Property equipment intelligent inspection and accurate predictive maintenance system and method based on multi-source data fusion

Abstract

The invention discloses a property equipment intelligent inspection and accurate predictive maintenance system and method based on multi-source data fusion, and relates to the technical field of Internet of Things and sensors, and the system comprises a multi-source data access module which synchronously accesses an ERP standing book, an IoT monitoring stream and an unstructured operation and maintenance record through a multi-protocol adapter; the dynamic fusion engine executes asynchronous data space-time alignment during operation, a sliding time window mechanism realizes space-time alignment, the unstructured data processing unit adopts a field fine tuning BERT model to extract a fault triple, and a two-channel convolutional network fuses visible light crack detection and thermal imaging abnormal region analysis; a distributed intelligent framework is adopted, lightweight LSTM real-time detection is deployed on an edge layer, and a cloud end realizes cross-project federated learning through a differential privacy gradient aggregator; and the decision execution interface is used for triggering three-level early warning based on a fault probability threshold value and an inventory safety coefficient, and optimizing a maintenance path by fusing GIS real-time road conditions.

Description

Technical Field

[0001] This invention relates to the fields of Internet of Things and sensor technology, specifically to a system and method for intelligent inspection and precise predictive maintenance of property equipment based on multi-source data fusion. Background Technology

[0002] Traditional maintenance methods have certain limitations. Regular maintenance is costly and prone to over-maintenance, such as mandatory daily elevator inspections which may lead to resource waste; maintenance after a malfunction results in significant downtime losses, such as air conditioning failures which may affect the comfort of the entire building.

[0003] With the maturity of IoT and AI technologies, the decline in sensor costs, and the widespread adoption of 5G communication, real-time data collection from equipment has become possible. AI models (such as LSTM and Transformer) can capture complex operating patterns of equipment, achieving a prediction accuracy of over 90%. The demand for digital transformation in the property management industry is prominent, with industry trends shifting from "firefighting maintenance" to "predictive maintenance," aiming to reduce maintenance costs by 30%-50% and extend equipment lifespan by 20%-40%. China's planning emphasizes the Industrial Internet and intelligent manufacturing, requiring the property management sector to respond to the demands for intelligent equipment management.

[0004] Therefore, there is an urgent need for a smart inspection and accurate predictive maintenance system and method for property equipment based on multi-source data fusion. Facing the challenges of data integration and quality, including heterogeneous data silos, inconsistent formats between equipment ledgers (ERP system) and real-time monitoring data (IoT) requiring manual cleaning; and unstructured data processing, such as maintenance work order text and images, which require natural language processing (NLP) and image recognition technology for parsing, increasing complexity. Summary of the Invention

[0005] To address the aforementioned technical challenges, this paper presents a system and method for intelligent inspection and precise predictive maintenance of property equipment based on multi-source data fusion. This technical solution resolves the issues of heterogeneous data silos and unstructured data processing.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A property equipment intelligent inspection and precise predictive maintenance system based on multi-source data fusion includes: Multi-source data access module: Built-in OPCUA / Modbus / HTTP2 multi-protocol adapter, synchronously accessing three types of heterogeneous data sources, including ERP equipment ledger data, IoT real-time monitoring streams, and unstructured operation and maintenance records; Dynamic fusion engine: performs asynchronous data spatiotemporal alignment at runtime, dynamically compensates for the transmission delay between ERP coordinate data and IoT real-time coordinates through a sliding time window; unstructured data processing unit: uses a domain-fine-tuned BERT model to parse maintenance work order text, outputs standardized fault triples, and identifies surface cracks and thermal anomaly areas in fault images based on a dual-channel attention convolutional network. Distributed modeling framework: Lightweight LSTM models are deployed at the edge layer for real-time anomaly detection; in the cloud, differential privacy gradient aggregator is used to fuse model parameters from across property projects. Decision execution interface: When the probability of equipment failure is greater than a predetermined proportion and the safety factor of spare parts inventory is less than a predetermined threshold, the spare parts supply chain early warning instruction is automatically triggered, and the dynamic scheduling scheme of maintenance personnel is integrated with GIS road conditions.

[0007] Preferably, the multi-source data access module specifically includes: Multi-protocol adapter architecture: The underlying communication protocol stack adopts a layered design to achieve parallel support for OPCUA, Modbus, and HTTP / 2 protocols. The physical layer supports adaptive switching between RS-485 / Ethernet / 5G wireless transmission media; the protocol conversion layer converts Modbus RTU messages into the OPCUA information model; and the security layer enables TLS1.3 encryption and two-way certificate authentication for HTTP / 2 connections. Three types of data source access: ERP equipment ledger data access: structured relational data, with key fields including equipment ID, 3D spatial coordinates, warranty expiration date, and historical maintenance count; extraction mechanism, monitoring database binlog changes via HTTP / 2 long connection; incremental update frequency, triggering synchronization whenever the number of record changes exceeds a predetermined threshold; IoT real-time monitoring stream access: high-frequency timing signals, including vibration acceleration, winding temperature, and three-phase current; acquisition control logic: pre-configured acquisition strategy at the edge, normal sampling when the device is in operation, and accelerated sampling when the temperature is higher than the predetermined temperature; Unstructured maintenance records are accessed: maintenance work order text, which comes from JSON messages submitted by scanning a QR code on the mobile APP. Key fields include fault description text, urgency level label, and reporting personnel ID; fault images support formats including JPEG / PNG / infrared thermal imaging DICOM; voice logs are collected on-site through a smart safety helmet, with 16kHz sampling rate single-channel PCM audio.

[0008] Preferably, the multi-source data access module specifically includes: Synchronous access control mechanism: Time synchronization is guaranteed by using the IEEE 1588 Precise Time Protocol (PTP) to timestamp all data; resource conflict resolution is achieved by arbitrating based on priority when multiple data sources access the system concurrently, with priority in the following order: emergency fault alarm, real-time monitoring stream, maintenance record update, and ledger data synchronization. It supports resuming data transmission after network outages, with a local cache queue capacity that supports storing 72 hours of raw data. After the network is restored, it can resume transmission from the point of interruption through hash verification. Data preprocessing pipeline: Structured data cleaning, ERP coordinate data verification, and automatic mapping to the nearest floor if the z-coordinate exceeds the building height range; streaming data denoising, and wavelet threshold denoising for vibration signals; unstructured data encapsulation, binding maintenance work order text and associated images into a unified event ID, and storing it in Avro binary format.

[0009] Preferably, the dynamic fusion engine specifically includes: Asynchronous data time-space alignment: Dynamic compensation for transmission delay is achieved by using a timestamp correction algorithm to dynamically compensate for the original timestamps of the ERP coordinate data; sliding time window synchronization control is implemented with adaptive adjustment of the window size, and when the high-frequency IoT data stream arrives, matching ERP coordinate data is retrieved and output within the window; The spatial coordinate mapping engine completes coordinate system transformation based on the transformation model. The ERP system uses the WGS-84 geographic coordinate system, while the IoT positioning data uses a local rectangular coordinate system. Equipment displacement tolerance control triggers position anomaly verification when the rate of change of adjacent coordinates is greater than 0.5 m / s.

[0010] Preferably, the dynamic fusion engine specifically includes: Unstructured data processing unit: Maintenance work order text parsing: Domain-adaptive BERT model, where the base model is BERT-Large, and the fine-tuning dataset consists of property equipment failure work orders, covering elevator / HVAC / power supply systems; Fault triplet extraction and entity recognition are performed. Equipment components / fault phenomena are labeled using BI-LSTM-CRF layers. The standardized output structure includes equipment entity, fault type, and root cause. Dual-channel analysis of fault images: Visible light channel processing path, network structure is ResNet-50+ spatial attention module; crack detection process is input image, adaptive histogram equalization, edge detection, attention weight map generation, output crack pixel ratio; thermal imaging channel processing path, high temperature area is extracted through temperature field analysis algorithm; thermal anomaly area ratio is calculated.

[0011] Preferably, the distributed modeling framework specifically includes: Lightweight LSTM architecture for edge layer: Model compression, parameter control (total model weights less than 1MB), use depthwise separable convolutions instead of fully connected layers, weight pruning (removing weights less than 10). -4 Redundant parameters; 8-bit integer quantization, FP32 to INT8 conversion; The anomaly detection embedded in the physical mechanism includes input features such as extracting vibration signals through wavelet packet energy entropy and extracting temperature curves through first-order derivative mutation point detection. If the anomaly probability of the LSTM output is greater than a predetermined threshold, a physical constraint violation flag is determined, triggering an edge-level alarm and uploading the original data slice.

[0012] Preferably, the distributed modeling framework specifically includes: Cloud-based differential privacy gradient aggregator: Privacy protection is achieved through a noise injection model using Gaussian noise and a gradient pruning threshold to calculate the aggregated gradient. A cross-project knowledge fusion strategy is employed, with device similarity weighting, calculating topological similarity and weighted aggregation. Dynamic learning rate adjustment is implemented, where the learning rate is adjusted when the gradient variance of three consecutive aggregation rounds exceeds a predetermined threshold.

[0013] Preferably, the decision execution interface specifically includes: Multi-level early warning triggering mechanism: A dynamic threshold for failure probability and a tiered response strategy are implemented, with a base threshold of 85%. Dynamic adjustment rules are in place: if the equipment's age exceeds 80% of its design life, the threshold is lowered to 75%; if the monthly failure frequency of similar equipment exceeds 3, the threshold is raised to 90%. Confidence check: When the predicted probability is within a certain range, the vibration spectrum verification is initiated. Inventory safety factor modeling and intelligent correction factor: if the supplier's on-time delivery rate is less than 95%, the threshold is raised to 1.5; if the equipment is under warranty, the threshold is lowered to 0.9.

[0014] Preferably, the decision execution interface specifically includes: Spare parts supply chain early warning: The three-level early warning system is as follows: Yellow warning: The ERP system automatically generates a purchase requisition form; Orange warning: The backup supplier bidding process is activated; Red warning: The cross-property project allocation agreement is triggered. Intelligent sourcing engine: supply network analysis, local warehouse search for available inventory within its radius; cloud collaboration: matching inventory status of other property projects with the same equipment model; cost optimization decision-making: controlling procurement costs; Dynamic scheduling of maintenance personnel, GIS traffic fusion engine, real-time traffic data processing, integration with Gaode / Baidu real-time traffic APIs, and calculation of road resistance factors; route optimization algorithm, improved A Algorithm introduction, time window constraints, skill matrix matching; Multi-objective scheduling model: Decision variables include total response time, manpower cost, and service timeout risk; personnel skills are greater than or equal to the fault repair requirement level.

[0015] Furthermore, intelligent inspection and precise predictive maintenance methods for property equipment based on multi-source data fusion include: Synchronous data collection is achieved through multi-protocol adapters, including ERP equipment ledgers, IoT real-time monitoring streams, and unstructured operation and maintenance records; spatiotemporal dynamic alignment is performed, and a transmission delay compensation model is established for low-frequency ERP coordinate data; synchronization with high-frequency IoT data is achieved through a sliding time window; For maintenance work order text processing, a domain-fine-tuned BERT model is used to extract triple entity relationships, including equipment entity, fault type, and root cause; semantic correction is performed by embedding the ISO13379-2023 fault tree rule base; fault image analysis is performed, using the visible light channel and a spatial attention convolutional network to quantify the crack length ratio; and the thermal imaging channel is used to extract abnormal areas with temperatures exceeding a predetermined proportion of the rated value, and then fused and output. A lightweight LSTM model is deployed at the edge layer to perform real-time anomaly detection; cross-project model parameters are fused in the cloud through a differential privacy gradient aggregator. When the failure probability is greater than the dynamic threshold and the inventory safety factor is less than 1.2, the three-level early warning of the spare parts supply chain is automatically triggered; maintenance personnel are dispatched, and the route is optimized by integrating real-time GIS traffic conditions.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: The multi-source data access module breaks down the heterogeneous data barriers between ERP ledgers, IoT monitoring, and operation and maintenance records through OPCUA / Modbus / HTTP2 multi-protocol adapters, enabling full-dimensional data synchronization access, providing comprehensive data support for precise operation and maintenance, and solving the problem of fragmented traditional inspection data.

[0017] The asynchronous spatiotemporal alignment mechanism of the dynamic fusion engine compensates for data transmission delay through a sliding time window, ensuring the consistency of ERP and IoT coordinate data. The domain-fine-tuned BERT model transforms unstructured chemical documents into standardized fault triples, and the dual-channel attention convolutional network accurately identifies cracks and thermal anomalies in fault images, significantly improving the utilization rate of unstructured data and breaking through the bottleneck of low efficiency and large error in traditional manual analysis.

[0018] The distributed modeling framework adopts an "edge-cloud" collaborative mode. The lightweight LSTM model at the edge layer enables real-time anomaly detection and millisecond-level response to potential device risks. The differential privacy gradient aggregator in the cloud integrates parameters from across projects, which not only improves the model's generalization ability but also ensures data privacy and security, avoiding modeling bias caused by insufficient data in a single project.

[0019] The decision-making execution interface constructs a closed-loop operation and maintenance system, automatically triggering supply chain early warnings through the linkage of dual thresholds for fault probability and spare parts inventory; the scheduling scheme integrating GIS traffic conditions enables dynamic optimization of maintenance resources, shortening fault handling time. The overall system inspection efficiency and fault prediction accuracy are improved, significantly reducing unplanned downtime losses and operation and maintenance costs. Attached Figure Description

[0020] Figure 1 This is an internal framework diagram of a property equipment intelligent inspection and precise predictive maintenance system based on multi-source data fusion. Figure 2 This is a flowchart of a method for intelligent inspection and precise predictive maintenance of property equipment based on multi-source data fusion. Detailed Implementation

[0021] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0022] Reference Figure 1 As shown, the intelligent inspection and precise predictive maintenance system for property equipment based on multi-source data fusion includes: Multi-source data access module: Built-in OPCUA / Modbus / HTTP2 multi-protocol adapter, synchronously accessing three types of heterogeneous data sources, including ERP equipment ledger data, IoT real-time monitoring streams, and unstructured operation and maintenance records; Dynamic fusion engine: performs asynchronous data spatiotemporal alignment at runtime, dynamically compensates for the transmission delay between ERP coordinate data and IoT real-time coordinates through a sliding time window; unstructured data processing unit: uses a domain-fine-tuned BERT model to parse maintenance work order text, outputs standardized fault triples, and identifies surface cracks and thermal anomaly areas in fault images based on a dual-channel attention convolutional network. Distributed modeling framework: Lightweight LSTM models are deployed at the edge layer for real-time anomaly detection; in the cloud, differential privacy gradient aggregator is used to fuse model parameters from across property projects. Decision execution interface: When the probability of equipment failure is greater than a predetermined proportion and the safety factor of spare parts inventory is less than a predetermined threshold, the spare parts supply chain early warning instruction is automatically triggered, and the dynamic scheduling scheme of maintenance personnel is integrated with GIS road conditions.

[0023] It should be noted that the data silo circuit breaker mechanism achieves seamless connection between millisecond-level IoT streaming data and hourly-level ERP ledgers in property scenarios through the layered protocol stack design of multi-protocol adapters, compressing transmission latency fluctuations to ±100ms. Unstructured value extraction: Dual-channel image analysis breaks through the limitations of single sensors, and cross-modal fusion of visible crack detection and thermal imaging over-temperature early warning reduces the fault misjudgment rate.

[0024] The multi-source data access module specifically includes: Multi-protocol adapter architecture: The underlying communication protocol stack adopts a layered design to achieve parallel support for OPCUA, Modbus, and HTTP / 2 protocols. The physical layer supports adaptive switching between RS-485 / Ethernet / 5G wireless transmission media; the protocol conversion layer converts Modbus RTU messages into the OPCUA information model; and the security layer enables TLS1.3 encryption and two-way certificate authentication for HTTP / 2 connections. Three types of data source access: ERP equipment ledger data access: structured relational data, with key fields including equipment ID, 3D spatial coordinates, warranty expiration date, and historical maintenance count; extraction mechanism, monitoring database binlog changes via HTTP / 2 long connection; incremental update frequency, triggering synchronization whenever the number of record changes exceeds a predetermined threshold; IoT real-time monitoring stream access: high-frequency timing signals, including vibration acceleration, winding temperature, and three-phase current; acquisition control logic: pre-configured acquisition strategy at the edge, normal sampling when the device is in operation, and accelerated sampling when the temperature is higher than the predetermined temperature; Unstructured maintenance records are accessed: maintenance work order text, which comes from JSON messages submitted by scanning a QR code on the mobile APP. Key fields include fault description text, urgency level label, and reporting personnel ID; fault images support formats including JPEG / PNG / infrared thermal imaging DICOM; voice logs are collected on-site through a smart safety helmet, with 16kHz sampling rate single-channel PCM audio.

[0025] Synchronous access control mechanism: Time synchronization is guaranteed by using the IEEE 1588 Precise Time Protocol (PTP) to timestamp all data; resource conflict resolution is achieved by arbitrating based on priority when multiple data sources access the system concurrently, with priority in the following order: emergency fault alarm, real-time monitoring stream, maintenance record update, and ledger data synchronization. It supports resuming data transmission after network outages, with a local cache queue capacity that supports storing 72 hours of raw data. After the network is restored, it can resume transmission from the point of interruption through hash verification. Data preprocessing pipeline: Structured data cleaning, ERP coordinate data verification, and automatic mapping to the nearest floor if the z-coordinate exceeds the building height range; streaming data denoising, and wavelet threshold denoising for vibration signals; unstructured data encapsulation, binding maintenance work order text and associated images into a unified event ID, and storing it in Avro binary format.

[0026] It should be noted that the protocol stack is designed in layers: At the physical layer, multi-mode communication intelligent decision-making uses a channel quality assessment matrix to dynamically select the transmission medium. When the 5G signal strength is >-90dBm and the latency is <20ms, 5G wireless transmission is prioritized (bandwidth utilization is improved by 40%); when electromagnetic interference >30dBµV is detected (such as in a power distribution room environment), it automatically switches to the RS-485 anti-interference link; Ethernet is used as the baseline medium and PoE power supply is supported (compliant with IEEE 802.3bt standard). Protocol conversion layer, the core algorithm for Modbus to OPCUA conversion: Address space mapping: Encode the Modbus register address as OPCUANodeId; Strong data type validation: Sign extension is applied to 16-bit integers (to prevent negative value overflow errors). Real-time performance guarantee: Conversion delay ≤ 3ms (meets IEC61131-3 PLC control cycle requirements); Security layer, TLS 1.3 encryption policy, only TLS_AES_256_GCM_SHA384 cipher suite is enabled (meeting NISTSP800-193 standard); session key refresh cycle ≤15 minutes (exceeding GDPR default 30-minute requirement). Enhanced two-way authentication: device-side certificate embeds device fingerprint (integrating MAC address + CPUID hash value), and the Certificate Revocation List (CRL) is updated synchronously every 5 minutes.

[0027] Data source access: ERP data extraction, binlog monitoring fault tolerance mechanism, breakpoint resumption, recording synchronization position based on GTID (Global Transaction Identifier); data consistency verification, calculating MD5 digest comparison after each batch synchronization; Dynamic threshold adjustment rules, adaptive algorithm for change record thresholds, reduced synchronization frequency during low bandwidth, and real-time synchronization during peak business hours; IoT stream acquisition, overheating accelerated sampling strategy:

[0028] Vibration signal anti-aliasing processing, with a pre-filter of 200kHz low-pass filter (attenuation slope -120dB / oct) to eliminate high-frequency noise above 10kHz.

[0029] Unstructured data: Fault image format compatibility: Infrared thermal imaging DICOM parsing, extracting key metadata including emissivity setpoint, ambient temperature, and measurement distance; Lossless JPEG / PNG compression, transcoded using WebP format (compression rate improved by 35% while retaining EXIF ​​tags); Noise reduction optimization for voice logs: RNN-based impulse noise suppression to eliminate helmet impact noise (energy > 80 dBSPL); keyword-triggered recording, storing only when words such as "fault" and "alarm" are recognized.

[0030] Synchronization control mechanism: Timing synchronization, PTP master clock optimization algorithm: Clock source priority: GPS atomic clock > carrier NTP > local OCXO crystal oscillator; Path delay compensation: IEEE 1588v2 specification; Clock drift constraint: A PID control algorithm is used to suppress crystal oscillator drift within ±0.1ppm; Resource conflict arbitration: The hardware-accelerated priority decision-maker implements a 4-level priority encoding circuit in the FPGA, and the emergency fault alarm signal is directly connected to the DMA channel (preempting bus control). A bandwidth reservation mechanism is in place to reserve ≥50Mbps of dedicated bandwidth for real-time monitoring streams (using TSN time-sensitive network scheduling). Resuming data transfer after network outage: cached data persistence strategy, dual storage media: RAM cache (read / write at the μs level) + eMMC persistence (to prevent power loss); write balancing optimization: using the WL (WearLeveling) algorithm to extend Flash lifespan.

[0031] Pre-processing line: The system includes a rule engine for structured data cleaning, spatial coordinate error correction, a building height knowledge base, and storage of the baseline elevation for each floor (with periodic laser ranging calibration). Physical constraints for streaming data denoising, wavelet denoising parameters:

[0032] Industry compliance processing ensures that the vibration signals meet the vibration intensity analysis requirements of ISO10816-2025 after processing. Unstructured data encapsulation: Event ID generation specifications: Structure: Project ID (8 digits) + Device Category Code (4 digits) + Timestamp (Unix milliseconds); Global uniqueness is guaranteed, combined with the Snowflake algorithm for distributed IDs; Avro pattern evolution rules: backward compatibility design, new fields are filled with default values; legal evidence collection support, the original file hash value is recorded in the metadata.

[0033] The dynamic fusion engine specifically includes: Asynchronous data time-space alignment: Dynamic compensation for transmission delay is achieved by using a timestamp correction algorithm to dynamically compensate for the original timestamps of the ERP coordinate data; sliding time window synchronization control is implemented with adaptive adjustment of the window size, and when the high-frequency IoT data stream arrives, matching ERP coordinate data is retrieved and output within the window; The spatial coordinate mapping engine completes coordinate system transformation based on the transformation model. The ERP system uses the WGS-84 geographic coordinate system, while the IoT positioning data uses a local rectangular coordinate system. Equipment displacement tolerance control triggers position anomaly verification when the rate of change of adjacent coordinates is greater than 0.5 m / s.

[0034] Unstructured data processing unit: Maintenance work order text parsing: Domain-adaptive BERT model, where the base model is BERT-Large, and the fine-tuning dataset consists of property equipment failure work orders, covering elevator / HVAC / power supply systems; Fault triplet extraction and entity recognition are performed. Equipment components / fault phenomena are labeled using BI-LSTM-CRF layers. The standardized output structure includes equipment entity, fault type, and root cause. Dual-channel analysis of fault images: Visible light channel processing path, network structure is ResNet-50+ spatial attention module; crack detection process is input image, adaptive histogram equalization, edge detection, attention weight map generation, output crack pixel ratio; thermal imaging channel processing path, high temperature area is extracted through temperature field analysis algorithm; thermal anomaly area ratio is calculated.

[0035] It should be noted that asynchronous data requires time and space alignment: Transmission delay compensation, timestamp correction algorithm, and the mechanism for generating the dynamic compensation coefficient β=0.78: A bimodal Gaussian distribution was fitted based on 90 days of historical delay data. Use the upper limit of the 99% confidence interval as the benchmark value; A dynamic refresh mechanism recalculates the measured network latency Δt_lag every 5 minutes, using an exponentially weighted moving average: Δt_lag_new = 0.7 × Δt_lag_old + 0.3 × real-time latency. In the formula, Δt_lag_new is the new measured network latency value; Δt_lag_old is the old measured network latency value. Adaptive control of sliding time windows, dynamic equation for window size: Basic window width =2×max(Δt_lag), initial value 10 seconds Dynamic adjustment rule: If Δt_lag > 8s for 3 consecutive times: W = 1.5 If Δt_lag < 2s for 10 consecutive times: W = 0.5 Where w is the dynamically adjusted window width; Matching and retrieval are optimized by using a B+ tree index to store timestamps (retrieval complexity O(logn)), which meets the processing time limit of ≤10ms for IoT data; Spatial coordinate mapping, displacement tolerance threshold of 0.5m / s, exceeds the upper limit of safe movement speed for personnel in ISO13849-1 standard (0.3m / s), but is lower than the maximum operating speed of the elevator (1.0m / s).

[0036] Unstructured data processing unit: Domain-specific BERT model, property-specific dictionary injection:

[0037] Fine-tuning the data augmentation technique using an entity replacement generative adversarial network (ES-GAN): Generator; Change "Cooling pump bearing noise" to "Refrigeration pump bearing noise"; Discriminator; enforces semantic consistency (F1 score up to 0.93); Fault triple extraction and neighborhood optimization of BI-LSTM-CRF: Label set expansion, loss function improvement, and introduction of weight decay factor to address label imbalance: Loss = CE_loss + 0.1 × ||θ||² In the formula, Loss is the total loss function; CE_loss is the cross-entropy loss, which measures the difference between the probability distribution predicted by the model and the probability distribution of the true labels; θ is the set of model parameters. Physical constraints of dual-channel image analysis and industrial adaptation of the visible light channel:

[0038] Absolute calibration of thermal imaging channels: Temperature field correction formula: T_corrected=(T_raw-T_env) / ε+25 In the formula, T_corrected is the corrected temperature value; T_raw is the original measured temperature value; T_env is the measured value of the ambient temperature probe built into the infrared sensor, that is, the temperature of the environment around the sensor; ε is the emissivity of the material, which represents the ability of an object to emit infrared radiation. Different materials have different emissivity. For example, the emissivity of 316 stainless steel is 0.75 and the emissivity of copper is 0.05. Damage index calculation: D2 = Σ(>15℃) / S_total × Equipment aging coefficient In the formula, D2 is the damage index; ΔT is the temperature change, that is, the difference between the temperature at a certain measurement point and the reference temperature; ΔT is the total number of measurement points or the total number of samples. The distributed modeling framework specifically includes: Lightweight LSTM architecture for edge layer: Model compression, parameter control (total model weights less than 1MB), use depthwise separable convolutions instead of fully connected layers, weight pruning (removing weights less than 10). -4 Redundant parameters; 8-bit integer quantization, FP32 to INT8 conversion; The anomaly detection embedded in the physical mechanism includes input features such as extracting vibration signals through wavelet packet energy entropy and extracting temperature curves through first-order derivative mutation point detection. If the anomaly probability of the LSTM output is greater than a predetermined threshold, a physical constraint violation flag is determined, triggering an edge-level alarm and uploading the original data slice.

[0039] Cloud-based differential privacy gradient aggregator: Privacy protection is achieved through a noise injection model using Gaussian noise and a gradient pruning threshold to calculate the aggregated gradient. A cross-project knowledge fusion strategy is employed, with device similarity weighting, calculating topological similarity and weighted aggregation. Dynamic learning rate adjustment is implemented, where the learning rate is adjusted when the gradient variance of three consecutive aggregation rounds exceeds a predetermined threshold.

[0040] It should be noted that the edge layer uses a lightweight LSTM: Model compression and depthwise separable convolution replacement strategy: This replaces traditional fully connected layers with channel-separable convolution kernels (a combination of depthwise convolution and pointwise convolution). Parameter compression formula: Params_ratio=1 / N+1 / Dk² In the formula, N is the number of output channels, and Dk is the kernel size. 8-bit integer quantization, dynamic range calibration algorithm to solve activation value distribution offset, dynamic calculation of scaling factor, zero-point offset calibration; Physical threshold pruning, removing values ​​with an absolute value less than 10. -4 The weights are determined, and key physical feature connections are preserved (such as the forced preservation of vibration harmonic feature channels).

[0041] Anomaly detection based on physical mechanisms, multimodal feature extraction engine:

[0042] Two-factor alarm triggering mechanism: Condition 1: The probability of an anomaly in the LSTM output is greater than θ_edge (θ_edge = 0.92); Condition 2: Violation of physical constraints (e.g., vibration entropy increases but temperature remains constant); When both conditions are met, an edge alarm is triggered and a 10-second raw data slice is uploaded.

[0043] Cloud-based differential privacy aggregator: Privacy-accuracy balance equation, noise injection, aggregated gradient formula:

[0044] In the formula, This is an estimate of the aggregation gradient; Let S be the size of the set, that is, the number of elements in the set; To make gradient Clipping to the range [-C, C] to prevent gradient explosion; With a mean of 0 and a variance of The gradient follows a normal distribution, where I is the identity matrix and C is the pruning threshold used to retain 95% of the effective gradients. This represents the noise variance, used to meet the requirements of differential privacy. Key parameter settings are based on: clipping threshold C = 1.2 (retaining 95% of the effective gradient), noise variance σ = 0.5 (satisfying (1.2, 10)). -5 (-Differential privacy).

[0045] The decision execution interface specifically includes: Multi-level early warning triggering mechanism: A dynamic threshold for failure probability and a tiered response strategy are implemented, with a base threshold of 85%. Dynamic adjustment rules are in place: if the equipment's age exceeds 80% of its design life, the threshold is lowered to 75%; if the monthly failure frequency of similar equipment exceeds 3, the threshold is raised to 90%. Confidence check: When the predicted probability is within a certain range, the vibration spectrum verification is initiated. Inventory safety factor modeling and intelligent correction factor: if the supplier's on-time delivery rate is less than 95%, the threshold is raised to 1.5; if the equipment is under warranty, the threshold is lowered to 0.9.

[0046] Spare parts supply chain early warning: The three-level early warning system is as follows: Yellow warning: The ERP system automatically generates a purchase requisition form; Orange warning: The backup supplier bidding process is activated; Red warning: The cross-property project allocation agreement is triggered. Intelligent sourcing engine: supply network analysis, local warehouse search for available inventory within its radius; cloud collaboration: matching inventory status of other property projects with the same equipment model; cost optimization decision-making: controlling procurement costs; Dynamic dispatching of maintenance personnel, GIS road condition integration engine, real-time traffic data processing, access to Amap / Baidu real-time road condition API, and calculation of road impedance factor; path optimization algorithm, improvement of A Introduction of algorithm, time window constraint, skill matrix matching; Multi-objective scheduling model: The decision variables include total response time, labor cost, and service timeout risk; the personnel skills are greater than or equal to the fault repair requirement level.

[0047] It should be noted that the multi-level early warning trigger mechanism: Physical-statistical dual drive of dynamic threshold, equipment aging rate compensation formula:

[0048] In the formula, is the dynamic threshold, the finally calculated threshold, used to judge whether to trigger aging compensation or alarm; is the designed life of the equipment (elevator = 20 years, water pump = 15 years), and the lower limit is 75% to avoid overly sensitive triggering of false alarms; Adaptive rule for fault frequency:

[0049] Confidence level verification, vibration spectrum review trigger condition: If 75% < P_fault < 85%, it is in the confidence fuzzy interval, then extract the vibration signal in the 6 - 8 kHz frequency band; if the kurtosis coefficient > 4.0 and the envelope spectrum has the bearing fault characteristic frequency, then confirm the alarm; Risk modeling of inventory safety factor, supplier on-time delivery rate correction equation: S_threshold = 1.2 + 0.3×(95% - OTD_rate) In the formula, S_threshold is the supplier on-time delivery rate; OTD_rate is the historical on-time delivery rate of the supplier; Warranty period privilege design, threshold within the warranty period = 0.9 (to encourage the preferential use of original spare parts).

[0050] Spare parts supply chain early warning: Three-level response mechanism: Yellow warning (1.0 ≤ S < 1.2), automatically generate a purchase order and digital signature (in line with the Electronic Signature Law), trigger a 48-hour supplier confirmation time limit (automatic upgrade if overdue); Orange warning (0.6 ≤ S < 1.0), blockchain evidence storage bidding process, enable Monte Carlo tree search (MCTS) to evaluate supplier reliability; Red alert (S<0.6), cross-project allocation protocol enforces SSL / TLS encrypted transmission, prioritizes the use of the same model of equipment within a 50km radius (response time <2h).

[0051] Intelligent sourcing engine, supply network topology analysis, cost optimization model: min(Σ(transportation cost × urgency factor + inventory holding cost)) Urgency level: Yellow = 1.0, Orange = 2.0, Red = 3.0 Cloud-based collaboration utilizes an improved Gossip protocol to synchronize cross-project inventory (latency <5s), avoiding single points of failure in centralized servers.

[0052] Maintenance personnel dispatch: Real-time decision-making based on GIS traffic condition fusion, dynamic calculation of road resistance factors: R = 1 + 0.5 × (Congested Mileage / Total Mileage) + 0.3 × Accident Impact Factor Accident Factors: 1 = Minor Accident, 2 = Serious Accident (Real-time data obtained from Traffic Management Bureau API) Improvement A Algorithm: f(n) = g(n) + h(n) + α × max(0, ETA - T_window) In the formula, f(n) is the evaluation function value of node n, used to estimate the total cost from the starting node to the target node; g(n) is the actual cost from the starting node to node n; h(n) is the heuristic estimated cost from node n to the target node; max(0, ETA-T_window) is the time exceeding the expected arrival time if it exceeds the equipment downtime tolerance time T_window, otherwise it is 0; T_window is the equipment downtime tolerance time (elevator = 15min, distribution cabinet = 5min); α = 10 is the time window default penalty coefficient; Multi-objective scheduling model, decision variable constraint system:

[0053] Pareto solution set generation uses the NSGA-Ⅲ algorithm to output the optimal solution set within 3 seconds and binds it to federated learning logs for audit tracking.

[0054] Reference Figure 2 As shown, the predictive maintenance method for property equipment based on multi-source data fusion includes: Multi-source data is collected synchronously through a multi-protocol adapter, including ERP equipment ledgers, IoT real-time monitoring streams, and unstructured operation and maintenance records; spatiotemporal dynamic alignment is performed, and a transmission delay compensation model is established for low-frequency ERP coordinate data; synchronization with high-frequency IoT data is achieved through a sliding time window; Based on spatiotemporally dynamically aligned multi-source data, maintenance work order text processing is performed. A domain-fine-tuned BERT model is used to extract triple entity relationships, including equipment entities, fault types, and root causes. Semantic correction is performed by embedding the ISO13379-2023 fault tree rule base. Fault image analysis is performed, including spatial attention convolutional networks in the visible light channel to quantify the crack length ratio and abnormal areas in the thermal imaging channel with temperatures exceeding a predetermined proportion of the rated value. The results are then fused and output. A lightweight LSTM model is deployed at the edge layer to perform real-time anomaly detection; cross-project model parameters are fused in the cloud through a differential privacy gradient aggregator. When the failure probability is greater than the dynamic threshold and the inventory safety factor is less than the predetermined threshold, the three-level early warning of the spare parts supply chain is automatically triggered; maintenance personnel are dispatched, and the route is optimized by integrating real-time GIS traffic conditions.

[0055] It should be noted that heterogeneous data alignment uses a bimodal Gaussian delay compensation model (main peak). =1.2s, secondary peak =3.1s), eliminating the spatiotemporal deviation between ERP and IoT data (≤0.3 meters); adaptive sliding window dynamically adjusts the window width (expanding by 150% when the delay is >8s), ensuring that the synchronization delay of 100Hz high-frequency stream data is <10ms; solving industry problems such as coordinate drift and signal loss of synchronization in property scenarios, and providing spatiotemporal consistency guarantee for model input.

[0056] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A property equipment intelligent inspection and precise predictive maintenance system based on multi-source data fusion, characterized in that, include: Multi-source data access module: Built-in OPCUA / Modbus / HTTP2 multi-protocol adapter, synchronously accessing three types of heterogeneous data sources, including ERP equipment ledger data, IoT real-time monitoring streams, and unstructured operation and maintenance records; Dynamic fusion engine: performs asynchronous data spatiotemporal alignment at runtime, dynamically compensates for the transmission delay between ERP coordinate data and IoT real-time coordinates through a sliding time window; unstructured data processing unit: uses a domain-fine-tuned BERT model to parse maintenance work order text, outputs standardized fault triples, and identifies surface cracks and thermal anomaly areas in fault images based on a dual-channel attention convolutional network. Distributed modeling framework: Lightweight LSTM models are deployed at the edge layer for real-time anomaly detection; in the cloud, differential privacy gradient aggregator is used to fuse model parameters from across property projects. Decision execution interface: When the probability of equipment failure is greater than a predetermined proportion and the safety factor of spare parts inventory is less than a predetermined threshold, the spare parts supply chain early warning instruction is automatically triggered, and the dynamic scheduling scheme of maintenance personnel is integrated with GIS road conditions.

2. The intelligent inspection and precise predictive maintenance system for property equipment based on multi-source data fusion as described in claim 1, characterized in that, The dynamic fusion engine specifically includes: Asynchronous data time-space alignment: Dynamic compensation for transmission delay is achieved by using a timestamp correction algorithm to dynamically compensate for the original timestamps of the ERP coordinate data; sliding time window synchronization control is implemented with adaptive adjustment of the window size, and when the high-frequency IoT data stream arrives, matching ERP coordinate data is retrieved and output within the window; The spatial coordinate mapping engine completes coordinate system transformation based on the transformation model. The ERP system uses the WGS-84 geographic coordinate system, while the IoT positioning data uses a local rectangular coordinate system. Equipment displacement tolerance control triggers position anomaly verification when the rate of change of adjacent coordinates is greater than 0.5 m / s.

3. The intelligent inspection and precise predictive maintenance system for property equipment based on multi-source data fusion as described in claim 2, characterized in that, The dynamic fusion engine specifically includes: Unstructured data processing unit: Maintenance work order text parsing: Domain-adaptive BERT model, where the base model is BERT-Large, and the fine-tuning dataset consists of property equipment failure work orders, covering elevator / HVAC / power supply systems; Fault triplet extraction and entity recognition are performed. Equipment components / fault phenomena are labeled using BI-LSTM-CRF layers. The standardized output structure includes equipment entity, fault type, and root cause. Dual-channel analysis of fault images: Visible light channel processing path, network structure is ResNet-50+ spatial attention module; crack detection process is input image, adaptive histogram equalization, edge detection, attention weight map generation, output crack pixel ratio; thermal imaging channel processing path, high temperature area is extracted through temperature field analysis algorithm; thermal anomaly area ratio is calculated.

4. The intelligent inspection and precise predictive maintenance system for property equipment based on multi-source data fusion as described in claim 1, characterized in that, The multi-source data access module specifically includes: Multi-protocol adapter architecture: The underlying communication protocol stack adopts a layered design to achieve parallel support for OPCUA, Modbus, and HTTP / 2 protocols. The physical layer supports adaptive switching between RS-485 / Ethernet / 5G wireless transmission media; the protocol conversion layer converts Modbus RTU messages into the OPCUA information model; and the security layer enables TLS1.3 encryption and two-way certificate authentication for HTTP / 2 connections. Three types of data source access: ERP equipment ledger data access: structured relational data, with key fields including equipment ID, 3D spatial coordinates, warranty expiration date, and historical maintenance count; extraction mechanism, monitoring database binlog changes via HTTP / 2 long connection; incremental update frequency, triggering synchronization whenever the number of record changes exceeds a predetermined threshold; IoT real-time monitoring stream access: high-frequency timing signals, including vibration acceleration, winding temperature, and three-phase current; acquisition control logic: pre-configured acquisition strategy at the edge, normal sampling when the device is in operation, and accelerated sampling when the temperature is higher than the predetermined temperature; Unstructured maintenance records are accessed: maintenance work order text, which comes from JSON messages submitted by scanning a QR code on the mobile APP. Key fields include fault description text, urgency level label, and reporting personnel ID; fault images support formats including JPEG / PNG / infrared thermal imaging DICOM; voice logs are collected on-site through a smart safety helmet, with 16kHz sampling rate single-channel PCM audio.

5. The intelligent inspection and precise predictive maintenance system for property equipment based on multi-source data fusion as described in claim 4, characterized in that, The multi-source data access module specifically includes: Synchronous access control mechanism: Time synchronization is guaranteed by using the IEEE 1588 precision time protocol to timestamp all data; resource conflict resolution is achieved by arbitrating based on priority when multiple data sources access the system concurrently, with priority in the following order: emergency fault alarm, real-time monitoring stream, maintenance record update, and ledger data synchronization. It supports resuming data transmission after network outages, with a local cache queue capacity that supports storing 72 hours of raw data. After the network is restored, it can resume transmission from the point of interruption through hash verification. Data preprocessing pipeline: Structured data cleaning, ERP coordinate data verification, and automatic mapping to the nearest floor if the z-coordinate exceeds the building height range; streaming data denoising, and wavelet threshold denoising for vibration signals; unstructured data encapsulation, binding maintenance work order text and associated images into a unified event ID, and storing it in Avro binary format.

6. The intelligent inspection and precise predictive maintenance system for property equipment based on multi-source data fusion according to claim 1, characterized in that, The distributed modeling framework specifically includes: Lightweight LSTM architecture for edge layer: Model compression, parameter control, total model weights less than 1MB, depthwise separable convolutions instead of fully connected layers, weight pruning, removal of redundant parameters below a predetermined threshold; 8-bit integer quantization, FP32 to INT8 conversion; The anomaly detection embedded in the physical mechanism includes input features such as extracting vibration signals through wavelet packet energy entropy and extracting temperature curves through first-order derivative mutation point detection. If the anomaly probability of the LSTM output is greater than a predetermined threshold, a physical constraint violation flag is determined, triggering an edge-level alarm and uploading the original data slice.

7. The intelligent inspection and precise predictive maintenance system for property equipment based on multi-source data fusion according to claim 6, characterized in that, The distributed modeling framework specifically includes: Cloud-based differential privacy gradient aggregator: Privacy protection is achieved through a noise injection model using Gaussian noise and a gradient pruning threshold to calculate the aggregated gradient. A cross-project knowledge fusion strategy is employed, with device similarity weighting, calculating topological similarity and weighted aggregation. Dynamic learning rate adjustment is implemented, where the learning rate is adjusted when the gradient variance of three consecutive aggregation rounds exceeds a predetermined threshold.

8. The intelligent inspection and precise predictive maintenance system and method for property equipment based on multi-source data fusion according to claim 1, characterized in that, The decision execution interface specifically includes: Multi-level early warning triggering mechanism: A dynamic threshold for failure probability and a tiered response strategy are implemented, with a base threshold of 85%. Dynamic adjustment rules are in place: if the equipment's age exceeds 80% of its design life, the threshold is lowered to 75%; if the monthly failure frequency of similar equipment exceeds 3, the threshold is raised to 90%. Confidence check: When the predicted probability is within a certain range, the vibration spectrum verification is initiated. Inventory safety factor modeling and intelligent correction factor: if the supplier's on-time delivery rate is less than 95%, the threshold is raised to 1.5; if the equipment is under warranty, the threshold is lowered to 0.

9.

9. The intelligent inspection and precise predictive maintenance system and method for property equipment based on multi-source data fusion according to claim 8, characterized in that, The decision execution interface specifically includes: Spare parts supply chain early warning: The three-level early warning system is as follows: Yellow warning: The ERP system automatically generates a purchase requisition form; Orange warning: The backup supplier bidding process is activated; Red warning: The cross-property project allocation agreement is triggered. Intelligent sourcing engine: supply network analysis, local warehouse search for available inventory within its radius; cloud collaboration: matching inventory status of other property projects with the same equipment model; cost optimization decision-making: controlling procurement costs; Dynamic scheduling of maintenance personnel, GIS traffic fusion engine, real-time traffic data processing, integration with Gaode / Baidu real-time traffic APIs, and calculation of road resistance factors; route optimization algorithm, improved A Algorithm introduction, time window constraints, skill matrix matching; Multi-objective scheduling model: Decision variables include total response time, manpower cost, and service timeout risk; personnel skills are greater than or equal to the fault repair requirement level.

10. A method for intelligent inspection and precise predictive maintenance of property equipment based on multi-source data fusion, characterized in that, include: Multi-source data is collected synchronously through a multi-protocol adapter, including ERP equipment ledgers, IoT real-time monitoring streams, and unstructured operation and maintenance records; Perform spatiotemporal dynamic alignment and establish a transmission delay compensation model for low-frequency ERP coordinate data; achieve synchronization with high-frequency IoT data through a sliding time window. Based on spatiotemporally dynamically aligned multi-source data, maintenance work order text processing is performed. A domain-fine-tuned BERT model is used to extract triple entity relationships, including equipment entities, fault types, and root causes. Semantic correction is performed by embedding the ISO13379-2023 fault tree rule base. Fault image analysis is performed, including spatial attention convolutional networks in the visible light channel to quantify the crack length ratio and abnormal areas in the thermal imaging channel with temperatures exceeding a predetermined proportion of the rated value. The results are then fused and output. A lightweight LSTM model is deployed at the edge layer to perform real-time anomaly detection; cross-project model parameters are fused in the cloud through a differential privacy gradient aggregator. When the failure probability is greater than the dynamic threshold and the inventory safety factor is less than the predetermined threshold, the spare parts supply chain level three early warning will be automatically triggered. Maintenance personnel dispatching, integrating real-time traffic information from GIS to optimize routes.

Images