Urban infrastructure intelligent operation and maintenance method based on digital asset twinborn mapping

Abstract

The invention discloses an urban infrastructure intelligent operation and maintenance method based on digital asset twin mapping, and relates to the technical field of urban infrastructure intelligent operation and maintenance, and the method comprises the steps: an edge gateway carries out the multi-modal collection and semantic packaging under nanosecond synchronization; the correlation screening module carries out time domain-frequency domain cross correlation noise stripping to generate a net signal; the threshold value unit adaptively updates an alarm threshold value in real time by using the statistical manifold and the life cycle weight; the anomaly synthesizer fuses the net signal and the threshold map, pays attention to aggregation and output of a credible score, and calls a micro-diagnosis process to synchronize camera shooting and discharge sensing to verify a fault; the micro-diagnosis result and the operation log are written back to a semantic library through a multi-dimensional index, a routing strategy pushes a refined alarm to a hierarchical operation and maintenance main body based on an entropy weight, and a threshold value and an operation and maintenance strategy are driven to be synchronously evolved. According to the method, high-precision synchronization, low false alarm and self-evolution closed-loop operation and maintenance are realized, asset safety and maintenance efficiency are improved, and shutdown risk and cost are reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for urban infrastructure, specifically to an intelligent operation and maintenance method for urban infrastructure based on digital asset twin mapping. Background Technology

[0002] In the supporting construction of smart cities, multimodal sensing nodes are being deployed on a large scale in key infrastructure such as rail transit sections, power ring network switching stations, underground integrated pipe corridors, and pumping stations. These nodes are used to collect high-frequency data in real time, including three-phase voltage-current, vibration and shock, ambient temperature and humidity, and local electromagnetic discharge. Edge gateways typically handle preliminary filtering, event triggering, and uplink forwarding tasks, while the cloud utilizes digital twin models for operational status simulation and maintenance scheduling. To improve the accuracy of operational and maintenance decisions, the industry has successively introduced IEEE 1588 PTP high-precision time synchronization, distributed stream processing engines, knowledge graph semantic alignment, and deep learning anomaly detection models. However, due to limitations imposed by complex metropolitan area network links, diverse and heterogeneous equipment vendors, and hierarchical management by operations and maintenance teams, existing systems continue to evolve in a fragmented manner in areas such as data timeliness, threshold adaptation, fault location, and knowledge accumulation. Some solutions focus on centralized cloud analysis, with sampling cycles forcibly lengthened by network bandwidth and peak congestion, only capable of processing sparse offline data. Other solutions employ fixed-threshold alarms at the edge, with threshold updates relying on manual periodic correction, making it difficult to maintain robustness during seasonal shifts in operating conditions and the cumulative effects of equipment aging. Furthermore, the lack of a unified mapping between the ontology models and time bases of equipment from different vendors leads to a persistent alignment error between digital twin models and field data, hindering the reuse and reasoning of knowledge bases. Operations and maintenance personnel often need to manually compare historical records across systems and forms to complete traceability.

[0003] Typical problems occur in integrated utility tunnel scenarios with low nighttime illumination or high humidity during the rainy season:

[0004] When cable insulation aging causes short-term partial discharge, the three-phase current, vibration vector, and temperature and humidity signals will simultaneously jump within milliseconds. However, due to clock drift and link jitter between different gateways, the current and vibration phases often show microsecond-level misalignment. Fixed threshold algorithms misjudge seasonal drift as abnormal, thus drowning out high-frequency signals truly caused by insulation defects. At this time, high-priority alarms need to be issued promptly via the monitoring center-field team link, but the uplink of video streams and batch logs is prone to link congestion during night shifts, causing repair commands to fail to arrive within a minute window, creating a risk of downtime. Furthermore, fault handling feedback cannot be written back to the threshold model and strategy library in a timely manner due to the lack of a unified semantic index. The threshold remains too tight or too loose for a long time, and false alarms and missed alarms accumulate iteratively, causing maintenance personnel to lose trust in the automatic alarm system.

[0005] In summary, existing technologies struggle to provide low-latency, high-confidence, traceable, and continuously self-optimizing operation and maintenance support capabilities in urban infrastructure environments where complex loads and aging equipment coexist. Summary of the Invention

[0006] (a) Technical problems to be solved

[0007] To address the shortcomings of existing technologies, this invention provides an intelligent operation and maintenance method for urban infrastructure based on digital asset twin mapping. This method involves: an edge gateway performing multimodal data acquisition and semantic packetization at nanosecond-level synchronization; a correlation filtering module performing time-domain-frequency domain cross-correlation to remove noise and generate a net signal; a threshold unit using statistical manifolds and lifecycle weights to adaptively update alarm thresholds in real time; an anomaly synthesizer fusing the net signal and threshold graph attention aggregation to output a reliable score, and invoking a micro-diagnostic process to simultaneously verify faults using cameras and discharge sensors; the micro-diagnostic results, along with operational logs, being written back to a semantic database via multi-dimensional indexing; and a routing strategy based on entropy weights pushing refined alarms to hierarchical operation and maintenance entities, driving the synchronous evolution of thresholds and operation and maintenance strategies. This method achieves high-precision synchronization, low false alarms, and self-evolving closed-loop operation and maintenance, improving asset security and maintenance efficiency; thus solving the technical problems described in the background section.

[0008] (II) Technical Solution

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A smart operation and maintenance method for urban infrastructure based on digital asset twin mapping includes parallel collection of electrical parameters, vibration, temperature and humidity and electromagnetic parameters, and after processing, generating data packets with consistent fingerprints and version vectors and mapping them into a semantic library.

[0011] At the edge side, supercomplex time alignment, complex cepstral equalization, and fractional wave packet coherent noise removal are performed on the data packets to output the net signal and time label.

[0012] The alarm threshold set is adaptively calculated using Wasserstein statistical manifold combined with lifecycle risk coefficient, and a symptom monitoring summary is generated.

[0013] An anomaly synthesizer constructs a five-point evidence graph, obtains a confidence score through a gated residual graph attention network, generates a source index, outputs multi-level diagnostic results, and writes them back to the semantic database.

[0014] Knowledge blocks are encapsulated using hierarchical segmented hashing and pushed to the hierarchical operation and maintenance entity based on entropy weight-minimum entropy matching. At the same time, the policy gain is mapped to the risk coefficient and the threshold is updated synchronously.

[0015] Furthermore, nanosecond-level synchronous triggering is achieved through time base assimilation, enabling real-time synchronous acquisition of electrical parameters, temperature and humidity, vibration, and electromagnetic parameters.

[0016] During the acquisition process, electrical parameters are vectorized and extracted and zero-order compensation vectors are injected. Fractional filtering is used to suppress trend noise for environmental signals.

[0017] Furthermore, the collected information is embedded with lifecycle fragment codes and a unified semantic identifier is generated and encapsulated into a data packet;

[0018] Semantic mapping is performed on data packets, consistency fingerprints are calculated and recursive version vectors are generated, data packets are transmitted to the semantic library through elastic streaming segmentation, and the evidence chain is extended based on ontology to maintain semantic consistency.

[0019] Furthermore, after performing supercomplex time alignment on the multimodal sequences, pseudo-periodic resonances are suppressed by retuning the complex cepstral envelope. Then, coherent islands are extracted by cross-scale scanning of fractional Gabor-Meyer wavelet packets. A semantic pruning graph is constructed based on lifecycle risk weights, maintenance priority weights, and quality label vectors to output a candidate signal set.

[0020] Furthermore, after being decomposed by Hankel-Toeplitz sparse low-rank double decomposition, the candidate signal enters the temporal attention fusion unit to generate attention weights associated with the risk coefficient.

[0021] Attention weights are bound to the recursive version vector to form time tags, and the net output signal is encapsulated according to semantic pointer rules.

[0022] Furthermore, statistical manifold trajectories are generated using the second-order Wasserstein distance minimization criterion, and risk layers are then divided using Laplace-Beltrami projection based on the risk-lifetime tensor.

[0023] Each risk layer uses a g-and-h extrapolation kernel to obtain the tail partition mapping, and introduces the Fréchet-Hoeffding upper bound to achieve multimodal tail collaborative constraints.

[0024] Furthermore, the alarm threshold set is automatically migrated along the manifold gradient flow, and recentering is performed when the Kullback-Leibler divergence is detected to exceed the drift threshold;

[0025] Simultaneously, the life cycle risk coefficient is read and the threshold width is adjusted through an elastic shrinkage function. The corrected threshold and the recursive version vector are synchronously written into the semantic library and an additional symptom monitoring summary is attached.

[0026] Furthermore, the anomaly synthesizer constructs a five-part evidence graph using asset nodes, attribute nodes, version nodes, lifetime nodes, and threshold nodes.

[0027] After compensating for cross-modal time difference through minimum phase multi-frequency rotation, a gated residual graph attention network is used to output a reliable score, and the edge information that contributes the most to the score is written into the source index.

[0028] Furthermore, based on the source index, priority is given to allocating camera and discharge diagnosis windows, and video optical flow vectors and discharge pulse differential hashes are acquired simultaneously, and mutual information-cross-spectral coupling degree is calculated;

[0029] The diagnostic results are divided into three levels: confirmed, suspected, and excluded. The results, along with the fingerprint information, are written into the semantic database to trigger a sliding window update of the risk coefficient.

[0030] Furthermore, the runtime logs are recursively extracted from topics and bidirectionally filled in with actions to generate semantic fragments, which are then aggregated into knowledge atoms using Jensen-Shannon divergence.

[0031] If the knowledge graph coverage is lower than a specified threshold, isolated atoms are connected to the nearest structure center through substructure co-occurrence measurement, and finally encapsulated into knowledge blocks with high-dimensional index matrix, version vector and sliding chain code.

[0032] Furthermore, the alarm routing uses the information entropy weights of danger level, resource occupancy and response time limit as weights, and solves the minimum entropy matching under capacity and delay constraints to directly send high-weight alarms to the corresponding operation and maintenance level;

[0033] The feedback process updates the operation and maintenance strategy vector using an empirical-Bayesian method, and maps the strategy gain to a risk correction amount, which is then synchronized to the threshold calculation.

[0034] (III) Beneficial Effects

[0035] This invention provides an intelligent operation and maintenance method for urban infrastructure based on digital asset twin mapping, which has the following beneficial effects:

[0036] A unified time benchmark and semantic anchors connect all layers of knowledge accumulation from edge perception, ensuring one-click traceability and multiple uses of a single source. This is achieved through consistent fingerprinting. High-bit signature and version vector The dual verification mechanism ensures that any data fragment remains intact and immutable when migrating across network segments, fundamentally improving data credibility.

[0037] Secondly, net signal Lifetime correction threshold The coupling causes the threshold to dynamically change with the asset risk coefficient. Real-time tightening or loosening of thresholds enables differentiated protection for aging and newly commissioned equipment, avoiding false alarms and missed alarms caused by traditional static thresholds.

[0038] Anomaly synthesizer with credibility scoring Quantify the consistency of cross-modal bounds and use a source-tracing index. By locating key coupling edges, on-site micro-diagnostic resources can be precisely focused on high-value fault links, significantly reducing diagnostic latency. Multi-source joint verification integrates video frame optical flow with discharge pulse differential hashing, and determines the authenticity of faults through coupling degree, reducing the risk of single-mode misjudgment and demonstrating the creative synergy of visual-electrical dual-domain complementarity.

[0039] The knowledge rewriting stage utilizes a high-dimensional index matrix. With knowledge blocks Diagnostic experience is solidified into reasonable atoms, and the blind spot compensation algorithm automatically attaches isolated semantics based on co-occurrence probability, continuously improving the knowledge graph coverage and achieving continuous gain in diagnosis as knowledge. The routing strategy layer uses entropy-weighted trajectories to adaptively assess risk, resource occupancy, and response time, solving for minimum entropy matching under network latency and capacity constraints, and directly delivering critical alarms to the monitoring center, regional dispatch, or field teams, reducing information flooding and ensuring low congestion of emergency repair links.

[0040] Incremental policy evolution uses an empirical-Bayesian framework to absorb disposition feedback and output a policy gain vector. After risk adjustment Real-time synchronization back to the threshold unit forms a self-evolving closed loop of policy-threshold bidirectional coupling. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the intelligent operation and maintenance method for urban infrastructure of the present invention. Detailed Implementation

[0042] 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.

[0043] Please see Figure 1 This invention provides a smart operation and maintenance method for urban infrastructure based on digital asset twin mapping, including:

[0044] With the exponential growth in the density and resolution of urban infrastructure sensor networks, relying solely on centralized cloud aggregation and offline analysis is no longer sufficient to promptly capture transient failures and coupling degradation. The concept of digital asset twin mapping couples physical assets, data assets, and knowledge assets in a homogeneous semantic space in real time, injecting a dynamic gene of falsifable, predictable, and collaborative capabilities into intelligent operation and maintenance.

[0045] Step 1: Construct an edge-level perception-semantic-temporal three-element synchronization mechanism for urban infrastructure, generating highly consistent, highly reliable and traceable data packets within millisecond cycles to provide real-time data feeding for digital twins.

[0046] Step one includes the following:

[0047] Step 101: Real-time multimodal acquisition and temporal assimilation

[0048] Urban power distribution cabinets, rail sections, and drainage pumping stations are often distributed across different areas and operate under varying conditions. Using a unified time reference for collaborative sampling of electrical parameter vectors, temperature and humidity scalars, vibration vectors, and electromagnetic tensors is a prerequisite for any subsequent correlation analysis. Through end-side crystal oscillator frequency stabilization and pulse synchronization, analog-to-digital co-processing, multi-physical channel alignment, and time-series tag embedding, the first batch of raw packets was ultimately formed. .

[0049] To suppress jitter caused by long-distance fiber optic distributed networks, the edge gateway embeds a local frequency stabilization module based on an improved constant-parameter Kalman estimator. This module compares the acquisition control word with the messages sent by the IEEE 1588 PTP master clock in real time. Its tuning formula is:

[0050]

[0051] Drift gain, value range This is used to accelerate the recovery of slow frequency conversion differences;

[0052] Noise suppression gain, value range Used to filter out instantaneous jitter between master and slave clocks;

[0053] Master clock timestamp, UTC base; Read from the clock timestamp via a hardware counter; : Measure high-frequency jitter from the instantaneous derivative of the clock; Acquire control word, synchronous pulse trigger gating;

[0054] By employing Kalman-PTP dual correction, the maximum alignment error of the trigger sequences for the four types of sensors is reduced to less than 50 ns, thus preserving phase consistency for subsequent frequency domain correlation analysis.

[0055] For three-phase four-wire distribution nodes, a bidirectional Clarke-Park transform is used to convert the asymmetric voltage-current sampling sequence. Mapped to synchronous rotating coordinate system electrical parameter vector Its composite mapping is written as:

[0056]

[0057] Clarke transformation matrix with fixed coefficients; Park is a rotation matrix whose phase angle is calculated from the synchronous angular velocity. : Electrical parameter vector, which is a four-element vector; Three-phase instantaneous voltage and current;

[0058] Specifically, coordinate transformation is used to eliminate the fundamental wave rotation component, so that subsequent spectral decoupling only focuses on amplitude and phase deviation.

[0059] Regarding temperature and humidity With vibration amount To address the significant impact of seasonal drift in the low-frequency band, a fractional-order frequency domain window is employed. Adaptive filter:

[0060]

[0061] Fractional index, adaptively adjusted according to the diurnal cycle, with a value of 0.45–0.75; Fractional Fourier spectrum, used to extract trend-perturbation mixed components; For frequency variables; The imaginary unit; Original temperature and humidity scalar;

[0062] By introducing a non-integer exponential phase kernel, both trend drift and high-frequency sporadic occurrences are taken into account, enabling subsequent threshold learning to capture genuine anomalies rather than seasonal baseline changes.

[0063] The four types of multimodal quantities collected were subjected to mutual exclusion hashing. Generate 128-bit fingerprint Then the ontology mapping function is called. Map it to a unified semantic tuple:

[0064]

[0065] : Collect fingerprints; : A collection of scene contexts, including facility type and spatial location; Ontology mapping functions are maintained in the edge ontology cache;

[0066] The unified semantic tuple has a unique anchor point in the knowledge graph and can be directly referenced in any subsequent subprocess without secondary parsing. Step 101 uses a four-ring linkage of clock assimilation, vector extraction, fractional filtering, and semantic packetization to solidify the multimodal observations of the scene into the first batch of traceable, high-precision time-stamped data packets. This lays the foundation for semantic hash consistency and streaming segmentation in step 102.

[0067] In use, board-level jitter compensation extends "nanosecond-level synchronization" to "full-path latency consistency," providing a consistent triggering benchmark for sensor nodes across temperature zones and material hardware platforms; interactive zero-sequence modeling ensures that the electrical parameters of the traction motor remain highly recognizable in high-harmonic scenarios; the multi-source radiative heat feedforward mechanism enables the filter to remain resilient to short-term lighting heat flow, no longer mistaking lamp thermal shock for environmental anomalies; backplane delay mapping incorporates PCB-level thermomechanical effects into the clock synchronization design, breaking through the limitations of existing technologies that only compensate at the crystal oscillator and network layers; dynamic Furthermore, it achieves a lightweight implementation of on-chain lifespan information, enabling urban assets to maintain a single lifespan view in multi-entity collaboration, and ensuring that thresholds and operation and maintenance strategies iterate together.

[0068] Step 102: Unified Semantic Mapping and Streaming Segmentation Encapsulation

[0069] In the first batch of data packets After generation, the edge gateway needs to continuously map the frequently arriving data into the unified semantic library. And based on network bandwidth and granularity of operation and maintenance level, it performs flexible segmentation to prevent congestion and packet loss during peak periods;

[0070] To address the issues of long lifecycles and frequent upgrades of urban infrastructure, a dynamic ontology extension operator is introduced. For the unified semantic library Online additions and extensions can be implemented to avoid offline manual maintenance of the core asset, thereby improving the efficiency of rapid management of heterogeneous assets.

[0071]

[0072] Ontology augmentation operator, based on class differences Dynamically add attributes; For semantic library; Similarity threshold between new assets and existing classes, ranging from 0 to 1;

[0073] To ensure consistency in cross-gateway data merging, the first batch of data packets... Hierarchical hashing of inner fields Then use threshold signature Binding timeline, consistent fingerprint When multiple gateways merge traffic, conflict detection can be performed directly to prevent duplicate names or tampering:

[0074]

[0075] In the formula: Consistent fingerprint; Hierarchical hash functions are used to hash tuples in layers. : Timestamp signature, issued by a trusted time source. This is the first batch of data packets;

[0076] Based on the instantaneous throughput of the link subscription granularity at the operation and maintenance level For the first batch of data packets A predictable segmentation strategy is adopted to make packet segment length adapt to network and service priorities, ensuring that the operations and maintenance team continuously obtains critical segments.

[0077]

[0078] in, Segment length Mother package size (number of units); Current bandwidth (bit / s); Subscription level weight, ranging from 0.1 to 1;

[0079] Subsequently, using recursive version control vectors :

[0080]

[0081] In the formula: : No. Segment version vector; XOR mixing operation; Incremental hashing, hashing each segment of the differential field;

[0082] Among them, the XOR-incremental hash combination can achieve differential backtracking and state reconstruction without adding an extra index table.

[0083] If link quality indicators Falling below the threshold This triggers a self-healing retransmission control law, initiating retransmission at the packet level rather than the frame level, reducing latency and maintaining smooth queuing.

[0084]

[0085] In the formula: : The retransmission rate that needs to be compensated; : Retransmission gain factor, ranging from 0.2 to 0.5; : Positive cutoff function, negative values ​​are set to zero; Real-time link quality estimation;

[0086] Through the two-level deep processing in step one, the edge gateway has transformed the original multimodal observations into time-series data packets with unified semantic anchors, consistent fingerprints, and recursive versions. Each field within the system embeds a nanosecond-level timestamp, an asset-attribute bidirectional index, and a verifiable hash value, bridging the semantic gap between the sensor side, the network side, and the knowledge side. Therefore, when step two proceeds to time-domain-frequency-domain cross-correlation analysis, no additional time-series alignment and entity mapping are required; it can be directly based on time-series data packets. The hierarchical structure calls the net signal extraction algorithm; meanwhile, the threshold unit and the anomaly synthesizer can also utilize version vectors. Enables online threshold migration and abnormal trajectory tracing.

[0087] By leveraging trusted crowdsourcing, ontology visual verification, and hierarchical salting, the semantic mapping accuracy and information security of urban assets are improved simultaneously: the operations and maintenance team can confidently accept new asset data from outsourced teams without worrying about ontology contamination; reversible indexes maintain NVMe caching efficiency, and segmented security blocking makes it difficult for potential attacks to reconstruct the original data through hash collisions; round-robin scheduling and sliding window re-insertion ensure that key fields arrive first, while secondary fields are delayed but not lost, ensuring that backend diagnostics and version reconciliation maintain a consistent and continuous causal chain, truly achieving disaster resilience that integrates "network-semantics".

[0088] By introducing the OCR visual step, semantic anchoring obtains a third verifiable source of information, overcoming the weakness of traditional pure data similarity models being easily manipulated by scripts; hierarchical salting introduces the idea of ​​"dynamic key domain division", which subdivides the security granularity to the asset level, rather than simply using high-strength encryption to cause a full-link computational load; the combination of round-robin scheduling and sliding window re-insertion gets rid of the dilemma of "only retransmitting, not scheduling" or "only predicting, not caching".

[0089] The urban infrastructure sensor network has formed a consistent fingerprint at the edge gateway through step one. With recursive version vector Time-series data packets .

[0090] However, time-series data packets It still contains multiple sources of noise, such as equipment background ringing, environmental periodic drift, and communication glitches. If these noises are not distinguished and enter the threshold unit, they will cause statistical distribution to be skewed and amplify the cumulative error of neural alarm.

[0091] Step 2: Construct a three-dimensional cross-correlation link across time, frequency, and semantics to automatically output a structurally clean multimodal net signal. and time tags .

[0092] Step two includes the following:

[0093] Step 201: Multi-domain cross-correlation deep screening

[0094] The high sampling rate of sensor networks results in a large amount of redundant information in a single instant, which would lead to a waste of computing power and bandwidth if processed directly in the cloud; therefore, it is necessary to process time-series data packets at the edge. Perform multi-domain cross-correlation analysis to quantify the authenticity of physical coupling using correlation strength, thereby eliminating irrelevant noise and retaining the synergistic characteristics that truly reflect the asset status.

[0095] The electrical parameter vector is divided into four elements using analytic envelope. Vibration amount and electromagnetic tensor Embedded unified hypercomplex timing Then, with temperature and humidity Composition of observation tensor To synchronously eliminate cross-modal phase drift, a dynamic elastic registration function is introduced:

[0096]

[0097] In the formula: Phase-aligned hypercomplex sequences; Dynamic registration operator, element range ,according to Internal timestamp initialization; The time derivative of the sequence; It is the original hypercomplex sequence;

[0098] By performing minimum energy registration on the time derivative, different modes can obtain a common starting point at the microsecond level, thus establishing an in-phase reference for subsequent frequency domain processing.

[0099] To address the pseudo-period caused by the 150Hz superposition wave on the power distribution side and the cogging resonance of the motor, a complex cepstrum correction is introduced:

[0100]

[0101] In the formula: : Complex cepstral coefficient sequence; Fast Fourier Transform operator;

[0102] Complex cepstrum transforms convolutional resonances into additive peaks, causing high-amplitude pseudo-harmonics to appear as isolated spikes in the cepstrum domain, which facilitates gating filtering and achieves phase balance.

[0103] Furthermore, to capture coupled events at different time scales, a fractional-order Gabor-Meyer master packet is defined. Calculate the cross-scale coherence coefficient:

[0104]

[0105] In the formula: : Modal at scale Time shift The coherence coefficient under the given conditions, with a range of values. ;

[0106] Fractional Gabor-Meyer wavelet packet, order exponent shape factor ; and Corresponding to the first Time series of modal parameters (e.g., electrical parameters and vibrational quantities) after complex cepstral processing; Simultaneously measuring amplitude and phase coupling, the highly coherent islands left after cross-scale scanning are the multimodal resonance cores of potential anomalies.

[0107] Combine the coherence coefficient matrix with the lifetime fragment code Establish a mapping: When the asset lifetime segment drives the threshold region to tighten, retain the frequency-time period with low threshold and high coherence; otherwise, appropriately relax it to form a semantically plastic pruning graph. Semantic plasticity pruning diagram The final decision is made on which part of the signal enters step 202. This step integrates static wave packet coherence with dynamic information about the asset lifecycle, avoiding reliance on mathematical thresholds.

[0108] In step 201, using supercomplex time series as the basis, complex cepstral suppression and fractional wave packet coherence locking are performed, and the final output is a semantically plastic pruning graph. Candidate signal set to be screened This lays the foundation for high-coherence input for the low-rank purification and time-tag injection in step 202.

[0109] In use, the dynamic registration chain consisting of supercomplex time alignment and transient residual gating allows multimodal events to converge quickly even after experiencing millisecond-level mismatches; envelope retuning solves the energy loss that may be caused by complex cepstral filtering, ensuring that subsequent wave packet analysis has a true energy spectrum; the overlap-cancellation strategy not only reduces the amount of computation, but also enhances the interpretability of the results through ghost coordinates; the dual-mapping design of semantic pruning incorporates lifecycle and operation and maintenance priority into the screening decision for the first time, avoiding the embarrassing scenario of traditional algorithms mistakenly cutting data of "low-lifetime high-priority assets".

[0110] Step 202: Net Signal Construction and Time Tag Injection

[0111] Candidate signal set Although noise has been significantly reduced, it still contains dimensional redundancy and weakly correlated fragments, requiring further compression to a low-dimensional dominant subspace that characterizes the asset's operational mechanism, and further compression using time-stamping. Lock the evolution chain to ensure that the threshold unit only faces the necessary and sufficient statistical samples.

[0112] For candidate signal set Constructing the Hankel-Toeplitz array The minimum principal curvature decomposition is performed, and the double decomposition explicitly separates the periodic operating mode from the abrupt anomaly, assigning different weights to subsequent attention mechanisms, wherein:

[0113]

[0114] In the formula: The embedding matrix is ​​formed by the candidate signal set. The Hankel-Toeplitz matrix generated according to the delayed embedding rule has row and column sizes that depend on the window length and the embedding order.

[0115] A low-rank matrix representing the periodic or quasi-periodic skeleton components of a multimodal signal; with the lowest rank... Approximately captures the coupling between background trends and steady-state modes; : A sparse anomaly matrix containing transient shocks, fault pulses, and outlier spikes; its upper limit for the number of non-zero elements is . This reflects its rarity and locality.

[0116] : The upper limit of the preset rank is estimated from the number of relevant islands; The allowed number of sparse elements originates from... size

[0117] low-rank matrix With sparse anomaly matrix Input gated recursive unit, calculate timing attention weight :

[0118]

[0119] In the formula: Time step The temporal attention weight, ranging from (0,1); Learnable weight vectors with the same size as the concatenated vectors; attention weights automatically enhance salient segments and weaken stable backgrounds, forming a weight sequence. Drives subsequent net signal fusion.

[0120] For low-rank background row vectors, periodic or steady-state components in time The corresponding four-dimensional expansion; For sparse outlier row vectors: instantaneous shocks or outlier spikes in time Four-dimensional unfolding; Index of time steps within the window;

[0121] Based on the recursive version vector Binding timing attention at the segment level Generate time tags Ensure that threshold units can be directly derived from time stamp sets. Once the segmented version is retrieved, the threshold synchronization migration along the version axis is completed:

[0122]

[0123] A collection of time tags, recording time steps, weights, and version segment numbers; For sampling timestamps, Version number; The time set within the segment;

[0124] Finally, the order of attention will be determined according to time sequence. The weighted fusion sequence is defined as the net signal. ;

[0125] And inherit the time-series data packets according to semantic identification rules. Asset-attribute anchors in the output .

[0126] Step 202 outputs a net signal through double decomposition, timing attention, version binding, and packetization. and time tags and will and The original embedding provides a clean and traceable dual guarantee for the next step of dynamic threshold evolution.

[0127] Step two uses a cross-modal, cross-scale, and cross-lifecycle information fusion chain to process the time-series data packets output from step one. High-precision noise removal and clean signal generation are completed. Step 201 extends from hypercomplex time alignment to semantic pruning to ensure that the candidate signals entering step 202 have physical coupling authenticity. Step 202 then uses sparse low-rank decomposition and temporal attention weights to accurately extract anomaly-sensitive features, and uses recursive version vectors. Inject time tags The final output is a net signal. Therefore, in step three, the net signal can be directly used as a basis. The statistical distribution is used for adaptive threshold migration without the need for time-consuming alignment or secondary cleaning; at the same time, the anomaly synthesizer and micro-diagnostic process can utilize time stamps. With consistent fingerprints The event chain is reconstructed to ensure the integrity of the diagnostic causal loop. The entire processing logic achieves a seamless transition from coarse-grained noise filtering to refined anomaly characterization, providing a solid and unified net signal baseline for subsequent intelligent operation and maintenance loops.

[0128] In use, the curvature-decreasing hot start allows the sparse low-rank bi-decomposition to converge instantly even under highly dynamic asset conditions, avoiding real-time losses caused by frequent iterations; the conditional gating aggregator incorporates risk coefficients into the attention mechanism, achieving "risk-driven feature amplification," enabling threshold statistics to focus on truly high-risk segments; partial-order consistency auditing resolves potential misalignments between version chains and time chains, unifying threshold updates with event tracing benchmarks; the semantic pointer table and ghost coordinates jointly construct a complete retracement path, providing "one-hop location + cancel tracing" double insurance for subsequent micro-diagnostic processes. Therefore, the overall net signal simultaneously possesses the triple advantages of structural purity, prominent risk, and traceable retracement.

[0129] Urban infrastructure exhibits a complex failure chain throughout its entire lifecycle, encompassing charge migration, thermal stress creep, material fatigue, and environmental erosion. No single-dimensional threshold is sufficient to capture this dynamic evolution. Step two has already output a net signal. and time tags And retain consistent fingerprints Lifecycle fragment code However, how to continuously generate adaptive alarm threshold sets for different assets, different aging stages, and different operating conditions without relying on manual intervention remains a key challenge for intelligent operation and maintenance closed loop.

[0130] Step 3: Based on the real-time statistical distribution and lifecycle weight of the net signal, construct a dynamic alarm threshold set that spans time periods and asset semantics, so that the alarm judgment process can be freed from manual dependence and maintain high sensitivity and low false alarm rate.

[0131] Step three includes the following:

[0132] Step 301: Statistical manifold modeling and risk stratification

[0133] Net Signal It includes multi-dimensional features such as electrical parameter amplitude, phase perturbation, and environmental gradient. Its statistical morphology changes synchronously with operating conditions and aging. If a fixed parameter distribution is still used, mismatch will occur within a short period of time and induce misjudgment.

[0134] First, regarding time tags The defined continuous window construction condition density function Then, using the Wasserstein distance as a metric, the kernel-smooth manifold is obtained in the statistics space:

[0135]

[0136] In the formula: : Statistical manifold, representing the minimum Wasserstein locus that evolves over time; Net signal conditional density function; Second-order Wasserstein distance;

[0137] Conditional density function of net signal Net signal conditional density function Used to characterize a given time section Below, the multimodal net signal vector net signal The probability distribution shape. Combining the sparse-low-rank decomposition in step 202 and the statistical manifold modeling in step 301, a time-recursive Gaussian mixture model is chosen to give the explicit form:

[0138]

[0139] Where: weight of the mixed component :satisfy Based on the life cycle risk coefficient With the right to attention to time Joint weighting—the weight of outlier components is automatically increased during high-risk periods;

[0140] Mean vector A four-dimensional column vector, whose elements represent the components of each mode. Instantaneous central value under low-rank background; Adaptive smooth update;

[0141] covariance matrix : A symmetric positive definite matrix that characterizes intermodal coupling and noise levels; its diagonal follows the mass marker vector. Dynamically widen or narrow;

[0142] Gaussian kernel :

[0143]

[0144] Using a hybrid model can take into account both steady-state components (low-rank background components) and sudden anomalous components (sparse anomalous components). The structure remains consistent; weights With attention Binding enables statistical manifolds More sensitive to high-risk windows;

[0145] Based on the minimum Wasserstein manifold, a smooth evolution curve can be obtained without losing higher-order information of the signal, providing continuous geometric constraints for projection and threshold shift.

[0146] Utilizing lifecycle fragment codes With operation and maintenance priority weight (Values ​​1–5) Construct the risk tensor Then, on the kernel smooth manifold... Conditional density function of net signal Performing Laplace-Beltrami projection:

[0147]

[0148] In the formula: Risk stratification density function; Laplace-Beltrami inverse operator on manifolds

[0149] When used, curvature-weight coupled projection can be used to make different risk layer densities present differences in local curvature, thus accurately partitioning the tail model.

[0150] For the high-skewness segments of each risk layer, a g-and-h extrapolation model is used to capture heavy-tailed behavior:

[0151]

[0152] : Quantum tail mapping result, corresponding quantile ; Over time Changes occur in sync with the risk layer;

[0153] Position parameter, controls the starting level of the tail mapping, usually taken as the time moving average of the net signal conditional density; range is real number; : Scale parameter, with a value greater than 0, controls the starting level of the tail mapping, usually taken as the time moving average of the net signal conditional density; the range is real numbers; Skewness control parameter, the degree of asymmetry between the left and right tails of the section distribution;

[0154] Kurtosis control parameter, which controls the thickness of the tail section; Standard normal quantiles The results, among which It is the inverse normal distribution function; The risk stratification density is adaptively selected.

[0155] g-and-h separates skewness and kurtosis using an exponential-squared composite kernel, making the model both sensitive to tail thickness variations and numerically stable, which facilitates subsequent co-constraints.

[0156] In multimodal scenarios, tail models are prone to imbalance due to outlier dimensions. Here, the Fréchet-Hoeffding upper bound is introduced to establish a collaborative constraint: all tail partitions... The theoretical limit of the cumulative probability of the risk layer must not be exceeded. In conjunction with the iterative projection algorithm, each update back-projects onto the kernel smooth manifold. This ensures statistical consistency. Collaborative constraints provide a theoretical safety net for the multimodal tail, guaranteeing that threshold shifts remain globally consistent even in high-dimensional scenarios.

[0157] Step 301 produces a ternary set. This forms a density-tail combination with risk stratification, providing a benchmark for the dynamic threshold migration in step 302.

[0158] When used, incremental kernel bandwidth tuning allows the statistical manifold to... Maintaining a dynamic balance between dense windows that are not overly smooth and sparse windows that are not overly fragmented; hyperbolic tensor expansion increases the density of risk stratification. Automatic fine-grained mapping in high-risk areas and automatic coarse-grained mapping in low-risk areas saves computing resources while ensuring tail accuracy for high-risk assets; gradual scene adjustment enables tail mapping. Differential resilience is adopted for long-term gradual changes and sudden shocks to avoid false alarms caused by frequent threshold tightening; the confidence bay area convergence index provides a real-time safety isolation zone for the multimodal tail by dynamically refreshing the curvature fingerprint, preventing extreme coupling from causing overall threshold drift. The above mechanisms together improve the stability of manifold evolution and the sensitivity of risk stratification.

[0159] Step 302: Dynamic Threshold Shifting and Streaming Synchronization

[0160] With risk stratification density and tail mapping in place, a set of alarm thresholds needs to be generated in real time. This ensures that the threshold evolves continuously across both time and version axes; simultaneously, in a streaming environment, it is essential to provide elastic compensation for distribution drift, network jitter, and asset-level switching. This is illustrated by applying Wasserstein gradient flow to a kernel-smooth manifold. Upward thrust tail mapping:

[0161]

[0162] Alarm threshold set; Gradient operator for Wasserstein distance; For the tail portion;

[0163] Furthermore, gradient flow allows the threshold set to drift naturally along the direction of distribution deformation, without the need for manual resetting.

[0164] When Kullback-Leibler divergence is detected $Exceeding the threshold area At that time, reset the distribution center :

[0165]

[0166] Real-time central location, latest moment The central estimate, a four-dimensional vector, corresponds to the mean values ​​of electrical parameters, vibration quantities, electromagnetic expansion, and environmental quantities in the current risk layer; The lag interval;

[0167] Drift learning rate, ranging from 0.05 to 0.3. For the new observation center; : Drift threshold, estimated from historical stability window;

[0168] Among these measures, heavy centralization suppresses the erosion of the threshold by long-term trends, maintaining alarm sensitivity, and is combined with lifetime fragment codes. risk factor Add an elastic contraction function above the threshold:

[0169]

[0170] Lifecycle correction threshold; Risk coefficient, ranging from 0 to 1;

[0171] When used, high-risk assets cause the threshold to move towards the tail end, while low-risk assets maintain the normal threshold, thus achieving risk-driven contraction.

[0172] Finally, the recursive version vector is used. For windows, adjust the lifecycle threshold. Write to semantic library And generate a symptom monitoring summary after each write. The summary records the triggered risk layer, tail ternary number, and recentering displacement, used for anomaly synthesizer tracing. Streaming updates ensure the convergence of threshold, version, and signal, providing a summary for symptom monitoring. It allows the backend to quickly pinpoint the trigger source.

[0173] Step 302: Output the alarm threshold set Summary of Symptom Monitoring Together with the previous net signal and time label, they are written into the semantic library to provide complete input for the subsequent anomaly synthesizer.

[0174] Step 3, driven by both statistical manifold and gradient flow, achieves a seamless mapping from the net signal to the dynamic threshold. Step 301, through Wasserstein minimum manifold, Laplace-Beltrami projection, and g-and-h tail model, finely decomposes the multimodal statistical structure and recombines it according to lifetime-maintenance risk, thoroughly solving the two major problems of static threshold mismatch and difficulty in unifying multimodal tails. Step 302 uses gradient flow to allow the threshold to naturally migrate with the distribution shape, and employs heavy centralization and lifetime contraction methods to resist long-term drift and aging impacts. Furthermore, it uses a recursive version vector to write the threshold back to the semantic library, ensuring that the time chain, version chain, and risk chain are synchronized. Therefore, the anomaly synthesizer receives the net signal in step 4. Alarm threshold set Summary of Symptom Monitoring Afterwards, it can directly calculate the credibility score and initiate micro-diagnoses without having to verify the validity of the threshold, thus injecting continuous evolutionary driving force into the intelligent operation and maintenance of urban infrastructure.

[0175] In practice, the thermal diffusion regularization term and the diffusion spectrum together suppress gradient flow oscillations, achieving smooth threshold migration; the piecewise rebound threshold centralizes the weight and uses the recursive version vector. Tight coupling ensures that threshold updates naturally align with data segments, preventing erroneous convergence caused by sparse data at the end of segments. The two-factor product allows for elastic contraction of the lifecycle, providing differentiated responses to both "external environmental deterioration" and "internal aging progress," ensuring both immediate sensitivity to sudden scenarios and stable response to long-term aging trends. The hash-nested index digest writing mechanism avoids concurrent read / write conflicts and enables the anomaly synthesizer to quickly determine the direction, improving the smoothness of subsequent diagnosis.

[0176] Urban infrastructure, under conditions of extreme load, transient failures, and aging, often exhibits a three-phase migration from latent symptoms to apparent problems to deterioration. No single mode or threshold can provide a sufficiently reliable warning. Step three has already output the lifecycle correction threshold. Symptom monitoring summary and net signal This information establishes a three-axis coordinate system for risk identification, encompassing data, thresholds, and version. However, truly intelligent operation and maintenance decisions must further measure the credibility of potential faults and achieve virtual-real mutual calibration through closed-loop verification via on-site micro-diagnostics.

[0177] Step 4: Integrate net signal and threshold information to generate a reliable score, and verify potential faults with multi-source measured evidence in the micro-diagnosis closed loop. Finally, write the verified conclusions back to the semantic library to drive strategy evolution.

[0178] Step four includes the following:

[0179] Step 401 Anomaly Synthesis and Credibility Score Generation

[0180] Threshold hits only indicate that the metric has exceeded the limit, but cannot explain the reliability of the fault; the excess phase of different modalities, the weight of the lifetime stage, and the version synchronization status together determine the true probability of the fault.

[0181] Firstly, within the same version segment Internally, it collects net signal fragments, trigger threshold labels, and symptom summary fields, and assembles evidence vectors. Subsequently, using asset-attribute anchors as nodes and time continuity and modal coupling strength as weighted edges, an evidence graph is generated. Evidence diagram Integrating multimodal boundary crossings with version continuity provides topology support for subsequent consistent mapping.

[0182] Regarding the evidence diagram To address the time jitter caused by inter-segment latency, a minimum phase rotation mapping function is introduced. All edge weights are projected onto a unified time axis and phase ambiguity is eliminated. Phase compensation is applied to the edge weights to ensure that the subsequent reliable scoring algorithm is not contaminated by time difference.

[0183]

[0184] In the formula: This is a phase rotation mapping function that maps the original edge weights and time differences to a unified time axis.

[0185] Phase-aligned weighted edges are complex numbers with magnitudes between 0 and 1. The original edge weights are complex numbers with magnitudes between 0 and 1. : for nodes Time difference between nodes The reference angular frequency is determined by the time stamp. Alignment;

[0186] Using multi-layered graph attention networks for evidence graphs Perform feature aggregation and output node confidence embeddings. Subsequently, a Beta-soft logic function mapping was used as the reliability score:

[0187]

[0188] In the formula: : Credible score, range (0,1); It is the Beta-Sigmoid smoothing function;

[0189] Node confidence embedding, output by gated residual graph attention network; The query vector is obtained by concatenating risk weights, lifecycle fragment codes, and operation and maintenance priorities and then linearly mapping them. Note the weighting coefficients, which range from 0 to 1;

[0190] Among them, the soft logic function ensures that the confidence score changes continuously with the confidence embedding, avoiding decision jumps caused by hard thresholds.

[0191] The top [individuals] with the greatest impact on credibility scores Edge writing to the source index The index along with the confidence score Together, they provide a priority verification list for the micro-diagnostic process. Among them, the traceability index allows on-site diagnosis to focus on key coupling points, improving verification efficiency.

[0192] Step 401 outputs a credibility score With source index This indicates the diagnostic priority in step 402.

[0193] In use, the five-part graph structure, combined with the entropy equilibrium potential, gives the evidence graph a natural hierarchical interpretability. Note that the network can extract multi-layer semantics without additional regularization. Multi-frequency reference rotation enables cross-modal alignment to overcome fundamental wave limitations, maintaining phase uniformity even when dealing with complex electromagnetic scenarios involving subsynchronous sampling. Gated residuals and semantic edge chromatography together suppress deep oversmoothing and improve the edge differentiation of different asset types. Multi-level cache keys with link signatures ensure the credibility of the evidence-to-diagnosis link from the root, avoiding inconsistent interpretations caused by cache misalignment during the diagnosis stage.

[0194] Furthermore, multi-frequency reference rotation resolves phase misalignment caused by high-order harmonics and variable sampling rates; gated residuals introduce deep controllability in unlabeled graph attention networks; and link signatures endow traceability keys with native anti-tampering capabilities.

[0195] Step 402 Micro-diagnostic call and joint verification

[0196] A high credibility score does not necessarily mean that the fault has been confirmed. It is necessary to achieve a closed loop of virtual anomaly and physical evidence through joint verification of on-site sensors and comparison with historical semantics.

[0197] According to the source index The system allocates diagnostic windows between the camera gateway and the partial discharge sensor node based on edge weights. The asset corresponding to the heaviest weighted edge is given priority for the longest continuous imaging and discharge monitoring duration, while other assets receive progressively less weight, in order to utilize limited diagnostic bandwidth. Prioritized scheduling concentrates diagnostic resources on critical risk links, improving verification hit rate.

[0198] The camera captures high frame rate video streams in full-spectrum mode, while the partial discharge sensor synchronously records discharge pulse trains. Synchronous triggering is caused by time tags. Corresponding version segment base time The system ensures that video frames are consistent with the pulse train time base. Synchronous acquisition allows image features and electrical discharge features to be registered within milliseconds, facilitating subsequent fusion.

[0199] Extracting structured optical flow vectors from video frames ;

[0200] For discharge pulse train Perform FFT, Hilbert transform, or wavelet packet decomposition to extract the dominant frequency amplitude or energy distribution, and calculate the spectral fingerprint. Then, joint confidence is calculated using mutual information-cross-spectral coupling degree:

[0201]

[0202] Evidence coupling degree, scope ; Mutual information; Shannon entropy;

[0203] Among them, coupling degree measures the consistency of heterogeneous evidence, and a high $\kappa$ means that both visual and electrical signals point to the fault.

[0204] Generate combined diagnostic results This indicates that the fault has been confirmed. This indicates a false alarm.

[0205] Combined diagnostic results Along with video thumbnail fingerprints, discharge spectrum fingerprints, and confidence scores and the degree of coupling of evidence Write back to semantic library And trigger the threshold unit based on the joint diagnostic results. The result relates to the lifecycle risk coefficient of the next version segment. Adjustments are made to increase or decrease thresholds, enabling diagnosis-driven threshold evolution. Specifically, write-back synchronization creates a closed loop between the threshold and the diagnostic results, ensuring that the next round of alerts more closely aligns with physical reality.

[0206] Step 402: Output the combined diagnostic results It also collects multi-source fingerprints and triggers self-adjustment of threshold risk coefficients; at the same time, it persists the diagnostic content to the semantic library, completing the virtual-real integration closed loop.

[0207] Step four achieves a highly reliable closed loop from digital anomalies to physical verification through the collaboration of the anomaly synthesizer and the micro-diagnostic process. Step 401 uses multi-source evidence mapping, phase consistency mapping, and graph attention aggregation to map the net signal, threshold, and symptom summary into a quantitative reliability score, and outputs a source tracing index to guide on-site diagnosis. Step 402 relies on priority scheduling, synchronous acquisition, and mutual information-mutual spectrum fusion to complete visual-electrical dual-source verification within a minute-level time window, and writes the confirmation results back to the semantic database to drive the dynamic contraction or relaxation of the threshold risk weight. Thus, the intelligent operation and maintenance link forms an adaptive trapezoidal structure of acquisition—purification—threshold—scoring—diagnosis—evolution: any fault indication must be verified by both digital twin evidence and measured signals before it can be upgraded to an alarm, and the threshold strategy is iteratively updated after verification, realizing continuous self-evolution from data to knowledge to strategy.

[0208] Layered pop-up windows and dynamic bandwidth reallocation ensure that diagnostic links in wired and wireless hybrid network environments always prioritize serving high-risk assets; near-infrared and differential hashing technologies expand the diagnostic scenarios at night while minimizing discharge bandwidth usage; adaptive cross-spectral windows and threshold drift bands provide a reference and loop for coupling degree calculation, ensuring that high-confidence anomalies are not "diluted" by light or noise; frame block-aware hashing and synchronous writing to the discharge feature library allow diagnostic images and electrical features to accumulate over time, providing a "image-electrical" dual-perspective golden dataset for subsequent fault evolution research.

[0209] When intelligent operation and maintenance of urban infrastructure enters the diagnostic-evolution terminal stage, the system faces the complex challenge of two-way real-time collaboration: on the one hand, the micro-diagnostic process has already incorporated joint diagnostic results... Writing multi-source fingerprints to the semantic library This triggers a sliding window adjustment of the risk coefficient; on the other hand, the operation and maintenance command center must accurately push high-value alarms to the responsible parties within a second-level window based on different levels of response time limits, resource availability, and security levels, ensuring that the emergency repair link and the threshold link iterate in parallel. If there is a lack of an efficient knowledge write-back-routing strategy, the diagnostic results will only remain in the database, leading to a disconnect between the threshold unit and the policy library, resulting in duplicate alarms, misaligned instructions, or redundant resource allocation.

[0210] Step 5: Accumulate the diagnostic results and operation logs at the knowledge layer and route them at the policy layer to ensure that the threshold units and operation and maintenance policies evolve synchronously within a unified version coordinate system.

[0211] Step five includes the following:

[0212] Step 501 Knowledge Rewriting and Heterogeneous Blind Spot Resolution

[0213] Combined diagnostic results Fingerprint information is an isolated data fragment, and directly writing it into the database does not help with later retrieval and strategy deduction; the operation and maintenance team needs to quickly retrieve similar historical cases based on different knowledge perspectives (device perspective, scenario perspective, lifecycle perspective).

[0214] First, generate a ternary index sequence: asset-attribute index. Risk-Lifecycle Index Version-Time Index The system constructs a high-dimensional index matrix from the ternary index using Kronecker-Hadamard composite operations:

[0215]

[0216] In the formula: High-dimensional index matrix, encoding a three-dimensional retrieval path; Asset-Attribute Index; Risk-Lifecycle Index; Version-time index;

[0217] Among them, the composite index matrix ensures that the same diagnostic result maintains a unique anchor point from different perspectives, thus solving the problem of cross-domain retrieval collision.

[0218] Run log The segment is divided into three levels: topic, event, and action. Each segment is mapped to a semantic vector. Then, the Jensen-Shannon divergence was used to measure the heterogeneity between fragments:

[0219]

[0220] In the formula: Semantic divergence; , Semantic vector; : intermediate distribution;

[0221] Among them, fragments with divergence below the threshold are aggregated into knowledge atoms, improving the efficiency of cross-log semantic aggregation.

[0222] Even after clustering, diagnostic results may still leave semantic blind spots—isolated atoms without historical context. This necessitates calculating the knowledge graph node coverage. Regarding coverage Coverage threshold Complete the atomic execution structure transfer map:

[0223]

[0224] In the formula: Isolated semantic atoms; : Set of candidate structure centers; Best Compensation Center

[0225] Among them, minimum distance compensation connects isolated atoms to the nearest structural center, eliminating blind spots.

[0226] Finally, the knowledge atom sequence along with the index matrix Packaged into knowledge blocks and use consistent fingerprints High-level signatures prevent tampering. Version consistency encapsulation establishes a knowledge-version mapping, ensuring that threshold units can read historical knowledge along the version chain.

[0227] Step 501: Output knowledge blocks It provides structured input with multidimensional indexes for routing strategies.

[0228] In practice, hierarchical segmented hashing compresses the three-dimensional index and ensures reversible queries, significantly reducing query latency; topic recursive extraction and bidirectional action completion emphasize causal integrity, avoiding misjudgments due to missing drivers in future strategy deductions; substructure co-occurrence measurement gives isolated atoms "semantic neighbors," increasing the link density of the knowledge graph and further reducing the probability of cold starts; sliding chaincode integrates version, fingerprint, and semantic chains, supporting bidirectional verification after cross-node knowledge migration. The superposition of multiple mechanisms makes knowledge blocks not only lightweight and highly interconnected but also secure and reliable, providing complete features for subsequent entropy weight evaluation.

[0229] Step 502 Multi-level routing coordination and policy evolution

[0230] Urban operation and maintenance systems are typically divided into three levels: monitoring center, regional dispatch, and field teams. Each level has different requirements for alarm elements, response time limits, and resource constraints. To avoid information overload and resource mismatch, a knowledge-based approach is essential. The minimum entropy route is dynamically calculated based on real-time network status to ensure that high-risk alarms reach the responsible party directly, and the threshold unit and policy library are simultaneously driven to evolve adaptively. Specifically:

[0231] The risk level, resource consumption, and response time of the knowledge block are constructed into an evaluation vector. Calculate entropy weights :

[0232]

[0233] In the formula: : Index entropy weight Information entropy;

[0234] : No. The alarm is in the first The normalized value of the indicator; Number of alarms

[0235] Among them, the larger the entropy weight, the stronger the difference in indicators and the higher the priority.

[0236] For the three-tier operation and maintenance entity, a mapping matrix is ​​set up. ,element Indicates an alarm Towards hierarchy The matching weights, and the optimal route is obtained by solving the minimum entropy matching with capacity constraints:

[0237]

[0238] In the formula: : Routing matrix; : level capacity;

[0239] For the index entropy weight, This is a routing decision variable, with values ​​between 0 and 1. The matching weight takes a value between 0 and 1. For alarm indexing, For operation and maintenance level index; This is the entropy difference term;

[0240] Among them, minimum entropy matching ensures that the difference between the route distribution and the entropy weight distribution is minimized, while respecting capacity.

[0241] Routing matrix Triggering a policy library update: For each matching edge, record the response latency - processing time - effect feedback triple and write it into the policy evolution vector. Subsequently, the policy gain was estimated using Kalman-Bayes fusion. It also updates the operation and maintenance strategy table to achieve experience-based Bayesian incremental evolution. Incremental evolution uses real-world feedback to adjust strategy priorities, achieving rolling optimization.

[0242] Policy Gain Mapped to risk coefficient correction The threshold unit reads the correction term to update the sliding window mean, ensuring that the response improvement brought about by policy optimization is directly reflected in threshold contraction / widening. Among them, threshold synchronization allows the results of policy evolution to immediately feed back into subsequent alarms, forming a closed loop.

[0243] Step 502 outputs the optimal route. Strategy Gain With risk correction And push the alarm to each level.

[0244] Step 501 transforms the diagnostic results into knowledge blocks with version signatures through a high-dimensional index matrix, semantic fragmentation, and blind spot compensation. This solves the problem of cross-domain retrieval and historical playback; step 502 uses entropy weight evaluation and minimum entropy matching to realize dynamic routing of the three-level operation and maintenance entities, and uses the experience-Bayes incremental evolution strategy library to map the strategy gain to risk correction and provide real-time feedback of threshold units.

[0245] In practice, steady-state testing is combined with entropy weight trajectory, allowing entropy weight to better reflect long-term trends rather than short-term noise; delay-redundancy threshold incorporates network physical state into routing decisions, ensuring that links can still reach high-priority alarms even under extreme loads; trust threshold window dynamically adjusts prior exploration during policy evolution—utilizing balance to shorten the policy convergence cycle without being overly conservative; fault-state weighting function matches the threshold shrinkage gradient with the severity of the fault, preventing the system from becoming overly sensitive to continuous minor faults.

[0246] 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.

[0247] 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.

[0248] 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.

[0249] 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.

[0250] 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 smart operation and maintenance method for urban infrastructure based on digital asset twin mapping, characterized by: include, Parallel acquisition of electrical parameters, vibration parameters, temperature and humidity parameters, and electromagnetic parameters; after processing, data packets with consistent fingerprints and version vectors are generated and mapped into the semantic library. At the edge side, supercomplex time alignment, complex cepstral equalization, and fractional wave packet coherent noise removal are performed on the data packets to output the net signal and time label. The alarm threshold set is adaptively calculated using Wasserstein statistical manifold combined with lifecycle risk coefficient, and a symptom monitoring summary is generated. An anomaly synthesizer constructs a five-point evidence graph, obtains a confidence score through a gated residual graph attention network, generates a source index, outputs multi-level diagnostic results, and writes them back to the semantic database. Knowledge blocks are encapsulated using hierarchical segmented hashing and pushed to the hierarchical operation and maintenance entity based on entropy weight-minimum entropy matching. At the same time, the policy gain is mapped to the risk coefficient and the threshold is updated synchronously.

2. The intelligent operation and maintenance method for urban infrastructure according to claim 1, characterized in that: Nanosecond-level synchronous triggering is achieved through time base assimilation, enabling real-time synchronous acquisition of electrical parameters, temperature and humidity, vibration, and electromagnetic parameters. During the acquisition process, electrical parameters are vectorized and extracted and zero-order compensation vectors are injected. Fractional filtering is used to suppress trend noise for environmental signals.

3. The intelligent operation and maintenance method for urban infrastructure according to claim 2, characterized in that: The collected information is embedded with lifecycle fragment codes and a unified semantic identifier is generated and encapsulated into a data packet. Semantic mapping is performed on data packets, consistency fingerprints are calculated and recursive version vectors are generated, data packets are transmitted to the semantic library through elastic streaming segmentation, and the evidence chain is extended based on ontology to maintain semantic consistency.

4. The intelligent operation and maintenance method for urban infrastructure according to claim 3, characterized in that: After performing supercomplex time alignment on the multimodal sequences, pseudo-periodic resonances are suppressed by retuning the complex cepstral envelope. Then, coherent islands are extracted by cross-scale scanning of fractional Gabor-Meyer wavelet packets. A semantic pruning graph is constructed based on lifecycle risk weights, maintenance priority weights, and quality label vectors to output a candidate signal set.

5. The intelligent operation and maintenance method for urban infrastructure according to claim 4, characterized in that: After being decomposed by Hankel-Toeplitz sparse low-rank double decomposition, the candidate signal enters the temporal attention fusion unit to generate attention weights associated with the risk coefficient. Attention weights are bound to the recursive version vector to form time tags, and the net output signal is encapsulated according to semantic pointer rules.

6. The intelligent operation and maintenance method for urban infrastructure according to claim 5, characterized in that: The statistical manifold trajectory is generated using the second-order Wasserstein distance minimization criterion, and then the risk layer is divided by Laplace-Beltrami projection based on the risk-lifetime tensor. Each risk layer uses a g-and-h extrapolation kernel to obtain the tail partition mapping, and introduces the Fréchet-Hoeffding upper bound to achieve multimodal tail collaborative constraints.

7. The intelligent operation and maintenance method for urban infrastructure according to claim 6, characterized in that: The alarm threshold set is automatically migrated along the manifold gradient flow, and recentering is performed when the Kullback-Leibler divergence is detected to exceed the drift threshold. Simultaneously, the life cycle risk coefficient is read and the threshold width is adjusted through an elastic shrinkage function. The corrected threshold and the recursive version vector are synchronously written into the semantic library and an additional symptom monitoring summary is attached.

8. The intelligent operation and maintenance method for urban infrastructure according to claim 7, characterized in that: The anomaly synthesizer uses asset nodes, attribute nodes, version nodes, lifetime nodes, and threshold nodes to form a five-part evidence graph; After compensating for cross-modal time difference through minimum phase multi-frequency rotation, a gated residual graph attention network is used to output a reliable score, and the edge information that contributes the most to the score is written into the source index.

9. The intelligent operation and maintenance method for urban infrastructure according to claim 8, characterized in that: Based on the source index, priority is given to allocating camera and discharge diagnosis windows, and video optical flow vectors and discharge pulse differential hashes are acquired simultaneously, and mutual information-cross-spectral coupling degree is calculated. The diagnostic results are divided into three levels: confirmed, suspected, and excluded. The results, along with the fingerprint information, are written into the semantic database to trigger a sliding window update of the risk coefficient.

10. The intelligent operation and maintenance method for urban infrastructure according to claim 9, characterized in that: The runtime log is recursively extracted from topics and bidirectionally filled in with actions to generate semantic fragments, which are then aggregated into knowledge atoms using Jensen-Shannon divergence. If the knowledge graph coverage is lower than a specified threshold, isolated atoms are connected to the nearest structure center through substructure co-occurrence measurement, and finally encapsulated into knowledge blocks with high-dimensional index matrix, version vector and sliding chain code.

11. The intelligent operation and maintenance method for urban infrastructure according to claim 10, characterized in that: Alarm routing uses information entropy weights of danger level, resource occupancy and response time as weights, and solves the minimum entropy matching under capacity and delay constraints to directly send high-weight alarms to the corresponding operation and maintenance level; The feedback process updates the operation and maintenance strategy vector using an empirical-Bayesian method, and maps the strategy gain to a risk correction amount, which is then synchronized to the threshold calculation.

Images