, and microscale scaling factors are recorded in real time and written into the spatial calibration field of the observation records. After regularization, the multi-source displacement vectors from the same starting point are at a consistent millimeter scale across any time slice, allowing subsequent algorithms to directly perform cross-modal comparisons.

[0032] The core function of coordinate regularization is to ensure that millimeter-sized depressions detected by laser scanning and phase accumulation abrupt changes detected by radar fall on the same three-dimensional location, rather than two unrelated points in the ground model. Through a hybrid rigid and micro-rigid transformation strategy, spatial consensus is established across different resolutions, viewpoints, and imaging mechanisms, essentially providing all monitoring terminals with the same set of rulers and avoiding false displacement misjudgments caused by scale drift.

[0033] Source metadata binding and version snapshot: For each raw observation, whether it's a single radar phase grid or a water level time series, four elements are recorded simultaneously: observation file, sensor serial number, firmware version, and sampling time. A 256-bit chain fingerprint is then generated using an encrypted hash algorithm. Once written to the ancestor node, the chain fingerprint is immediately locked, and any subsequent processing will reference this fingerprint as the ancestor node.

[0034] Meanwhile, following the total data accumulation strategy, a version snapshot is automatically generated every gigabyte, mapping the snapshot to an entity version node in the knowledge graph. When engineers query in the future, they can trace back to the original observation files and hardware information along the entity version nodes, ensuring that records are traceable, responsibilities are pinpointable, and errors are explainable.

[0035] When regulatory authorities need to verify the basis of an early warning, engineers only need to retrieve the entity version node for that day to read the corresponding chain fingerprint, further trace back all ancestor nodes, and reproduce the original observation file and processing script. This mechanism not only avoids human deletion and modification but also improves legal credibility.

[0036] The unified spatiotemporal semanticization process for multi-source observation streams addresses three major pain points in traditional monitoring systems: opaque channel vulnerabilities, conflicting coordinate frameworks, and missing data lineage. This is achieved through a combination of three technical features: link reliability screening, coordinate regularization, and source-tracing metadata binding. Link reliability screening ensures that the original data entering the semantic web has a stable transmission background; coordinate regularization allows multimodal deformation information to appear within the same reference frame, achieving millimeter-level alignment; and source-tracing metadata binding and version snapshots ensure that each piece of data has an irrefutable origin and timestamp, meeting the needs of engineering, regulations, insurance, and other parties.

[0037] Step 102: Constructing a disaster ternary entity-relationship network based on indicators. Entity extraction and unique identifier mapping: To ensure the knowledge graph accurately describes the slope target, a deep attention sequence annotation algorithm is used to batch parse monitoring logs, geological survey reports, and historical landslide archives, extracting dual entities: observation indicator descriptions and geological structure descriptions. After extraction, the minimum description length principle is applied to evaluate entity label complexity, and embedding similarity is used to determine name conflicts. If synonyms or duplicate names are found, the label that minimizes the overall description complexity and maintains the clearest semantics is selected as the retained item, and other labels with the same name are mapped to aliases.

[0038] Ultimately, each entity is assigned a unique four-part code, comprising the slope number, entity category, semantic tag hash value, and creation timestamp, fundamentally eliminating ambiguity. In massive amounts of text, terms like "primary slip surface" and "secondary slip surface" frequently appear. Directly inputting these into the knowledge graph would generate a vast number of overlapping nodes during subsequent reasoning. By employing a dual filtering approach—minimum description length and embedding similarity—similar items are merged at the entity level, ensuring that queries can locate unique nodes with a single click, avoiding unnecessary detours in reasoning.

[0039] Relation instantiation and graph connectivity optimization: After the entity nodes are established, it is necessary to define the index association to construct this core relationship. The strength of the relationship is driven by three sets of evidence: geological confidence, observation quality score, and historical disaster frequency. Geological confidence is obtained through borehole core imaging and structural surface scanning scores; the observation quality score is the weighted result of the link reliability index and the three-dimensional registration error; the historical disaster frequency is calculated by counting the number of landslide events recorded within the same structural zone within the past fifty years.

[0040] The three sets of evidence are summed according to set weights, and then compressed to the range of zero to one using a logical normalization function to generate relation weights. When the relation weight is below 0.2, the corresponding edge is automatically downweighted or even pruned; when the weight is greater than 0.8, it is marked as a priority reasoning path in the graph database. In this way, high-confidence monitoring indicators will be prioritized for matching the main structural control factors, while weak evidence will be marginalized, avoiding noise affecting decision-making.

[0041] Suppose a fault zone has triggered multiple landslides over the past half-century, and the monitoring link has maintained high quality over a long period. In this case, the edge constructed by this indicator will be assigned extremely high weight. During the model training phase, the gradient boosting algorithm will pay more attention to this path, forming an adaptive threshold with high risk even with few samples, preventing blind spots caused by traditional fixed thresholds.

[0042] Temporal persistence and query optimization: Knowledge graphs expand rapidly over time, and without a proper indexing structure, online queries quickly become slow. To address this, a dual-indexing system—time window index and entity hash index—is designed at the graph database layer. All relation edges are first segmented by time windows and then arranged according to entity hash prefixes. Each query first locates the time segment, then narrows the range based on the entity prefix, reducing query complexity from linear traversal of the entire graph to logarithmic levels. Tests show that even when the entire graph reaches 100 million edges, this dual-indexing structure maintains sub-second response times, meeting real-time alert requirements.

[0043] When a sudden query request is triggered by a nighttime rainstorm, it is only necessary to retrieve the edge set from the most recent 24 hours and use a hash prefix within the set to match the specific entity. Since the query does not need to access the old edge set, the number of disk seeks is greatly reduced, latency is sharply reduced, and the early warning logic chain is ensured to be smooth.

[0044] The knowledge graph leverages three mutually reinforcing technical features: a unique code system to resolve ambiguity conflicts, multi-evidence weighting to enable self-evolution of relationships, and a dual-index structure to maintain real-time searchability of the massive graph. The unique code system binds any semantic reference to a single object, eliminating ambiguity; multi-evidence weighting welds real-time monitoring, geological background, and disaster experience into a single relationship, forming a dynamic, adaptive risk view; and the dual-index structure ensures that the vast network of relationships does not slow down inference. Combined, the knowledge graph is not merely a static storage table, but a dynamic inference foundation that self-adjusts with time and data quality. Traditional approaches often disperse entity disambiguation, relationship weighting, and query acceleration across different software modules. Deeply coupling these three elements creates a synergistic benefit greater than the sum of its parts, significantly improving early warning accuracy and response efficiency.

[0045] The core foundation of the semantic web was built through two main steps: unified spatiotemporal semanticization of multi-source observation streams and the construction of a three-element entity-relationship network for disaster indicators. For the first time, slope monitoring data achieved comprehensive consistency across spatial, temporal, and legal dimensions, forming a real-time computable dynamic knowledge structure within the graph database. Subsequent steps can directly utilize unified entity codes, relation weights, and temporal indexes to generate displacement-pore-pressure coupling trends, instability triggering coefficients, and multi-channel early warning templates, without requiring any coordinate transformations or data lineage verification.

[0046] For steep slopes, displacement curves and pore water pressure curves often differ significantly in sampling frequency, phase delay, and amplitude scale, and exhibit non-stationary characteristics due to multiple factors such as rainfall, temperature, blasting, and human interference. If the two curves are processed separately and then simply added together, the coupling information will be diluted due to scale misalignment; if the data reliability is not rigorously evaluated, individual distorted peaks may be mistaken by the algorithm for real instability symptoms. In step one, a complete spatiotemporal entity network has been constructed and the data lineage has been solidified. Step two further integrates the displacement and pore pressure sequences associated in the network into a coupling trend quantity. At the same time, information theory is used to evaluate the signal-to-noise level of each monitoring sequence. Then, the coupling trend quantity and the reliability index are fed into an adaptive gradient boosting model to generate interpretable instability triggering coefficients and write them back into the knowledge graph.

[0047] Step 2: Utilize cross-scale coupling quantity, information theory credibility, and adaptive gradient boosting algorithm to transform real-time monitoring data into interpretable instability triggering coefficients, and simultaneously write them into the early warning information entity to achieve seamless integration of digital risk assessment and source tracing records.

[0048] Step 201: Calculation of Coupling Trend Quantity and Assessment of Data Reliability Construction of displacement-orifice pressure coupling trend: Based on the fundamental assumption that deformation acceleration characterizes the external destructive force and pore pressure increment describes the internal seepage force, the displacement sequence is first differentially calculated twice and accumulated to upgrade the original displacement curve to an acceleration curve of the same dimension; the abrupt changes of minute displacements are amplified so that the slow extrusion that is difficult to distinguish with the naked eye appears as obvious pulses in numerical terms.

[0049] Subsequently, a segmented gradient transformation was performed on the pore water pressure curve: first, the increment at each time resolution was extracted using a separable filter, and then the increments at different levels were stored as scale slices. After time synchronization of the two increment curves was completed, a continuous cross-correlation scan was performed on the two curves using a pre-defined wavelet family as the kernel function. During the scan, both the time axis step size and the scale axis step size adopted an adaptive grid to ensure that there was no overfitting in the high sampling stage and no omissions in the low sampling stage. After energy normalization, the cross-correlation output was compressed to between 0 and 1. This normalization process completely preserved the amplitude ratio relationship while eliminating dimensional differences, enabling the system to directly determine the coupling strength using a single threshold.

[0050] The first few heavy rains of the rainy season often rapidly increase infiltration pressure, but surface displacement only becomes apparent after pore water seeps to the slip surface and weakens shear strength. Traditional methods using fixed-time-window cross-correlation cannot capture this cross-window information. By simultaneously sliding the kernel function along both the scale and time axes, it's equivalent to simultaneously translating and scaling a capture net in a two-dimensional plane. Whenever seepage and deformation forces occur synchronously at any scale, they will be marked as highly coupled by the energy normalization result. In this way, the system effectively extends the time for early identification of instability signs and avoids misjudging irrelevant noise as high coupling.

[0051] When the monitoring scenario includes high-frequency artificial vibrations or micro-vibrations of the equipment itself, short-duration spikes will appear on the displacement acceleration curve. If these spikes cannot be matched with corresponding pulses on the pore water pressure increment curve, the cross-correlation energy will be dispersed across multiple grid cells on the scale and time axes, causing the normalized output to drop to near zero. This means that the spikes are independent noise rather than actual coupling. This scenario verifies the ability of cross-correlation scanning to simultaneously dilute heterogeneous noise in the spatial and frequency domains, providing a unified criterion for distinguishing between laboratory simulations and actual field monitoring.

[0052] Multi-source residual entropy and mutual information weight fusion: While calculating the coupling trend, the quality of the input data must be quantified to prevent low-quality curves from artificially inflating the coupling score through accidental peaks.

[0053] Adhering to the dual principle that higher randomness equates to lower reliability, and higher health weight equates to higher reliability, the residual entropy of each monitoring sequence is first calculated on a daily basis. During the calculation, the residual probability distribution is statistically analyzed within a fixed observation window using the autoregressive residuals as a basis, and then the sequence randomness is obtained using the information entropy formula. Randomness close to zero indicates a highly predictable curve, i.e., low noise; randomness close to extreme values ​​indicates that the curve is approximately white noise, i.e., low information value.

[0054] Next, the sensor health attribute from the knowledge graph is invoked; this attribute is pre-evaluated based on indicators such as sensor manufacturing date, recent maintenance records, and link reliability. Subsequently, the system uses mutual information to quantify the overlap between the two distributions. If the residual entropy is high and the health is low, the overlap is low, and the credibility index is suppressed after multiplication; if the residual entropy is low and the health is high, the overlap is high, and the credibility index approaches one after multiplication. Finally, the credibility index is normalized to between zero and one and written into the real-time feature table.

[0055] Groundwater well levels often drift slowly due to temperature gradients and sediment deposition. If this drift coincides with a rainfall cycle, ordinary mean square error (MSE) is insufficient to distinguish between true trends and false drifts. Residual entropy, by focusing on probability distribution rather than absolute error, exhibits significant randomness in dealing with slow drifts. Furthermore, knowledge graphs provide information on sensor maintenance delays, leading to reduced mutual information and a substantial decrease in the reliability index. Thus, the algorithm does not overestimate its reliability simply because the data appears smooth.

[0056] In high-altitude and frigid regions, radar phase curves exhibit periodic "breathing" due to surface frost heave. This phenomenon appears ordered in the frequency domain but displays high randomness in the time domain. Residual entropy can accurately identify this implicit randomness, while the health weight is actively reduced due to frequent icing. The product of the two results in a confidence index of around 0.4. The system then reduces the weight of this curve to a secondary feature to prevent frost heave noise from dominating the coupled judgment.

[0057] Cross-validation of coupling quantity and credibility index: Having obtained the two core features, coupling trend and credibility index, it's important to understand that high coupling does not necessarily equate to high risk. A typical counterexample is the simultaneous shutdown of the water level well and displacement monitoring line during pump maintenance. This operation simultaneously weakens the dynamic changes of both curves; upon restarting, they simultaneously recover, creating a pseudo-coupling that appears highly correlated. To avoid such misjudgments, cross-validation thresholds are introduced: separate thresholds are set for coupling trend and credibility index. Only when both indicators simultaneously exceed their respective thresholds will the data row be marked as passed and moved to the machine learning training sample queue; otherwise, the monitoring backend will automatically push a secondary verification task, requiring manual or redundant sensor confirmation of the anomaly source.

[0058] The threshold values ​​are not hard-coded, but rather calculated by cross-validating the accumulated real unstable samples and safe samples from the previous calendar year. The recall and false alarm rates of real positive examples under various threshold combinations are statistically analyzed each year. The optimal dual thresholds are determined by constraining the recall rate to be no less than 80% and keeping the false alarm rate below 20%. In this way, the cross-validation thresholds are automatically updated based on the historical performance of the entire site, evolving in tandem with the monitoring environment.

[0059] When a single artificial blasting test causes significant oscillations across the entire monitoring sequence, the coupling trend can jump rapidly. However, the blasting duration is short, and sensor stability deteriorates instantaneously, leading to a simultaneous decrease in the confidence index. The threshold mechanism detects the contradiction of high coupling but low confidence and marks the event as pending confirmation. Subsequently, redundant accelerometers return a construction blasting label, and the event is deemed a risk-free signal.

[0060] By employing three mutually supportive technical features—cross-scale cross-correlation, information-theoretic quality assessment, and dual-threshold cross-validation—this approach achieves simultaneous quantification and dynamic verification of coupling strength and data reliability. Cross-scale cross-correlation ensures that asynchronous phenomena such as seepage precedence and deformation follow-up are not missed; information-theoretic quality assessment quantitatively reduces random noise and hardware drift; and dual-threshold cross-validation adds an automatic verification process for contradictory situations involving high coupling and low quality. This combination ensures that training samples are both representative and reliable, significantly reducing the risk of mislearning by the model due to anomalous noise, enhancing the physical rationality of feature mapping to the model, and exhibiting unique multidimensional self-consistency among similar solutions.

[0061] Step 202: Gradient boosting model training and writing instability trigger coefficients Feature tensor construction and scale uniformity: Deep tree models are prone to splitting bias when dealing with multi-scale data due to inconsistencies in feature dimensions. To address this, a three-step integrated operation is implemented during the sample construction phase. The first step involves vertically concatenating the coupling trend quantity, real-time reliability index, and first-order time-delay reliability index to form a three-element column vector. This preserves both the strength of coupling and incorporates the historical inertia of sensor stability.

[0062] The second step is to stack the continuous column vectors in chronological order within the sliding window to form a two-dimensional time-feature matrix. Then, the sliding window is used as a building block to continuously translate and construct a three-dimensional tensor, where the third dimension is the sample batch number.

[0063] The third step is to perform quantile scaling on all elements of the tensor: first calculate the difference between the 75th percentile and the 25th percentile, and then divide this difference by all elements to make the tensor fall within the symmetrical distribution range, thereby avoiding the influence of the long tail of data during extreme rainfall or extreme drought on the tree model splitting threshold.

[0064] When storing feature tensors, a label field is also attached to indicate the source entity key, graph version number, and sliding window index. These labels help upstream algorithms trace the source across versions and provide key material for subsequent fingerprint writing in consortium blockchains.

[0065] During the rainy season, coupling trend values ​​often cluster at high levels, while during the dry season they remain low for an extended period. If the two data points are directly concatenated, the first tree in the gradient boosting model tends to use the low values ​​from the dry season as global split points, resulting in all high-risk samples from the rainy season falling on the same leaf, leading to low information utilization. After quantile scaling, the high and low values ​​are pruned by the same amount, allowing the model to distinguish the coupling peaks from the rainy season and the low values ​​from the dry season into several risk levels, thus enriching the decision tree hierarchy.

[0066] In scenarios adhering to data security rules, feature tensors need to be anonymized before being written into the model. Quantile scaling, based on percentiles, is essentially a reversible linear transformation that requires quantile statistics to recover, preserving relative relationships while concealing absolute magnitudes. This approach protects sensitive commercial or governmental data while ensuring the model learns the correct relative risk gradient.

[0067] Dynamic hyperparameter tuning for gradient boosting models: Instability events are rare in nature and involve a high degree of class imbalance. Traditional static hyperparameters (such as fixed learning rate or fixed tree depth) struggle to simultaneously address recall and false positive rate. A differential evolution algorithm is employed to jointly search for three hyperparameters: learning rate, maximum depth, and subsample ratio. The differential evolution algorithm updates candidate solutions through three key operations: mutation, crossover, and selection; where the fitness function is defined as the percentage improvement in recall relative to the previous training epoch minus the false positive rate multiplied by a penalty coefficient.

[0068] Through this custom objective, the algorithm automatically compares the performance of the old and new models at the end of each training cycle. If the recall rate increases significantly while the false positive rate increases moderately, it is considered a positive gain; otherwise, it is rolled back. The entire process requires no manual intervention, yet maintains the model's resilience in extreme scenarios where positive samples are scarce.

[0069] In this scenario, the differential evolution algorithm also introduces a seasonal memory mechanism: at the end of each quarter, the algorithm writes the best-performing set of hyperparameters within the quarter into the memory pool, and uses these parameters as the parent parameter when the search begins in the next quarter to speed up convergence.

[0070] During the dry season, the number of unstable samples is almost zero. If the model training uses the high learning rate of the rainy season, it is prone to overfitting on a small number of negative samples, leading to a surge in the false positive rate. Dynamic hyperparameter tuning automatically lowers the learning rate and limits the tree depth when it detects that the false positive rate has exceeded the limit, making the model's decision boundary more conservative. It has been verified that although the recall rate has decreased slightly, the false positive rate has converged significantly, and the overall fitness has improved.

[0071] On edge nodes, computational resources are limited. Differential evolution algorithms only need to perform a forward evaluation of candidate solutions once each time, without having to backtrack all gradients, making them well-suited for devices with low computing power. By aggregating the evaluation results from multiple nodes back to the parent node, the entire computing cluster obtains the globally optimal set of hyperparameters with minimal resource cost.

[0072] Instability trigger coefficient generation and knowledge graph writing: The gradient boosting model outputs the probability value of each sample belonging to the instability category during the prediction phase. To ensure the engineering interpretability of the output, the probability values ​​are mapped to instability trigger coefficients via log-odds transformation. Log-odds have the advantages of monotonicity and additivity, which aligns with the reasoning habits of multi-source evidence superposition. Trigger coefficient range thresholds are set based on field experience; for example, 0.25 to 0.5 is designated as a medium-level warning, and greater than 0.5 is designated as a high-level warning.

[0073] To ensure that the trigger coefficients remain within the context of the input features, the one to three leaf node paths that contribute most to the prediction are extracted simultaneously. The corresponding splitting conditions are parsed, and textual descriptions of the trigger factors are generated, such as high coupling trend and stable confidence index. Subsequently, the knowledge graph interface is called to write the trigger coefficients and trigger factors as attributes into the warning information entity, and the entity version number and version fingerprint are updated.

[0074] The write timing is controlled by an event-driven mechanism: as long as the trigger coefficient crosses the threshold, the new version is pushed immediately; if the trigger coefficient fluctuates within the threshold, the write is delayed to prevent the knowledge graph version from exploding.

[0075] After a rainstorm, the model gives a probability value of 0.7 or higher for five consecutive time steps. After mapping using logarithmic probability, the trigger coefficient falls between 0.9 and 1, indicating a high-level warning. The new version of the knowledge graph is immediately implemented, with strong textual coupling to the archived trigger factors and stable sensor readings. Because the version fingerprint is generated based on the entire entity content, any subsequent tampering will lead to fingerprint mismatch, meeting the regulatory requirement that disaster information cannot be tampered with.

[0076] Different supervisory units can access knowledge graph entities according to their permissions. Trigger coefficients are presented in pure numbers, while trigger factors are presented in natural language. The combination of the two allows people with different professional backgrounds to quickly understand the warning level and cause; at the same time, version fingerprints provide legal credibility and prevent inconsistencies in information between multiple units.

[0077] By employing a cascading system of scale-uniform feature tensors, dynamic hyperparameter tuning, and trigger coefficient writing into the knowledge graph, the statistical learning output is strongly bound to the semantic entities of the knowledge. Feature tensor scaling addresses model skew caused by differences in distribution over time periods, dynamic hyperparameter tuning ensures high recall and controllable false positives even in scenarios where positive examples are scarce, and trigger coefficient writing back makes the digital quantification results readable and traceable.

[0078] Tensor scaling provides a stable feature space for the model, hyperparameter tuning ensures the model's effectiveness in dynamic scenarios, and trigger coefficient writing is responsible for persisting the results to the semantic network, achieving a closed-loop update of data, model, and knowledge. This design avoids the redundant steps of manually synchronizing thresholds to the database after model upgrades in traditional processes, improving information consistency and engineering efficiency, and demonstrating an automatic semantic alignment capability rarely seen in the industry.

[0079] Step two, using the coupling trend quantity and credibility index as its dual cores, successfully infuses physical mechanisms and data health into machine learning input, supported by continuous cross-correlation and information theory quality assessment. In the model stage, scale consistency ensures uniformity of feature channel dimensions, and adaptive hyperparameters enable the model to self-adjust according to seasonal and sample distribution fluctuations, making the prediction boundary both sensitive and robust. Finally, the log-probability trigger coefficient and natural language trigger factor are simultaneously written into the knowledge graph, generating version fingerprints to provide authoritative digital evidence for subsequent multi-media consistent early warning and consortium blockchain notarization.

[0080] Step three focuses on maintaining equivalence of the same risk semantics across multiple communication media—a crucial but often overlooked aspect of the slope early warning chain that is most prone to misunderstanding. The first two steps have transformed multi-source monitoring data into instability trigger coefficients, triggering factors, and early warning information entities. Only by converting these structured results into different formats, such as SMS messages, emergency broadcast scripts, and social media posts, can we truly guide on-duty technicians, nearby residents, and temporary tourists to take consistent actions. Meanwhile, each medium has unique limitations in character length, content review, feedback channels, and interaction methods; relying solely on manual template splitting inevitably introduces semantic shifts due to further editing.

[0081] Step 3: Through semantic mapping engine, typesetting and splicing algorithm and summary fingerprint generation mechanism, the same early warning entity is transcribed into a message that is readable, recyclable and verifiable through multiple channels, and the efficiency and credibility of the release link are ensured by priority scheduling and real-time consistency monitoring.

[0082] Step 301: Multi-media early warning template generation and digest code binding Semantic layer and presentation layer bidirectional mapping engine: First, fields such as instability trigger coefficient, trigger factor, risk level, and version fingerprint from the knowledge graph are input into the template parser. The parser queries the template rule base based on three parameters: audience profile, media characteristics, and urgency threshold, and selects the template skeleton that best matches the current scenario.

[0083] Subsequently, the instability trigger coefficient was rewritten into concise phrases such as extremely high risk and medium risk, based on the risk level threshold. The triggering factor was broken down into one or two sentences describing the cause and the terminology was automatically replaced according to the professional density allowed by the channel. For example, professional text messages retained the wording of shear stress reduction, while public channels were rewritten as "once the landslide surface loses resistance." The engine also automatically reduced irrelevant modifiers to ensure that text messages did not exceed the seventy-character limit, broadcast scripts did not exceed the twenty-second reading time, and social media post titles did not exceed thirty Chinese characters.

[0084] After the text is generated, reverse semantic parsing is immediately performed to verify whether the text can recover the original fields; only when the bidirectional recovery error is within one percent will the template enter the typesetting process.

[0085] A high-level warning was triggered late at night on a steep slope beside a certain primary highway. The engine generated a professional text message for the on-duty technician: "Extremely high risk; please arrive at the monitoring room within two minutes to confirm the beam status." A broadcast was generated for residents: "Evacuate south along the town's central road and remain calm." A post with pictures and text was generated for social media platforms: "Significant landslide activity; tourists are advised to suspend entry." Reverse verification confirmed that all three text messages could derive the same trigger coefficient and trigger factor before being published to avoid confusion between multiple versions.

[0086] Adaptive typesetting and multimodal stitching based on media characteristics: After semantic filling, the typesetting and splicing algorithm reads parameters such as the channel's character limit, multimedia limit, screen resolution, and synchronization bandwidth. It then performs verb compression, synonym merging, and punctuation rearrangement on the text, inserting line breaks or pause markers as needed. For social media, the algorithm accesses historical disaster images or real-time slope heatmaps from the knowledge graph. Through resolution resampling, gamma correction, and file compression, it keeps the total image and text size within the platform's 10-megabyte limit. Simultaneously, it adjusts the image color saturation to a broadcast-safe range to prevent reduced recognizability after image-text mixing. For broadcast scripts, the system divides continuous sentences into segments of up to eight seconds according to the frequency response curve of a mono audio device, inserting half-second pauses before and after high-risk verbs to guide the announcer to reread key information.

[0087] Simulation tests in the valley scenic area showed that an unoptimized broadcast script was difficult to distinguish numerical information in a noisy environment. The layout algorithm rewrote trigger coefficients higher than 0.75 as extremely high risk, and split them into two short sentences with extremely high risk and pause, with a one-second silence before and after each sentence, significantly improving the accuracy of speech recognition.

[0088] Digest generation and cross-media uniqueness verification: Each formatted message will have its instability trigger coefficient, channel identifier, generation timestamp, and template version number concatenated into a string in a fixed order, and then a 256-bit digest code will be generated using a secure hash algorithm. The digest code is published simultaneously with the message or embedded in a hyperlink. If anyone deletes, adds, modifies, or reorders the message text or multimedia, the hash result will immediately change completely. When the system retrieves content returned from the channel, it uses the original digest code as the comparison benchmark; if a match fails, the message is considered to have undergone unacceptable editing. In addition to the Chinese environment, the system will uniformly adopt a UTF-8 encoding process before international release to ensure that the description content still points to a unique hash value across languages, avoiding misjudgments caused by language differences.

[0089] During a social media sharing session, a user cut out an evacuation route map. A data crawler detected a discrepancy in the digest code and immediately flagged the share as potentially edited. Meanwhile, the original post retained a valid fingerprint, allowing the source of the edit to be traced and a clarification proactively issued, thus improving information transparency.

[0090] By combining three technologies—mapping engine, typesetting and splicing, and abstract fingerprinting—early warning information achieves, for the first time, a triple guarantee of semantic reversibility, format adaptation, and unique fingerprinting in multi-media dissemination. The mapping engine resolves the contradiction between technical terminology and public language, which is difficult to balance manually; the typesetting and splicing logic compresses complex multimedia content into the most suitable presentation form without the need for secondary manual editing; and the abstract code provides a zero-cost, high-precision automatic comparison benchmark for subsequent consistency verification.

[0091] Step 302: Parallel release scheduling and real-time monitoring of semantic consistency Channel Priority Quantification and Batch Segmentation Strategy: Each delivery route is evaluated using a score based on three metrics: channel coverage, end-to-end latency, and interaction activity. Coverage represents the proportion of potential audiences reached, latency reflects message delivery speed, and interaction activity measures the frequency of receipts and forwards. These three metrics are weighted and fed into a non-linear mapper to generate a priority score between zero and one. The scheduler packages messages into multiple batches according to their scores, from highest to lowest, with each batch controlled within 40% of available bandwidth, reserving a margin for receipts and monitoring streams. If the first batch is sent and the receipt feedback indicates abnormal latency for a high-scoring channel, the scheduler automatically lowers the weight of that channel in subsequent batches and transfers the remaining messages to the next priority carrier.

[0092] During a nighttime downpour, the SMS gateway experienced high online rates while social media platforms experienced network instability, resulting in the initial SMS allocation being the first batch. After sending, a surge in SMS delivery delays was observed. The scheduler immediately reduced the weight of these SMS messages and brought forward the second batch of broadcast messages to ensure rapid delivery of critical information. Semantic consistency embedding alignment and drift alerts: To capture semantic shifts caused by platform compression, user editing, and even broadcasting errors, a semantic embedding alignment model is constructed: first, the original template text is encoded into a vector, and then the real-time crawled online text is encoded into a vector of the same dimension, and the cosine similarity between the two is calculated. When the similarity is lower than a preset threshold and the digest code verification fails, an information deviation node is generated and its source identifier, deviation degree, and crawling time are written into the knowledge graph. The subsequent step four will use graph traversal to quickly locate the deviation propagation chain.

[0093] To attract attention, Weibo users changed their "immediate evacuation" message to "temporary observation." If the similarity score dropped sharply to below 80% and the hash code didn't match, the police were immediately alerted. On-duty personnel issued a clarification within ten minutes and marked the false content with a guiding entry, effectively curbing the spread of rumors.

[0094] Release link latency monitoring and dynamic feedback weighting: Each message is immediately timestamped upon leaving the gateway, and the system listens for the first client-side receipt or platform read event and records the response time. The difference between these two times represents the propagation latency metric. The scheduling module feeds this metric back to the channel latency index in real time and dynamically adjusts the next batch order accordingly. If three consecutive samples show that the latency exceeds the static threshold, the system writes a channel health decline label into the knowledge graph, providing a basis for adjustment to the next layer of priority quantification logic.

[0095] Priority quantization algorithms enable different media to dynamically share bandwidth and attention resources during network congestion or equipment failures, preventing high-value channels from being paralyzed due to local bottlenecks. The embedded alignment model and digest code double-insurance reduce the detection time of oversimplification, malicious editing, or broadcast errors from minutes to seconds, significantly compressing the window for rumor propagation. A latency monitoring closed loop enables self-adjustment of the release strategy based on real-time indicators, ensuring high availability even in complex disaster scenarios. The release pipeline not only accurately delivers the latest warnings to those who need them most but also promptly detects and corrects information deviations, forming a planning-execution-monitoring-adjustment closed loop, greatly improving the robustness and transparency of the digital early warning system.

[0096] The early warning information has evolved from structured entities to cross-channel consumable content, and then to real-time monitorable and schedulable streams, realizing a closed loop across the entire chain from the data layer, presentation layer to the operation and maintenance layer.

[0097] The semantic mapping engine ensures a one-to-one correspondence between technical terms and public language, the typesetting and splicing algorithm ensures optimal presentation on any medium, and the digest code mechanism provides an undeniable fingerprint throughout the entire process. Priority quantification, semantic consistency monitoring, and latency feedback together construct a dynamic, resilient, and self-healing publishing network, making the goal of information reaching the right audience verifiable, measurable, and traceable even in resource-constrained and volatile public opinion scenarios.

[0098] The core significance of step four lies in bringing the distributed multi-media warning messages back to the same reliable semantic coordinate system and using quantitative rules to instantly reveal any content deviations caused by automatic compression, character transcoding, user deletion or modification, or even verbal errors. Unlike traditional post-event sampling and verification, this approach initiates backhaul capture, difference detection, and deviation tracking early in the message's dissemination process, thereby compressing potential public opinion risks to the minute level.

[0099] Step 4: Align the real-time returned content with the original rendered text one-to-one in the knowledge graph, and instantly label the information deviation nodes through difference quantization and graph traversal algorithms, and output traceable propagation paths and impact measurement results.

[0100] Step 401: Data aggregation and difference quantification Semantic alignment multi-threaded timing reconstruction mechanism: First, separate capture threads are created for the three types of sources: SMS gateway, emergency broadcast text stream, and social platform interface. The digest code, channel identifier, and receiver timestamp in the header of each capture packet are written into the thread's dedicated buffer.

[0101] Subsequently, three key anchor points are constructed on the global timeline: the first is the sending time anchor, which is the moment the message leaves the publishing gateway; the second is the first receipt time anchor, which is the moment the first message with the same digest code arrives; and the third is the last interaction time anchor, which is the moment the last comment, like, or share appears under that digest code. The time-series reconstructor uses these three anchor points to stretch or compress scattered receipt segments into a standard time window of the same length, and then uses this window as an index to merge text and metadata. In this way, all the returned fragments of each digest code, regardless of network latency, are mapped to a unified start and end scale, ensuring that subsequent difference measurements only target the content itself and are not affected by time misalignment.

[0102] Taking a mountain disaster scenario as an example, broadcast messages are frequently repeated within twenty seconds, with the same digest code corresponding to multiple text transcriptions; while SMS messages are received with a delay of more than ten seconds due to microwave blind spots in valleys. If the two are simply concatenated according to the order of reception, the broadcast text will be misjudged as a duplicate receipt, and the SMS message will be treated as a delayed receipt, resulting in baseline drift. The time-series reconstruction mechanism uses anchor point alignment to place broadcast and SMS messages in the same unified window frame, thereby accurately capturing the content changes themselves without introducing false differences due to network characteristics.

[0103] Fingerprint-level coding difference stripping process: Immediately after time-series alignment, the digest code of each returned text is verified. If the digest codes match perfectly, the text enters the fingerprint qualification bucket without deep comparison. If the digest codes do not match, the text undergoes a unified character encoding conversion, unifying all possible UTF, GB, or ISO encodings into a comparable universal encoding, and then the digest code is recalculated. If the recalculated digest matches the original archive, the difference is determined to be an encoding layer difference and no deviation alarm is triggered. Only when the recalculation still does not match is the text sent to the content difference quantification engine. This hierarchical process prevents false positives caused by device character set differences or SMS center gateway transcoding, significantly reducing the false alarm rate.

[0104] Many low-end terminals default to saving Chinese SMS messages as UTF-16, while cloud templates use UTF-8. Direct comparison will generate log-level red warnings. The encoding difference stripping process quickly identifies these harmless differences by transcoding and then comparing them, reducing invalid alarms faced by on-duty engineers and allowing them to focus on actual content tampering or deletion.

[0105] Difference quantization algorithm for semantic vector and weighted edit distance fusion When text is identified as potentially altered, the difference quantization engine is activated. The engine first calls a pre-trained semantic model to convert both the original and returned texts into 1000-dimensional contextual semantic vectors, and calculates the cosine similarity between the two vectors. Subsequently, a weighted edit distance calculation is performed at the character sequence layer: the replacement cost of frequently used high-frequency characters is set to a low value, while the replacement cost of key verbs and place names is set to a high value, highlighting modifications that have the greatest impact on risky instructions. The two scores are weighted, summed, and then inverted to obtain the content difference index, which ranges from zero to one; a higher value indicates a greater difference. The weight ratio is set during system initialization based on a 70% weighting for semantic impact and a 30% weighting for character precision, and can be adaptively fine-tuned by subsequent model iterations.

[0106] Assuming the returned text simply changes "Evacuate immediately" to "Leave right away," the semantic vectors are highly similar, and the character replacements are also synonyms, resulting in a difference index of less than 0.1. However, if someone deletes the entire evacuation route and inserts "Temporarily observe," the semantic vector difference amplifies, and the character deletion is significant, causing the difference index to rise rapidly to over 0.7, successfully distinguishing the risk level difference from harmless text rearrangement.

[0107] Multi-level clustering-driven conflict cluster compression strategy: During peak batch data scraping periods, a single digest code may correspond to hundreds or even thousands of returned texts. To reduce computational load and improve the efficiency of manual review, a two-level clustering process is performed after calculating the difference index: First, coarse clustering is performed based on channel identifier and priority score to distinguish between high-reliability and second-reliability sources; then, density clustering is performed within each coarse cluster using the difference index as a distance metric, automatically aggregating the text into several conflict clusters, with the text with the highest difference index selected as the representative sample from each cluster. After clustering, only the representative sample is saved for subsequent graph traversal, while the remaining text is mounted via references, maintaining information integrity while significantly reducing memory and computational consumption.

[0108] A trending Weibo topic may generate hundreds of reposts, most of which are only slightly modified or not modified at all. Clustering compression strategies fold over 90% of repetitive or similar text into the same conflict cluster, allowing difference tracking and manual review to focus on the very few samples that truly deviate from the original text, thereby identifying high-risk variants at minimal cost.

[0109] The traditional manual sampling and comparison process has been upgraded to a multi-threaded, fingerprint-based, hierarchical, semantic-character dual-domain quantization, two-level clustering, and compression four-dimensional integrated automatic link. Temporal reconstruction ensures consistent comparison perspectives; fingerprint-level filtering reduces the false alarm rate from exponential to a controllable level; semantic-character hybrid indicators accurately capture substantial content shifts; and clustering compression maintains sub-second processing speeds even in high-concurrency environments.

[0110] Step 402: Deviation Node Labeling and Propagation Path Backtracking A three-rule-driven bias hierarchy annotation engine: Based on the existing ternary relationship between the carrier and the receiving subject of the early warning information dissemination carrier in the knowledge graph, the carrier node of each differential over-threshold text can be quickly located.

[0111] Subsequently, the rules engine performs three consistency checks on the node: first, digest code consistency, i.e., whether the returned text digest code is consistent with the original digest code; second, timestamp order, i.e., whether the returned time is earlier than the original sending time or later than the last interaction closing time; and third, whether the content discrepancy exceeds a threshold equal to the discrepancy in the previous hop multiplied by a tolerance coefficient. If any rule is not satisfied, the node is identified as an information deviation node, and a deviation level value is assigned based on the product of the content discrepancy and the propagation depth. The higher the level value, the higher the priority of operational intervention.

[0112] In an extreme incident of rampant rumors, a forum post was deemed to have a large content discrepancy due to the deletion and modification of evacuation routes, but it was at the end of the third forwarding chain, indicating a depth of three. Another WeChat public account article only had a text order misalignment, indicating a moderate discrepancy, but it was only one hop away from the source. The former and the latter had similar hierarchical values, so the high-risk node detection algorithm would push both to the on-duty engineer to ensure that both the impact and the discrepancy were considered.

[0113] Minimum cost backtracking path algorithm for multi-weighted composition: After marking the deviation nodes, it is necessary to find the originating node that introduced the difference. To this end, the algorithm assigns three types of weights to each edge in the propagation graph: propagation delay weight, channel reliability weight, and semantic similarity weight. These three weights are linearly normalized and then added together to form the comprehensive edge cost. A minimum cost path search is performed starting from the deviation node. When the comprehensive cost along the path begins to increase and the content difference amount first falls below the warning threshold, the originating node is considered found. This path not only indicates which link introduced the difference but also provides a decomposition of the contribution of each hop, facilitating subsequent rectification.

[0114] A misinformation spread from an anonymous social media platform to Weibo, and was then reposted by several news apps. The algorithm, in its weighting calculation, determined that the anonymous platform had extremely low reliability weight, and coupled with the high degree of content variation, the overall cost increased dramatically, quickly identifying the anonymous platform as the source. Meanwhile, the semantics remained almost unchanged between Weibo and the news apps, resulting in lower overall costs, and the information was considered a misrepresentation and reposted. This allows for precise and targeted governance measures, avoiding a blanket ban on subsequent reposters.

[0115] Audience influence measurement and risk level dynamic adjustment mechanism: After identifying the source node, it is also necessary to assess the actual scope of the deviation information. The impact measurement model comprehensively calculates three factors: the number of users covered, the platform's interaction activity, and the information re-propagation index. The number of users covered comes directly from the platform's feedback statistics on reading or clicking; interaction activity is measured by the density of behaviors such as comments and likes; and the re-propagation index is calculated by combining the average forwarding depth and the width of the forwarding chain. The three scores are summed according to their weights to obtain the impact score. If the impact score crosses the system threshold, the current risk level is immediately raised by half a level, and the scheduler in step three is triggered to generate an urgent correction message to ensure that the negative spread is contained before the turning point of public opinion.

[0116] During the National Day holiday, an erroneous announcement at a scenic spot was recorded by a tourist and uploaded to a short video platform, reaching tens of thousands of users. Due to the high volume of visitors and increased interaction during the holiday, the impact score skyrocketed. The system automatically raised the risk level and pushed out an official clarification video, effectively alleviating the tourists' panic.

[0117] Trust Level Dynamic Write-back and Adaptive Deployment Adjustments: The deviation level value, source node identifier, and audience impact score are written into the knowledge graph entity and relation attributes. For propagation paths containing high-level deviation nodes, the system automatically lowers the priority score of related channels in subsequent alerts; for gold channels with stable and unbiased feedback, the priority score is increased proportionally. In this way, the release scheduling in step three can prioritize the use of high-trust channels in the next round, gradually eliminating problematic links and achieving self-healing optimization.

[0118] A broadcasting station in a mountainous area was repeatedly marked as having a level 1 deviation due to aging equipment. The system lowered its priority for three consecutive rounds, ultimately limiting its use to off-peak hours. Meanwhile, the local government's app provided consistent and consistent feedback, which was prioritized and became part of the first batch of push notifications, significantly improving the alert delivery rate.

[0119] By upgrading information governance from static auditing to a dynamic closed loop, the rule engine ensures that deviation nodes have clear hierarchical labels and are searchable; minimal-cost backtracking quantitatively identifies the primary source, eliminating the need for traditional manual in-depth investigation; impact measurement couples the scale of public opinion with risk level in real time; and trust write-back allows the release strategy to self-correct, forming an immediate feedback mechanism for adjusting strategies upon discovering problems.

[0120] Step four uses a six-step sequence of capturing alignment differences, quantifying hierarchical annotation paths, backtracking impact assessments, and trust write-back to compress any text or speech deviations during cross-media propagation into a quantifiable, localizable, and manageable range. Step 401 ensures that input information undergoes rigorous alignment and clustering before entering deep analysis, significantly reducing false positives and resource consumption. Step 402 projects the deviation detection results back to the knowledge graph and enables the publishing chain to have self-healing capabilities through dynamic priority adjustment.

[0121] Consortium blockchains, with their characteristics of transparent and verifiable processes, immutable historical records, and multi-center node permissions, have become the de facto foundation for disaster information governance, ensuring auditable processes and traceable results. However, simply storing the core text of instability warnings on the blockchain cannot meet the dual demands of project operation and continuous optimization: on the one hand, evaluating whether a warning is truly effective requires examining performance indicators such as its dissemination breadth, initial reach speed, and subsequent correction efficiency; on the other hand, only by writing these performance metrics back to the knowledge graph can a self-correcting closed loop be formed when generating templates or adjusting channel weights in the future.

[0122] Step 5: Solidify the verified key early warning semantics and dissemination performance into the consortium blockchain in one go, and write the confirmed information on the blockchain back into the knowledge graph to drive the continuous evolution of the three business lines of templates, channels and audits.

[0123] Step 501: On-chain fingerprint generation and consortium blockchain writing Dual-domain digest fusion and fingerprint hash construction: To ensure that the semantics of the warning and the effectiveness of governance are presented on-chain simultaneously without revealing too much sensitive information, the warning text is first semantically simplified, retaining three types of elements: First, the numerical description of the instability trigger coefficient exceeding the set threshold and the corresponding risk level phrase in the text; second, the overview of the triggering factors, that is, the dominant physical mechanism that caused the alarm; and third, the evacuation instructions containing action verbs.

[0124] Simultaneously, three metrics—time spent on initial reach, time spent on difference correction, and coverage improvement ratio—are extracted from the dissemination performance matrix. These values ​​are discretized into predefined character ranges and converted into text format. Two summaries are concatenated in a fixed order, and then a template version number and chain write time description are added. Finally, the entire plaintext is used to generate a unique fingerprint using a collision-resistant hash algorithm recommended by the consortium blockchain. The fingerprint is a 384-character hash string, sufficient to prevent birthday attacks or rainbow table reverse engineering.

[0125] In forensic identification scenarios, identification agencies can export broadcast transcripts and performance logs from their operational backend, independently execute the same summary fusion rules and calculate fingerprints, and then verify them against on-chain records. If the fingerprints match, it can be confirmed that the broadcast text and performance indicators have not been tampered with afterward; if the fingerprints do not match, the discrepancies can become the starting point for accountability. This text plus performance fingerprint mechanism significantly improves the completeness of evidence compared to the traditional method of only storing text hashes, eliminating the gray area where the text remains unchanged but performance indicators are falsified.

[0126] Cross-organization node consensus and write transaction locking: The consortium blockchain consists of four types of nodes: supervisory units, emergency management bureaus, insurance institutions, and research institutions. Since any single party holding write rights would undermine credibility, the system employs multi-signature consensus: supervisory nodes initiate transactions and sign them using their own private keys; emergency management bureau nodes double-sign, primarily checking the matching of timestamps and administrative regions; insurance nodes double-sign, focusing on checking the completeness of performance field formats; and research nodes double-sign to confirm sufficient metadata. Once at least 75% of valid signatures are collected, the smart contract enters a write-locked state, prohibiting further modification of the transaction content, before packaging the transaction into a new block. If a node is offline for an extended period, the contract provides an emergency simplified mode, allowing other nodes to quickly complete the write operation using three-party signatures after leaving offline proof in the monitoring log, ensuring timeliness in disaster scenarios.

[0127] During the rainy season, when communication in valleys is unstable, local nodes may go offline. Without an emergency simplification mechanism, the entire chain would be shut down for an extended period, losing its value for evidence storage. By using multi-signature and timeout degradation, the chain maintains a high security threshold while remaining available in extreme situations, thus strengthening the system's resilience.

[0128] On-chain metadata indexing and fast retrieval interface: Evidence solidification is only the first step; efficient querying is equally important. A composite key is generated based on three elements: administrative region code, calendar date, and risk level, serving as an inverted index in the side-view database. The querying party can input any combination to retrieve the block height and fingerprint string. The client then uses the knowledge graph fingerprint-entity mapping to find the complete original text and performance matrix. The interface uses a universal network format for output, compatible with web, mobile, and third-party visualization platforms.

[0129] Dual-domain summarization integrates risk connotation and governance extension into a single fingerprint, greatly improving the granularity of evidence; multi-signature consensus combined with timeout downgrade balances security and timeliness, which is particularly crucial in high-risk slope scenarios; composite indexes transform on-chain data from dormant archives into active resources, which can be accessed at low cost by scientific research, judicial, and insurance entities.

[0130] Step 502: On-chain write-back and performance closed-loop update On-chain receipt monitoring and knowledge graph write-back: After the chain write is completed, the consortium blockchain publishes a transaction confirmation event. Upon capturing the event, the listener writes the block height, confirmation time description, and on-chain fingerprint back to the knowledge graph, specifically into the relationship of the on-chain fingerprint in the warning information entity, and writes the chain write time value in the attribute field. Chain write time becomes an important indicator for evaluating the operational efficiency of the consortium blockchain and is periodically summarized in the system performance entity for analyzing the optimal combination of node count, network bandwidth, and block generation time.

[0131] If the average chain write time increases from three seconds to eight seconds over three months, the performance panel will immediately issue a red warning. The operations and maintenance team can use this information to investigate whether the node load is too high or the network is congested, and then solve the bottleneck by scaling up or chain-level sharding.

[0132] Communication performance indicator mapping and template weight optimization: In the knowledge graph, the performance table for each message template receives two sets of metrics in real time: the initial delivery time and the time for error correction. The delivery time is first normalized, and then a real-time score is calculated. This real-time score is quickly fed back to the template selector, influencing the probability of template extraction for subsequent events. The scoring algorithm is set with a slightly higher weight for delivery time than for error correction time, adhering to the logic of prioritizing notification before correction.

[0133] If a template exhibits a short average reach time but a slightly longer correction time in multiple landslides, it can still achieve a high real-time score due to its rapid reach, making it suitable for use in emergency situations. Conversely, templates with slow reach but fast correction will be assigned to non-emergency scenarios by the system, forming a fine division of labor.

[0134] Channel trust score reassessment and priority rebalancing: The success rate of digest code verification and chain writing latency for each propagation link are periodically calculated, and a channel trust score is synthesized according to a fixed weight. Channels with high scores are given higher priority in subsequent scheduling, while channels with low scores are given lower priority. This dynamic mechanism encourages operators to proactively optimize the links, creating a positive feedback ecosystem where data drives decision-making.

[0135] After the local broadcasting station underwent equipment upgrades, the chain write latency dropped from six seconds to two seconds, the verification success rate rose to 100%, and the trust score increased significantly. It was immediately upgraded to a high-priority channel by the system, validating the value of the upgrade investment.

[0136] Automatic generation of compliance audit reports and external interfaces: Once the write-back is complete, the report generator automatically retrieves the on-chain fingerprint relationships of all early warning information for the current period, along with fields such as delay, success rate, and impact score, using a knowledge graph query script. This generates a machine-readable compliance report, which is then pushed to an external regulatory interface. External systems can directly render this report as a visual dashboard or archive it as an audit PDF.

[0137] Insurance institutions no longer need to manually apply for data; they can directly subscribe to the reporting API to instantly learn about information dissemination and corrective actions during each disaster, thus creating a green channel for the claims process where data is evidence.

[0138] Chain write receipts not only provide proof of completion for legal evidence preservation but also bring quantifiable metrics for self-optimization. Template weights and channel priorities fluctuate in real time with performance indicators, enabling the dissemination system to continuously self-calibrate; performance dashboards and automatic reports translate blockchain-native events into a language that operations and supervision can directly understand, achieving zero-latency decision-making for technical event management.

[0139] By constructing a highly reliable, transparent, and iterative closed-loop pathway at both ends of the consortium blockchain and the knowledge graph, the on-chain portion ensures the long-term immutability of early warning semantics and governance performance, while the knowledge graph portion absorbs on-chain feedback in real time, driving template optimization, channel reassessment, and compliance auditing. In this way, the slope early warning system not only becomes robust in the four-level flow of observation, inference, release, and verification, but also develops self-growth capabilities at the evidence storage, feedback, and evolution levels. This truly achieves a leap from data credibility to governance credibility, and from a single early warning to continuous improvement, laying the foundation for a benchmark paradigm for digital governance of disaster information.

[0140] Please see Figure 2 This invention provides a slope early warning system based on a knowledge graph, comprising: The data acquisition unit, based on lidar, satellite remote sensing and groundwater level monitoring flow, constructs a spatiotemporal network system of observation indicators, geological structures and historical disasters containing source traceability metadata in the knowledge graph; The data processing unit calculates the coupling trend between displacement and pore water pressure and the reliability of the data, uses a gradient boosting model to output the instability triggering coefficient, and writes it along with the triggering factor into the early warning information entity; The generation unit automatically generates SMS, broadcast, and social media templates based on the semantic relationship between the warning information entity, the transmission carrier, and the receiving subject, injects a unique digest code, and publishes them in parallel to maintain semantic consistency across media. The backtracking unit acquires text returned from various channels in real time, categorizing it by digest code, timestamp, and content differences. Figure 1 Consistency rule comparison; if the threshold is exceeded, the node with information deviation is marked and the propagation path is traced back. The feedback unit generates hash fingerprints from the verified core early warning segments and propagation performance, writes them into the consortium blockchain through multi-node consensus, and fully synchronizes the on-chain fingerprints and performance back to the knowledge graph.

[0141] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0142] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0143] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0144] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0145] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A slope early warning method based on a knowledge graph, characterized in that: comprising, Based on laser radar, satellite remote sensing and groundwater level monitoring flow, the knowledge graph is constructed to include observation index-geological structure-historical disaster spatial network system with traceability metadata metadata; The coupling trend of displacement and pore water pressure is calculated, and the gradient boosting model is used to output the instability trigger coefficient, which is written into the early warning information entity together with the trigger factor; According to the semantic relationship between the early warning information entity and the transmission carrier and the receiving subject, short message, broadcast and social media templates are automatically generated, unique abstract codes are injected and cross-media semantic consistency is maintained; Real-time acquisition of each channel return text, comparison according to abstract code, timestamp and content difference graph consistency rule, exceeding threshold marks information deviation node and backtracks propagation path; The checked core early warning segment and propagation performance are generated into a hash fingerprint, which is written into a consortium chain through multi-node consensus, and the on-chain fingerprint and performance are completely synchronized and written back to the knowledge graph.

2. The slope early warning method of claim 1, wherein: The link reliability index is used to dynamically screen the laser, radar, satellite remote sensing and underground water level collection channels. When the link reliability index is less than 0.85, the corresponding data packet is automatically removed, and the removal reason and timestamp are registered in the knowledge graph to ensure the timing alignment accuracy and data traceability integrity.

3. The slope early warning method of claim 2, wherein: The relationship edge weight between the observation index node and the geological structure node is established by weighting the geological confidence, observation quality and historical disaster frequency, and the edge weight is limited to the interval of 0 to 1 through logical normalization mapping, dynamically weakening the weak correlation path and improving the subsequent reasoning accuracy.

4. The slope early warning method of claim 3, wherein: The displacement sequence implements second-order difference accumulation, and the pore water pressure sequence implements multi-scale gradient calculation, then continuous wavelet cross-correlation is performed and energy normalization is performed to generate a coupling trend quantity in the range of 0 to 1; The residual entropy and sensor health mutual information are used to calculate the data reliability index, which is used as the input of the gradient boosting model.

5. The slope early warning method of claim 4, wherein: The coupling trend quantity, real-time reliability index and one-lag reliability index are scaled by quantile distance to form a stereoscopic feature tensor, and the learning rate, maximum tree depth and subsample ratio are jointly searched by adaptive differential evolution algorithm, so that the gradient boosting model has high recall rate and controlled false positive rate in the class imbalance scene.

6. The slope early warning method of claim 5, wherein: The early warning information entity is converted into short message, broadcast and social media templates by using a semantic layer-presentation layer mapping engine, and a unique abstract code is embedded and medium feature adaptive layout is performed; The template can be published only after the semantic restoration error is verified to be less than one percent by reverse analysis.

7. The slope early warning method of claim 6, wherein: The priority score is calculated according to the coverage, real-time index and interaction coefficient for all messages to be sent, and the priority score is batched and pushed in parallel. Each message is recorded after the first reply time and real-time feedback is updated for the next batch of channel sorting, realizing dynamic load balancing under bandwidth limited conditions.

8. The slope warning method according to claim 7, characterized in that: First, the returned text is checked for summary code, and then a content difference index is generated by fusing the semantic vector cosine similarity and the weighted edit distance. If the content difference index is higher than the preset threshold, the corresponding propagation carrier node is marked as an information bias node and written into the bias level attribute; Based on the time weight, channel weight and semantic similarity weight, the edge cost is synthesized, the minimum cost path search is performed from the bias node upwards to determine the difference source node, and the influence score is calculated by combining the covered user number, interaction activity and re-propagation index, which is used to dynamically adjust the risk level.

9. The slope warning method according to claim 8, characterized in that: The trigger coefficient, trigger factor and three propagation performance indicators are textually spliced, and the template version number and write chain time description are appended. A 384-bit on-chain fingerprint is generated by an anti-collision hash algorithm, and then a multi-signature confirmation is completed by the supervision, emergency, insurance and scientific research nodes to write into the alliance chain; After listening to the transaction confirmation event on the chain, the block height, confirmation time description and on-chain fingerprint are written back to the knowledge graph, and the template real-time score is updated according to the first batch of touch time consumption and difference correction time consumption. At the same time, the channel trust score is re-estimated according to the summary code verification success rate and chain writing delay, which is used to adaptively optimize the subsequent template selection and release priority.

10. A knowledge graph-based slope early warning system, characterized in that: Including, The data acquisition unit constructs a spatiotemporal network system containing observation index-geological structure-historical disaster data with traceability metadata based on laser radar, satellite remote sensing and underground water level monitoring flow; The data processing unit calculates the displacement and pore water pressure coupling trend and data reliability, and uses a gradient boosting model to output the instability trigger coefficient, which is written into the warning information entity together with the trigger factor; The generation unit automatically generates SMS, broadcast and social media templates according to the semantic relationship between the warning information entity and the propagation carrier and the receiving subject, injects a unique summary code and releases it in real time to maintain cross-media semantic consistency; The backtracking unit obtains real-time returned text from each channel, compares it according to the summary code, timestamp and content difference consistency rules, and labels information bias nodes and traces the propagation path if the threshold is exceeded; The feedback unit generates a hash fingerprint of the checked core warning segment and propagation performance, writes it into the alliance chain through multi-node consensus, and synchronously writes the on-chain fingerprint and performance back to the knowledge graph.

Images