Distributed non-inductive intelligent operation and maintenance and resource value-added method based on old benefit recycling

Abstract

This invention provides a distributed, seamless intelligent operation and maintenance and resource value-added method based on reuse, comprising: reusing existing cameras through edge adapters, using a built-in deep network super-resolution reconstruction module to enhance the original video stream in real time, generating a high-resolution standardized bitstream; performing spatiotemporal context analysis on the standardized bitstream to extract structured features from the ambient diffuse light, performing perception recognition on the original video stream, and generating lightweight feature data; constructing a graph structure containing structured features and inputting it into a multi-layer graph convolutional network to generate feature vectors for each unit region for dynamic baseline comparison, marking regions with differences exceeding a dynamic threshold as abnormal regions for selective unsharpened mask compensation, and outputting operation and maintenance early warning signals; cross-validating the lightweight feature data with the application data from external business systems to form a business early warning closed loop, and simultaneously encapsulating the extracted structured features into data assets according to a preset catalog.

Description

Technical Field

[0001] This invention relates to an administrative supervision technology, and in particular to a distributed, seamless intelligent operation and maintenance and resource value-added method based on the reuse of existing resources. Background Technology

[0002] With the rapid development of smart cities and the Industrial Internet, automated operation and maintenance management using video surveillance systems has become mainstream. Current video operation and maintenance solutions typically rely on high-specification hardware investment, acquiring structured feature data by deploying high-resolution sensors, adding active lighting systems, or requiring monitored objects to wear specific sensors (such as RFID, smart bracelets, etc.). In terms of data processing, traditional technologies mostly adopt a full-stream video transmission mode, directly pushing raw data to the cloud for centralized identification and processing, and issuing anomaly warnings based on fixed thresholds. Improving monitoring capabilities requires replacing existing low-resolution cameras or rewiring, resulting in poor "reuse" capabilities and significant resource waste. Existing identification solutions require active lighting in low-light or complex environments, increasing energy consumption and potentially causing light pollution. Furthermore, the mandatory requirement to wear devices limits the flexibility of application scenarios. Summary of the Invention

[0003] The purpose of this invention is to provide a distributed, seamless intelligent operation and maintenance and resource value-added method based on reuse, comprising:

[0004] Step S100: The existing camera is reused through the edge adapter. Without changing the original wiring structure, the original video stream is acquired and enhanced in real time using the built-in deep network super-resolution reconstruction module to generate a high-resolution standardized bitstream.

[0005] Step S200: Without adding an active supplementary light source and without relying on the wearable device of the test object, perform spatiotemporal context analysis on the standardized bitstream, extract the structured features in the ambient diffuse light, perform perception recognition on the original video stream, and generate lightweight feature data.

[0006] Step S300: Construct a graph structure containing structured features and input it into a multi-layer graph convolutional network to generate feature vectors for each unit region for dynamic baseline comparison. Regions with differences exceeding the dynamic threshold are marked as abnormal regions and selective unsharpened mask compensation is performed. Operation and maintenance early warning signals are output.

[0007] Step S400 involves cross-validating the lightweight feature data with the application data from external business systems to form a business early warning closed loop. Simultaneously, the extracted structured features are packaged into data assets according to a preset catalog and output to the application platform.

[0008] Furthermore, step S200 specifically includes:

[0009] Step S210: Receive the high-resolution normalized bitstream output by the edge adapter, and use the image segmentation unit to segment the single frame image into several unit regions in the spatial dimension.

[0010] Step S220: For each segmented unit region, without turning on the active supplementary light source, the structured features formed by the ambient diffuse reflection light in that region are extracted by the gray-scale feature management unit and the texture feature management unit respectively. The structured features include gray-scale feature values ​​and texture feature values.

[0011] Step S230: Concatenate the grayscale feature values ​​and texture feature values ​​dimensionally to generate the region feature vector corresponding to each unit region;

[0012] Step S240: Calculate the Euclidean distance between the centroids of adjacent unit regions, and mark the regions with a distance less than a preset distance threshold as neighborhoods; traverse the neighborhoods to calculate the gray-level gradient, and record the unit region corresponding to the maximum gray-level gradient as the adjacent region, while defining the maximum gradient value as the edge weight.

[0013] Step S250: Collect all region feature vectors to form a node feature matrix, and collect all edge weights to construct an adjacency matrix.

[0014] Step S260: Use the spatiotemporal context association algorithm to perform cross-frame and cross-device feature matching on the graph structure data, perform cross-device weak association moving target re-identification, and output cross-device trajectory association features;

[0015] Step S270: The node feature matrix and trajectory association features are fused to generate lightweight feature data.

[0016] Furthermore, the specific process of step S260 includes:

[0017] Step S261: Using the spatiotemporal context association algorithm, the structured features of the current frame of the same sensing device and the historical feature baseline of the previous frame are combined to perform correlation calculation and obtain a multimodal spatiotemporal context feature vector with temporal continuity and spatial semantic information.

[0018] Step S262: The structured features of the current frame are used as the target to be searched. Based on the physical distribution relationship between the current device and other devices in the local area network, the logical offset time from the target moving from the neighboring device to the current device is calculated. Based on the offset time, feature data of other devices within the corresponding time period are called to construct the target spatiotemporal sliding window reference baseline pool. The similarity between the target feature sequence and each candidate feature vector in the target spatiotemporal sliding window reference baseline pool in the time domain is obtained by normalized cross-correlation calculation, resulting in a set of correlation coefficient value sequences distributed over time. If the frequency of correlation coefficient values ​​exceeding the preset confidence threshold within the time step reaches a preset proportion, the system confirms that the target detected by the current device and the target in the reference baseline pool are the same physical entity, completing the cross-device identity alignment and outputting cross-device trajectory association features.

[0019] Step S263: Obtain the weak association re-identification result signal and output the cross-camera trajectory association features; the weak association re-identification result signal includes the target unique identifier, state measure and confidence score; the state measure includes the target's real-time motion vector; the confidence score indicates the degree of certainty in determining the target's identity.

[0020] Furthermore, the confidence score acquisition process in step S263 includes:

[0021] Step S2631: Map the correlation coefficient values ​​exceeding the preset confidence threshold from step S262 to the [0,1] interval, and use them as the initial visual confidence score S. vision ;

[0022] Step S2632: Obtain the target's trajectory using the Kalman filter algorithm; calculate the overlap G between the target's current trajectory and the predicted trajectory using the motion direction prediction algorithm. motion ;

[0023] Step S2633: Calculate the logical offset time G of the target appearance based on the topology of the sensing devices. topology ;

[0024] Step S2634: The final execution score (Stotal) is obtained using a non-linear weighted summation algorithm.

[0025] Stotal=αS vision +βG motion +γG topology ,

[0026] Where α, β, and γ are weights, respectively.

[0027] Furthermore, step S300 specifically includes:

[0028] Step S310: Receive the node feature matrix and adjacency matrix as input, and use the spatial adjacency relationship and edge weight defined by the adjacency matrix to construct a graph structure data model representing the spatiotemporal context relationship of the current video frame.

[0029] Step S320: The graph structure data model is input to a pre-defined multi-layer graph convolutional network (GCN) for forward propagation calculation, and the output is a unit region evolution feature vector with high-order semantic enhancement.

[0030] Step S330: Using the unit region evolution feature vector as the first input and the historical feature baseline of the previous frame as the second input, the spatial deviation component between the evolution feature vector and the historical feature baseline is calculated using the second-order norm distance as the anomaly intensity. When the anomaly intensity exceeds the preset dynamic threshold, the corresponding coordinate space is marked as an anomaly region. The anomaly area ratio is obtained based on the ratio of the number of anomaly regions to the total number of unit regions N.

[0031] Step S340: Using the coordinates of the locked abnormal region as input, perform selective unsharpened mask compensation on the corresponding local image block in the original bitstream.

[0032] Step S350: Construct a time-domain status observation window backtracking. If the number of consecutive frames in an abnormal region exceeds a preset warning trigger threshold, an operation and maintenance warning is triggered; calculate the area change rate ΔP=P between adjacent frames. t -P t-1 And the area variance σ within the window P 2 If there are k consecutive frames whose area change rate ΔP is greater than a preset surge threshold θ, then... jump Furthermore, within the consecutive n frames after ΔP approaches 0, the area variance σ P 2 Less than the preset stability threshold If the abnormal intensity within the window is not found, it is determined to be an environmental mutation; if the slope k' of the first-order linear regression fitting line satisfies 0 < k' < θ, then it is considered an environmental mutation. slow θ slow If the threshold for slow growth is reached and the fitted correlation coefficient is greater than the preset confidence threshold, then the equipment is considered to be in a state of degradation.

[0033] Furthermore, step S310 specifically includes:

[0034] Step S311: Assign a globally unique spatial index number to each unit region, and use the relative positions of the center coordinates of each region to establish an initial zero-weight complete graph in the computation space.

[0035] Step S312: Sequentially fill each region feature vector into the corresponding index node of the zero-weight complete graph to generate the initial node feature matrix V;

[0036] Step S313: Obtain the center coordinate sequence P={p1,p2,...,p} for each unit region. n} and preset distance threshold D th For any two nodes in a zero-weight complete graph, calculate the Euclidean distance d between their centers. ij The Euclidean distance is compared with a preset distance threshold; if d ij <D th If the j-th node belongs to the logical neighborhood of the i-th node, then a binary connectivity matrix B is generated to describe the connectivity of the graph structure. If i and j satisfy the logical neighborhood relationship, then B... ij =1, otherwise B ij =0;

[0037] Step S314, for B in the binary connection matrix B ij For elements with a value of 1, retrieve the corresponding node U. i with U j Locate the adjacent edge regions of the two regions in the original image and set them as the sampling range for gradient calculation; traverse all pixel pairs of the adjacent edges between node i and node j, calculate their gray-level differences, and extract the maximum gray-level difference as the local maximum gray-level gradient value G corresponding to that specific node pair. max (i,j), using the Gaussian kernel function to extract the local maximum gray-level gradient value G. max (i,j) is mapped to the edge weight W connecting nodes i and j. ij ; Set the edge weight W ij Replace the 1 value in the binary connectivity matrix to obtain the adjacency matrix W;

[0038] Step S315: Encapsulate the node feature matrix V and the adjacency matrix W into a graph structure data model G.

[0039] Furthermore, step S400 specifically includes:

[0040] Step S410: If the corresponding unit area does not trigger an operation and maintenance warning, the feature data is determined to be high-confidence valid data. The irreversible hash algorithm operator is called to de-identify the sensitive attributes in the high-confidence data and generate a highly reliable de-identified feature set.

[0041] Step S420: Using a spatiotemporal consistency discrimination algorithm, the target behavior trajectory and occurrence time in the high-reliability desensitization feature set are matched with the reservation information in the application data; if the two are inconsistent within a preset logical threshold, an anomaly judgment is triggered at the business level, and a handling instruction is generated.

[0042] Step S430: Package the business-level anomaly judgment, handling instructions, and high-reliability de-identified feature set according to a preset format to generate a structured value-added data asset package.

[0043] Step S440: Through the service subscription distribution mechanism, the packaged data assets are pushed to the third-party application platform.

[0044] Furthermore, step S420 specifically includes:

[0045] Step S421: Extract the sampling timestamp and spatial index coordinates from the desensitized features, and simultaneously extract the reservation time interval and preset physical location from the declaration data; use the time synchronization operator to map the two to the same standard time reference system to construct a unified spatiotemporal comparison benchmark.

[0046] Step S422: Chain-connect the desensitized features with the same characteristic attributes that appear at different time points to generate the target behavior trajectory;

[0047] Step S423: Invoke the spatiotemporal consistency discrimination algorithm to calculate the dynamic time warping distance between the measured target behavior trajectory and the preset path in the reservation information, and compare whether the measured occurrence time is within the reservation time interval; if the dynamic time warping distance is greater than the preset spatial deviation threshold, or the measured occurrence time deviates from the reservation time interval, it is determined that the two are inconsistent, triggering the business-level anomaly judgment.

[0048] Step S424: Retrieve the corresponding response strategy preset library based on the dynamic time warping distance to generate a processing instruction.

[0049] Compared with existing technologies, this invention has the following advantages: (1) It directly reuses existing equipment through edge adapters and uses deep network super-resolution reconstruction technology (DLSR) to achieve real-time enhancement of image quality without changing the wiring, which greatly extends the service life of old equipment; (2) It extracts structured features (grayscale, texture, etc.) in ambient diffuse light and achieves accurate recognition without adding supplementary lights or relying on wearable devices, reducing the intrusiveness of the system to the physical environment; (3) It adopts a perception recognition strategy to convert the original video stream into lightweight feature data, reducing the data volume to less than 5% of the original bitstream, which significantly reduces bandwidth pressure and storage costs; (4) It introduces multi-layer graph convolutional network (GCN) to perform spatial topology modeling of unit areas and combines it with temporal state observation window for dynamic baseline comparison, which can accurately distinguish between "environmental mutation" and "device state degradation", effectively reducing the false alarm rate; (5) It realizes cross-device identity alignment and trajectory solidification through spatiotemporal context association algorithm and logical offset time calculation, which solves the problem of target loss in weak association environment of traditional schemes.

[0050] The present invention will now be further described with reference to the accompanying drawings. Attached Figure Description

[0051] Figure 1This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0052] Combination Figure 1 A distributed, seamless intelligent operation and maintenance and resource value-added method based on reuse and recycling, characterized by comprising:

[0053] Step S100, Lightweight reuse access and bitstream reconstruction: By reusing existing sensing devices through edge adapters, the original video stream of the sensing devices is collected without changing the original wiring structure, and real-time enhancement is performed using built-in or external deep network super-resolution reconstruction modules to generate a high-resolution standardized bitstream.

[0054] Step S200, Non-contact feature extraction and lightweight compression: Without adding an active supplementary light source and without relying on the wearable device of the test object, perform spatiotemporal context analysis on the standardized bitstream, extract the structured features in the ambient diffuse light, perform perception recognition on the original video stream, and generate lightweight feature data.

[0055] Step S300, Graph Convolutional Anomaly Detection and Selective Sharpening: Construct a graph structure containing structured features and input it into a multi-layer graph convolutional network to generate feature vectors for each unit region for dynamic baseline comparison. Regions with differences exceeding the dynamic threshold are marked as abnormal regions and selective unsharpened mask compensation is performed to output an operation and maintenance early warning signal.

[0056] Step S400, Multi-source data fusion and data asset value-added: Cross-validate lightweight feature data with the declaration data of external business systems to form a business early warning closed loop. At the same time, encapsulate the extracted structured features into data assets according to a preset catalog and output them to the application platform.

[0057] In step S100, the Deep Learning-based Super-Resolution (DLSR) module is a technique that uses deep convolutional neural networks (CNNs) or generative adversarial networks (GANs) to upscale, repair, and reconstruct low-resolution (LR) images / videos into high-resolution (HR) images.

[0058] The specific process of step S200 includes:

[0059] Step S210: Spatiotemporal feature domain initialization and unit region segmentation;

[0060] Step S220: Extraction of original features under ambient diffuse light;

[0061] Step S230: Synthesis of region feature vectors;

[0062] Step S240: Adjacency relationship establishment and edge weight calculation;

[0063] Step S250: Graph structure modeling and feature matrix construction;

[0064] Step S260: Cross-camera weak correlation re-identification and trajectory solidification;

[0065] Step S270: Perception recognition and lightweight feature data generation.

[0066] Step S210, Spatiotemporal Feature Domain Initialization and Unit Region Segmentation: Receive the high-resolution normalized bitstream output from the edge adapter, and use the image segmentation unit to segment the single-frame image into N unit regions in the spatial dimension, serving as the basic spatial index for subsequent feature extraction. Here, the unit region refers to the smallest computational unit after the original image is divided into a grid according to a preset pixel size (e.g., 32×32 pixels).

[0067] Step S220, Original Feature Extraction under Ambient Diffuse Light: For each segmented unit region, without enabling active supplementary light sources, the structured features formed by ambient diffuse light within that region are extracted using the grayscale feature management unit and the texture feature management unit, respectively. These features are grayscale feature values ​​and texture feature values. The grayscale feature values ​​are obtained through the Gray-Level Co-occurrence Matrix (GLCM), and the texture feature values ​​are obtained through the Local Binary Pattern (LBP) algorithm.

[0068] Step S230, Region Feature Vector Synthesis: The grayscale feature values ​​and texture feature values ​​are concatenated dimensionally to generate the region feature vector corresponding to each unit region. This vector serves as the underlying data representation describing the structured information of the unit region under ambient light.

[0069] Step S240, Adjacency relationship establishment and edge weight calculation: Calculate the Euclidean distance between the centroids of adjacent unit regions, and mark the regions with a distance less than a preset distance threshold as neighborhoods; traverse the neighborhoods to calculate the gray-level gradient, and record the unit region corresponding to the maximum gray-level gradient as the adjacent region, and define the maximum gradient value as the edge weight to characterize the correlation strength of the spatial context.

[0070] Step S250, Graph structure modeling and feature matrix construction: Collect all region feature vectors to form a node feature matrix, and collect all edge weights to construct an adjacency matrix; the node feature matrix and the adjacency matrix together constitute a graph structure data model describing the spatiotemporal context relationship of the current video frame.

[0071] Step S260, Cross-camera weak association re-identification and trajectory solidification: The spatiotemporal context association algorithm is used to perform cross-frame and cross-camera feature matching on the graph structure data, perform cross-camera weak association moving target re-identification, and output cross-camera trajectory association features.

[0072] Step S270, Perception and Recognition and Lightweight Feature Data Generation: The node feature matrix and trajectory association features are fused to form structured metadata after perception and recognition of the original video stream, i.e., lightweight feature data is generated; the amount of lightweight feature data is less than 5% of the original video stream, and it serves as the core input for subsequent intelligent decision-making and data asset transformation.

[0073] The specific process of step S260 includes:

[0074] Step S261: Generate multimodal spatiotemporal context feature vectors;

[0075] Step S262: Perform cross-device normalized cross-correlation calculation;

[0076] Step S263: Weak correlation re-identification result signal, output cross-camera trajectory correlation features.

[0077] Step S261: Generate a multimodal spatiotemporal context feature vector: Using the Spatio-Temporal Context Correlation Algorithm, the structured features of the current frame from the same sensing device are combined with the historical feature baseline of the preceding frames to perform correlation calculation, thereby obtaining a multimodal spatiotemporal context feature vector with temporal continuity and spatial semantic information. In this embodiment, the historical feature baseline of the preceding frames refers to a statistical reference vector generated by a weighted moving average algorithm based on the structured features of the preceding N frames in the time series. The correlation calculation in this embodiment uses second-order norm distance calculation in the spatial dimension and normalized cross-correlation calculation in the temporal dimension. In the second-order norm distance calculation, the deviation value of each unit region is determined by calculating the second-order norm distance between the current frame's structured feature vector and the historical feature baseline vector, thus determining whether the device has degraded or the image is abnormal.

[0078] Step S262: Perform cross-device normalized cross-correlation calculation: On the spatial axis, compare the multimodal spatiotemporal context feature vector of a sensor with the real-time feature pool of other sensor devices in the local area network. Using the structured features of the current frame as the target to be searched, the system calculates the logical offset time from the target moving from a neighboring device to the current device based on the physical distribution relationship between the current device and other devices in the local area network, i.e., the topological distance. Based on this offset time, feature data within the corresponding time period is extracted from the real-time feature pool cached by the middle platform to construct a target spatiotemporal sliding window reference baseline pool. The system performs normalized cross-correlation calculation between the target feature sequence and each candidate feature vector in the target spatiotemporal sliding window reference baseline pool, calculating the similarity in the time domain between feature vectors generated by two different devices, obtaining a set of correlation coefficient values ​​distributed over time. If the frequency of correlation coefficient values ​​exceeding a preset confidence threshold within the time step reaches a preset proportion, the system confirms that the target detected by the current device and the target in the reference baseline pool are the same physical entity, completing cross-device identity alignment and outputting cross-device trajectory association features.

[0079] Step S263: Obtain the weak association re-identification result signal and output cross-camera trajectory association features: The weak association re-identification result signal is essentially a structured data packet, a set of digital labels and confidence weights that can uniquely identify the target. The weak association re-identification result signal includes a unique identifier, a state measure, and a confidence score. A globally unique number is assigned to targets of the same physical entity. The state measure includes the target's real-time motion vector and the logical state of the pedestrian skeleton points. Confidence score: Due to the weak association, this signal carries a confidence value, representing the degree of certainty with which the system determines the target's identity based on spatiotemporal logic. The weak association re-identification result signal is aggregated with the geographic coordinate information and timestamps of each camera, mapped to a unified spatiotemporal coordinate system, and the target's movement path between different monitoring areas is fitted to finally form and output cross-camera trajectory association features.

[0080] The confidence score acquisition process includes:

[0081] Step S2631: Map the correlation coefficient values ​​exceeding the preset confidence threshold from step S262 to the [0,1] interval, and use them as the initial visual confidence score S. vision ;

[0082] Step S2632: Obtain the target's trajectory using the Kalman filter algorithm; calculate the overlap G between the target's current trajectory and the predicted trajectory using the motion direction prediction algorithm. motion ;

[0083] Step S2633: Calculate the logical offset time G of the target appearance based on the topology of the sensing devices. topology ;

[0084] Step S2634: The final execution score (Stotal) is obtained using a non-linear weighted summation algorithm.

[0085] Stotal=αS vision +βG motion +γG topology ,

[0086] Where α, β, and γ are weights, respectively.

[0087] Step S300 specifically includes:

[0088] Step S310: Construct a cross-domain spatiotemporal feature map structure model;

[0089] Step S320: Perform multi-layer graph convolution feature evolution;

[0090] Step S330: Perform dynamic baseline comparison and anomaly region locking;

[0091] Step S340: Perform selective unsharpened mask compensation;

[0092] Step S350: Generate and output maintenance early warning signals.

[0093] Step S310: Construct a cross-domain spatiotemporal feature graph structure model: Receive node feature matrices and adjacency matrices as input; use the regional feature vectors of each unit region as nodes, and utilize the spatial adjacency relationships and edge weights defined by the adjacency matrix to construct a graph structure data model representing the spatiotemporal context of the current video frame. This model transforms isolated local features into global feature representations with topological relationships.

[0094] The specific method for obtaining the graph structure data model in step S310 includes:

[0095] Step S311: Assign a globally unique spatial index number to each unit region, and use the relative position of the center coordinates of each region to establish an initial zero-weight complete graph in the computation space to carry the node features and connection relationships generated subsequently.

[0096] Step S312: Sequentially fill the corresponding index nodes of the zero-weight complete graph with the feature vector of each region to generate the initial node feature matrix V. A node is represented as a unit region. Mathematically, this matrix transforms the physical information of the unit region, such as grayscale and texture, into the original state attributes of the nodes in the graph.

[0097] Step S313: Obtain the center coordinate sequence P={p1,p2,...,p} for each unit region. n} and preset distance threshold D th For any two nodes in a zero-weight complete graph, calculate the Euclidean distance d between their centers.ij The Euclidean distance is compared with a preset distance threshold; if d ij <D th If the j-th node belongs to the logical neighborhood of the i-th node, then a binary connectivity matrix B is generated to describe the connectivity of the graph structure. If i and j satisfy the logical neighborhood relationship, then B... ij =1, otherwise B ij =0;

[0098] Step S314, for B in the binary connection matrix B ij For elements with a value of 1, retrieve the corresponding node U. i with U j Locate the adjacent edge regions of the two regions in the original image and set them as the sampling range for gradient calculation; traverse all pixel pairs of the adjacent edges between node i and node j, calculate their gray-level differences, and extract the maximum gray-level difference as the local maximum gray-level gradient value G corresponding to that specific node pair. max (i,j), using the Gaussian kernel function to extract the local maximum gray-level gradient value G. max (i,j) is mapped to the edge weight W connecting nodes i and j. ij ; Set the edge weight W ij Replace the 1 value in the binary connectivity matrix to obtain the adjacency matrix W;

[0099] Step S315: Encapsulate the node feature matrix V and the adjacency matrix W into a graph structure data model G.

[0100] Step S320: Perform multi-layer graph convolutional feature evolution: Using the graph structure data model as input, input it into a pre-defined multi-layer graph convolutional network (GCN) for forward propagation computation. A multi-layer graph convolutional network (GCN) is a deep learning model capable of performing convolution operations directly on a graph structure. It achieves deep fusion of node features and their neighborhood features through Laplacian matrix transformation; it is obtained through supervised learning on a training set pre-labeled with "device normal / abnormal" samples; GCN uses convolutional kernels to aggregate and transform node features, outputting a unit region evolution feature vector with high-order semantic enhancement.

[0101] Step S330: Perform dynamic baseline comparison and anomaly region locking. Using the unit region evolution feature vector output in step S320 as the first input and the historical feature baseline of the previous frame as the second input, calculate the spatial deviation component between the evolution feature vector and the historical feature baseline using second-order norm distance calculation, i.e., the anomaly intensity. When the deviation component exceeds a preset dynamic threshold, the corresponding coordinate space is marked as an anomaly region. Obtain the anomaly area percentage based on the ratio of the number of anomaly regions to the total number of unit regions N. The dynamic threshold is a judgment boundary that automatically adjusts according to ambient light intensity and historical noise distribution; it is calculated by statistically analyzing the mean and variance of the second-order norm distance of the past N frames and combining it with a preset sensitivity coefficient.

[0102] Step S340: Perform Selective Unsharp Mask Compensation: Using the coordinates of the locked anomalous region as input, selective unsharp mask compensation (USM) is performed on the corresponding local image blocks in the original bitstream. A local image block refers to a local pixel matrix segmented from a single frame image corresponding to the normalized bitstream, perfectly aligned with the anomalous grid coordinates; its pixel dimensions are completely consistent with the unit region. Selective unsharp mask compensation (USM) is an image enhancement algorithm that targets only specific local regions. It extracts high-frequency details by subtracting a blurred copy (low-pass filtering) from the original image and proportionally superimposing it back into the original image; this operation is performed only within the grid cells marked as anomalous. By enhancing the edge contrast of the anomalous region, it compensates for feature blurring caused by backlighting or insufficient resolution, providing clear image support for subsequent accurate identification.

[0103] Step S350: Generate and output maintenance early warning signal: Construct a time-domain status observation window backtracking; if the number of consecutive frames of an abnormal area exceeds the preset early warning trigger threshold, then trigger the maintenance early warning; calculate the area change rate ΔP=P between adjacent frames. t -P t-1 And the area variance σ within the window P 2 If there are k consecutive frames whose area change rate ΔP is greater than a preset surge threshold θ, then... jump Furthermore, within the consecutive n frames after ΔP approaches 0, the area variance σ P 2 Less than the preset stability threshold If the abnormal intensity within the window is within a range of slowly increasing positive values ​​(i.e., 0 < k' < θ), it is considered an environmental anomaly, such as camera shift or human-caused obstruction. slow θ slow If the threshold increases slowly and the fitted correlation coefficient is greater than the preset confidence threshold, it is determined that the equipment condition has deteriorated, such as a dirty lens or a degraded image sensor performance.

[0104] This embodiment achieves real-time self-healing and operation and maintenance monitoring of the physical performance of the reuse sensing nodes through step S300, ensuring high consistency of the output structured features. Step S400 takes the highly consistent feature data and realizes the application of business assets through spatiotemporal coupling with external business dimensions. Step S400 specifically includes:

[0105] Step S410: Perform confidence filtering on lightweight feature data using operation and maintenance early warning signals;

[0106] Step S420: Multi-source data cross-validation and business early warning closed loop;

[0107] Step S430, encapsulation of structured features;

[0108] Step S440: Output value-added service signal to application platform.

[0109] Step S410: Utilize maintenance early warning signals to perform confidence filtering on lightweight feature data: If the corresponding unit area does not trigger maintenance early warning, the feature data is determined to be high-confidence valid data; call the irreversible hash algorithm operator to desensitize the sensitive attributes in the high-confidence data and generate a highly reliable desensitized feature set.

[0110] Step S420, Multi-source data cross-validation and business early warning closed loop: Using a spatiotemporal consistency discrimination algorithm, the target behavior trajectory and occurrence time in the highly reliable de-identified feature set are matched with the reservation information in the application data; if the two are inconsistent within a preset logical threshold, a business-level anomaly judgment is triggered. Specifically, this includes:

[0111] Step S421: Extract the sampling timestamp and spatial index coordinates from the desensitized features, and simultaneously extract the reservation time interval and preset physical location from the declaration data; use the time synchronization operator to map the two to the same standard time reference system to construct a unified spatiotemporal comparison benchmark; the time synchronization operator is a mathematical correction function that calibrates time records from different sources to a unified standard atomic time.

[0112] Step S422: Using the target association operator, desensitized features with the same characteristic attributes appearing at different time points are chained together to generate the target behavior trajectory.

[0113] Step S423: Invoke the spatiotemporal consistency discrimination algorithm to calculate the dynamic time warping (DTW) distance between the measured target behavior trajectory and the preset path in the reservation information, and compare whether the measured occurrence time is within the reservation time interval; if the DTW distance is greater than the preset spatial deviation threshold, or the measured occurrence time deviates from the reservation time interval, it is determined that the two are inconsistent, triggering the business-level anomaly judgment.

[0114] Step S424: Generate handling instructions by retrieving the corresponding response strategy preset library based on DTW distance; the response strategy preset library is a data warehouse that stores standardized handling processes for different types of business anomalies; it includes instruction templates for on-site verification, automatic alarm, secondary review, etc.

[0115] Step S430, Encapsulation of structured features: Encapsulate the business-level anomaly judgment, handling instructions, and high-reliability de-identified feature set according to a preset format to generate a structured value-added data asset package.

[0116] Step S440: Output value-added service signal to application platform: Push the packaged data assets to third-party application platform through service subscription distribution mechanism.

Claims

1. A distributed, seamless intelligent operation and maintenance and resource value-added method based on reuse and recycling, characterized in that, include: Step S100: The existing camera is reused through the edge adapter. Without changing the original wiring structure, the original video stream is acquired and enhanced in real time using the built-in deep network super-resolution reconstruction module to generate a high-resolution standardized bitstream. Step S200: Without adding an active supplementary light source and without relying on the wearable device of the test object, perform spatiotemporal context analysis on the standardized bitstream, extract the structured features in the ambient diffuse light, perform perception recognition on the original video stream, and generate lightweight feature data. Step S300: Construct a graph structure containing structured features and input it into a multi-layer graph convolutional network to generate feature vectors for each unit region for dynamic baseline comparison. Regions with differences exceeding the dynamic threshold are marked as abnormal regions and selective unsharpened mask compensation is performed. Operation and maintenance early warning signals are output. Step S400: Cross-validate the lightweight feature data with the application data from external business systems to form a business early warning closed loop; The extracted structured features are packaged into data assets according to a preset directory and then output to the application platform.

2. The method according to claim 1, characterized in that, Step S200 specifically includes: Step S210: Receive the high-resolution normalized bitstream output by the edge adapter, and use the image segmentation unit to segment the single frame image into several unit regions in the spatial dimension. Step S220: For each segmented unit region, without turning on the active supplementary light source, the structured features formed by the ambient diffuse reflection light in that region are extracted by the gray-scale feature management unit and the texture feature management unit respectively. The structured features include gray-scale feature values ​​and texture feature values. Step S230: Concatenate the grayscale feature values ​​and texture feature values ​​dimensionally to generate the region feature vector corresponding to each unit region; Step S240: Calculate the Euclidean distance between the centroids of adjacent unit regions, and mark the regions with a distance less than a preset distance threshold as neighborhoods; traverse the neighborhoods to calculate the gray-level gradient, and record the unit region corresponding to the maximum gray-level gradient as the adjacent region, while defining the maximum gradient value as the edge weight. Step S250: Collect all region feature vectors to form a node feature matrix, and collect all edge weights to construct an adjacency matrix. Step S260: Use the spatiotemporal context association algorithm to perform cross-frame and cross-device feature matching on the graph structure data, perform cross-device weak association moving target re-identification, and output cross-device trajectory association features; Step S270: The node feature matrix and trajectory association features are fused to generate lightweight feature data.

3. The method according to claim 2, characterized in that, The specific process of step S260 includes: Step S261: Using the spatiotemporal context association algorithm, the structured features of the current frame of the same sensing device and the historical feature baseline of the previous frame are combined to perform correlation calculation and obtain a multimodal spatiotemporal context feature vector with temporal continuity and spatial semantic information. Step S262: The structured features of the current frame are used as the target to be searched. Based on the physical distribution relationship between the current device and other devices in the local area network, the logical offset time from the target moving from the neighboring device to the current device is calculated. Based on the offset time, feature data of other devices within the corresponding time period are called to construct the target spatiotemporal sliding window reference baseline pool. The similarity between the target feature sequence and each candidate feature vector in the target spatiotemporal sliding window reference baseline pool in the time domain is obtained by normalized cross-correlation calculation, resulting in a set of correlation coefficient value sequences distributed over time. If the frequency of correlation coefficient values ​​exceeding the preset confidence threshold within the time step reaches a preset proportion, the system confirms that the target detected by the current device and the target in the reference baseline pool are the same physical entity, completing the cross-device identity alignment and outputting cross-device trajectory association features. Step S263: Obtain the weak association re-identification result signal and output the cross-camera trajectory association features; the weak association re-identification result signal includes the target unique identifier, state measure and confidence score; the state measure includes the target's real-time motion vector; the confidence score indicates the degree of certainty in determining the target's identity.

4. The method according to claim 3, characterized in that, The confidence score acquisition process in step S263 includes: Step S2631, mapping the correlation coefficient value exceeding the preset confidence threshold in step S262 to the interval [0, 1] as an initial visual confidence score S vision ; Step S2632, obtaining the moving track of the target by using Kalman filtering algorithm; calculating the coincidence degree G of the current track and the predicted track of the target by using motion direction prediction algorithm motion ; Step S2633, according to the topological relationship of the sensing device, the logical offset time G of the target appearance is calculated topology ; Step S2634: The final execution score (Stotal) is obtained using a non-linear weighted summation algorithm. Stotal=αS vision +βG motion +γG topology , Where α, β, and γ are weights, respectively.

5. The method according to claim 2, characterized in that, Step S300 specifically includes: Step S310: Receive the node feature matrix and adjacency matrix as input, and use the spatial adjacency relationship and edge weight defined by the adjacency matrix to construct a graph structure data model representing the spatiotemporal context relationship of the current video frame. Step S320: The graph structure data model is input to a pre-defined multi-layer graph convolutional network (GCN) for forward propagation calculation, and the output is a unit region evolution feature vector with high-order semantic enhancement. Step S330: Using the unit region evolution feature vector as the first input and the historical feature baseline of the previous frame as the second input, the spatial deviation component between the evolution feature vector and the historical feature baseline is calculated using the second-order norm distance as the anomaly intensity. When the anomaly intensity exceeds the preset dynamic threshold, the corresponding coordinate space is marked as an anomaly region. The anomaly area ratio is obtained based on the ratio of the number of anomaly regions to the total number of unit regions N. Step S340: Using the coordinates of the locked abnormal region as input, perform selective unsharpened mask compensation on the corresponding local image block in the original bitstream. Step S350: Construct a time-domain status observation window backtracking. If the number of consecutive frames in an abnormal region exceeds a preset warning trigger threshold, an operation and maintenance warning is triggered; calculate the area change rate ΔP=P between adjacent frames. t -P t-1 And the area variance σ within the window P 2 If there are k consecutive frames whose area change rate ΔP is greater than a preset surge threshold θ, then... jump Furthermore, within the consecutive n frames after ΔP approaches 0, the area variance σ P 2 Less than the preset stability threshold If the abnormal intensity within the window is not found, it is determined to be an environmental mutation; if the slope k' of the first-order linear regression fitting line satisfies 0 < k' < θ, then it is considered an environmental mutation. slow θ slow If the threshold for slow growth is reached and the fitted correlation coefficient is greater than the preset confidence threshold, then the equipment is considered to be in a state of degradation.

6. The method according to claim 5, characterized in that, Step S310 specifically includes: Step S311: Assign a globally unique spatial index number to each unit region, and use the relative positions of the center coordinates of each region to establish an initial zero-weight complete graph in the computation space. Step S312: Sequentially fill each region feature vector into the corresponding index node of the zero-weight complete graph to generate the initial node feature matrix V; Step S313: Obtain the center coordinate sequence P={p1,p2,...,p} for each unit region. n } and preset distance threshold D th For any two nodes in a zero-weight complete graph, calculate the Euclidean distance d between their centers. ij The Euclidean distance is compared with a preset distance threshold; if d ij <D th If the j-th node belongs to the logical neighborhood of the i-th node, then a binary connectivity matrix B is generated to describe the connectivity of the graph structure. If i and j satisfy the logical neighborhood relationship, then B... ij =1, otherwise B ij =0; Step S314, for B in the binary connection matrix B ij For elements with a value of 1, retrieve the corresponding node U. i with U j Locate the adjacent edge regions of the two regions in the original image and set them as the sampling range for gradient calculation; traverse all pixel pairs of the adjacent edges between node i and node j, calculate their gray-level differences, and extract the maximum gray-level difference as the local maximum gray-level gradient value G corresponding to that specific node pair. max (i,j), using the Gaussian kernel function to extract the local maximum gray-level gradient value G. max (i,j) is mapped to the edge weight W connecting nodes i and j. ij ; Set the edge weight W ij Replace the 1 value in the binary connectivity matrix to obtain the adjacency matrix W; Step S315: Encapsulate the node feature matrix V and the adjacency matrix W into a graph structure data model G.

7. The method according to claim 5, characterized in that, Step S400 specifically includes: Step S410: If the corresponding unit area does not trigger an operation and maintenance warning, the feature data is determined to be high-confidence valid data. The irreversible hash algorithm operator is called to de-identify the sensitive attributes in the high-confidence data and generate a highly reliable de-identified feature set. Step S420: Using a spatiotemporal consistency discrimination algorithm, the target behavior trajectory and occurrence time in the high-reliability desensitization feature set are matched with the reservation information in the application data; if the two are inconsistent within a preset logical threshold, an anomaly judgment is triggered at the business level, and a handling instruction is generated. Step S430: Package the business-level anomaly judgment, handling instructions, and high-reliability de-identified feature set according to a preset format to generate a structured value-added data asset package. Step S440: Through the service subscription distribution mechanism, the packaged data assets are pushed to the third-party application platform.

8. The method according to claim 7, characterized in that, Step S420 specifically includes: Step S421: Extract the sampling timestamp and spatial index coordinates from the desensitized features, and simultaneously extract the reservation time interval and preset physical location from the declaration data; use the time synchronization operator to map the two to the same standard time reference system to construct a unified spatiotemporal comparison benchmark. Step S422: Chain-connect the desensitized features with the same characteristic attributes that appear at different time points to generate the target behavior trajectory; Step S423: Invoke the spatiotemporal consistency discrimination algorithm to calculate the dynamic time warping distance between the measured target behavior trajectory and the preset path in the reservation information, and compare whether the measured occurrence time is within the reservation time interval; if the dynamic time warping distance is greater than the preset spatial deviation threshold, or the measured occurrence time deviates from the reservation time interval, it is determined that the two are inconsistent, triggering the business-level anomaly judgment. Step S424: Retrieve the corresponding response strategy preset library based on the dynamic time warping distance to generate a processing instruction.

Images