An edge-cloud collaborative intelligent inspection data processing system for municipal maintenance
Through an intelligent patrol data processing system that integrates edge and cloud, the system adjusts data dimensionality reduction and flow control mechanisms in real time, solving the problem of unstable data processing caused by wireless channel fading under high-speed movement of vehicle-mounted equipment. This achieves data throughput balance and information integrity in a multi-source high-dimensional feature acquisition environment.
Patent Information
- Application Number
- CN202611081580.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-21
- Publication Date
- 2026-11-17
AI Technical Summary
In distributed computing systems, when vehicle-mounted mobile devices collect multi-source high-dimensional spatial geometric features while moving at high speed, wireless channel fading leads to uncertainty in data throughput boundaries, causing data flow compression in collaborative computing, resulting in nonlinear gradient oscillations in the buffer queues of edge nodes. Traditional solutions cannot effectively reconcile the temporal mismatch between front-end feature processing and back-end cloud collaborative computing, leading to wasted computing resources and data loss.
An intelligent patrol data processing system that adopts edge-cloud collaboration monitors channel bandwidth and calculates cascaded flow control response coefficients in real time through the data dimensionality reduction and mapping module and runtime flow control module of the vehicle-mounted edge computing node. It adjusts the data dimensionality reduction and flow control mechanism to form a two-way linkage control closed loop of forward feature adaptive dimensionality reduction and backward runtime flow control. It cuts off data transfer and activates the local loop overlay stack, converting high-dimensional features into 1-dimensional scalar text features while retaining key location coordinates and category codes.
It achieves a balance in computing data throughput under non-ideal operating conditions, reduces waste of computing resources, prevents data loss, and improves the stability and efficiency of data processing. In particular, it can effectively preserve core inspection feature flow data when the communication network is disconnected.
Smart Images

Figure CN123114643A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of distributed computer system technology, and in particular relates to an intelligent municipal maintenance inspection data processing system with edge-cloud collaboration. Background Technology
[0002] In current cloud-based collaborative computing architectures for distributed computer systems and multi-node collaborative data processing, the non-steady-state distributed data processing and computing resource scheduling of high-concurrency data streams across multiple nodes are common technical challenges. Edge node caching structures combined with static filtering thresholds are used to remove computational redundancy during cross-node distributed data processing, thereby reducing the information entropy of global synchronization. However, when in-vehicle mobile devices collect multi-source high-dimensional spatial geometric features at high speeds, the inherent fading effect of wireless channels leads to dynamic uncertainty in the data throughput boundaries during cloud-based collaborative interaction. Sudden surges in collaborative computing data streams cause nonlinear oscillations in the backlog gradient of the local cache queues at edge nodes, thus disrupting traditional multi-node collaborative data processing... The control pre-sets the static predictability of the data flow of computing nodes. Under this condition, the traditional solution completely decouples the front-end feature processing from the back-end cloud collaborative computing state. It can only implement unidirectional data block discarding based on the real-time data processing residual measured by the collaborative computing layer. This passive truncation mechanism cannot track the evolution of the throughput loss of the underlying data processing, which can easily cause high-frequency jumps in the collaborative processing gating boundary. This oscillation not only destroys the continuity of cross-node data flow, but also consumes a lot of edge computing resources due to distributed computing renegotiation, causing secondary computing power overhead. When the device enters the isolated island condition where computing nodes are disconnected from each other, the entire cloud collaborative data feedforward compensation loop will experience parameter update distortion and loop failure due to the loss of input reference.
[0003] To alleviate the aforementioned shortcomings, linear solutions such as increasing cache capacity or increasing chip frequency fail to achieve the desired results. This is because increasing storage space only delays queue overflow and does not eliminate the high-dimensional tensor flow backlog, nor can it reconcile the deep mismatch between front-end dimensionality reduction and back-end scheduling timing. These methods are mainly limited to linear hardware-level scaling. In practical applications, existing software control methods also have shortcomings. For example, Chinese invention patent application CN121256196A discloses a battery online monitoring and fault prediction system based on BSCA-SPCA collaborative optimization. This technology is based on battery state parameters... The slow-changing characteristics of numbers allow for the search of feature subsets within a stable time series domain. The implicit premise for this technique is that the input data stream possesses high steady-state convergence. However, in the scenario of multi-source high-dimensional feature acquisition during high-speed vehicle patrols, this architecture, which relies on slow-changing patterns and iterative optimization, suffers from a fundamental lag due to the uncertainty of throughput boundaries caused by severe channel fading. Furthermore, since its threshold adaptive adjustment is entirely tied to the battery's own health status, it cannot perceive the transient degradation of communication network bandwidth or the nonlinear oscillations of buffer queue backlog. The high-frequency alternating optimization calculations may induce secondary computational overhead under degraded network conditions, leading to loss of control over traffic regulation boundaries.
[0004] Therefore, how to construct a two-way linkage control closed loop that couples forward data feature adaptive dimensionality reduction with backward runtime flow control, so as to maintain the computing data throughput balance of the cloud collaborative system under the conditions of deterioration of distributed computing node collaboration and disconnection, has become the technical problem to be solved by this invention. Summary of the Invention
[0005] This invention aims to solve the problems of high-frequency gating oscillations and feature data loss under non-ideal operating conditions caused by the decoupling of forward feature adaptive dimensionality reduction and backward runtime flow control.
[0006] In this technical solution, an edge-cloud collaborative intelligent municipal maintenance inspection data processing system includes a central server, a communication network, and vehicle-mounted edge computing nodes. The vehicle-mounted edge computing node includes an inspection and perception module, a data acquisition module, a data dimensionality reduction and mapping module, a runtime flow control module, and a topology adaptive scheduling module; The data acquisition module is used to acquire multi-source high-dimensional geometric disease feature data collected by the patrol perception module during its movement. The data dimensionality reduction mapping module is used to determine the variability of spatial geometric features by introducing a pre-dimensionality reduction mapping process and based on the safety threshold value of the geometric variation coefficient. The runtime flow control module is used to monitor the channel bandwidth and calculate the cascaded flow control response coefficient. Based on the cascaded flow control response coefficient, it generates rewriting instructions and injects them into the front-end dimensionality reduction image process to rewrite the geometric variation coefficient safety threshold value upward according to the proportion of the cascaded flow control response coefficient, thus linking the front-end dimensionality reduction and the back-end flow control. The topology adaptive scheduling module is used to cut off data transfer and activate the local cyclic overlay stack when the communication network is disconnected, based on at least one of the following: the cascaded flow control response coefficient exceeds the cascaded flow control threshold and the available bandwidth residual returns to zero. The local cyclic overlay stack is used to degrade high-dimensional image features into 1-dimensional scalar text features, discard redundant background data, and retain the location coordinates and category codes of multi-source high-dimensional geometric disease feature data.
[0007] Preferably, the vehicle-mounted edge computing node also includes a local storage module and a local clock protocol stack; when the communication network is disconnected and the local cyclic overlay stack is activated, the topology adaptive scheduling module starts a two-stage qualification stripping mechanism and cuts off the real-time network parameter input, reads the spatiotemporal overlap factor of the edge node accumulated before the disconnection from the local storage module, and uses the historical gradient decreasing trend of the spatiotemporal overlap factor of the edge node as a control benchmark to perform a step-by-step fixed-step correction on the logical rack bounce of the local clock protocol stack.
[0008] Preferably, the inspection and perception module includes a laser scanning module and an image acquisition module; when the inspection and perception module is in a high-speed moving state, the data acquisition module acquires multi-source high-dimensional geometric defect feature data synchronously collected by the laser scanning module and the image acquisition module; the multi-source high-dimensional geometric defect feature data includes the three-dimensional point cloud coordinates of pavement cracks and the pixel matrix of apparent defect images.
[0009] Preferably, the local loop overlay stack includes a text feature conversion module, which is used to degrade high-dimensional image features into 1-dimensional scalar text features, so that the local storage space of the vehicle edge computing node can additionally carry 30 times the number of 1-dimensional scalar text features.
[0010] Preferably, the runtime flow control module calculates the cascaded flow control response coefficient by extracting the available bandwidth residual and delay fluctuation index of the channel bandwidth in real time; when the cascaded flow control response coefficient generates a rewrite instruction, the distributed network topology and computing power allocation parameters are adjusted.
[0011] Preferably, when the data dimensionality reduction mapping module rewrites the geometric coefficient of variation safety threshold upwards, it reduces the number of data frame sequences to be sent based on the geometric coefficient of variation safety threshold to reduce the input load.
[0012] Preferably, when the topology adaptive scheduling module reads the spatiotemporal overlap factor of the edge nodes accumulated before the disconnection from the local storage module, it extracts the historical gradient change rate of the spatiotemporal overlap factor of the edge nodes.
[0013] Preferably, the topology adaptive scheduling module uses a stepped fixed-step correction to control the logic rack bounce of the local clock protocol stack, limiting the logic rack bounce to a preset range in microseconds.
[0014] Preferably, when the communication network connection is restored, the central server receives the 1D scalar text features carried from the local circular overlay stack by the vehicle-mounted edge computing node and updates the global consistent state topology.
[0015] Compared with existing technologies, the intelligent municipal maintenance inspection data processing system of the present invention, which is based on edge-cloud collaboration, has the following advantages: 1. In the intelligent inspection data processing of municipal maintenance, a second-order differential interlock relationship is established between the backlog gradient parameters of the cache queue read by the processor and the feedback flow control instructions temporarily stored in the local memory. This generates a cascade flow control response coefficient that fluctuates dynamically with the communication link. When the cascade flow control response coefficient exceeds the cascade flow control threshold, the system triggers a high-frequency event-driven mechanism and shortens the feature slice period within the current fixed time window to 1 / 2 of the current running period. This eliminates the high-frequency oscillation of the gating logic generated when relying on real-time bandwidth residual filtering of data. Without introducing external hardware, a stable balance of the concurrent data flow throughput of multiple nodes on the distributed heterogeneous edge side is achieved.
[0016] 2. By generating in-situ rewriting instructions from cascaded flow control response coefficients and injecting these instructions back into the preceding data dimensionality reduction mapping process, the safety threshold value of the geometric variation coefficient for determining the variability of spatial geometric features is rewritten upwards according to the linear positive correlation ratio of the cascaded flow control response coefficients. This forms a two-way collaborative response closed loop of forward feature dimensionality reduction and backward runtime sequence flow control, changing the control defects of isolated decoupling of data collection and transmission control in traditional technologies. It adaptively reduces the input load when sending data frame sequences from the data source, thereby stably controlling the transmission path degradation of the distributed multi-vehicle edge computing node system within the preset safety boundary.
[0017] 3. When the cascaded flow control response coefficient continuously exceeds the cascaded flow control threshold or the available bandwidth residual returns to zero, indicating that the communication network is in a disconnected state, the system automatically cuts off the transfer of feature data between multiple nodes and activates the local cyclic overlay stack. This stack automatically degrades the received high-dimensional image features into 1-dimensional scalar text features. By discarding redundant background data and retaining only the position coordinates and category codes of geometric features, the carrying capacity of local storage space is increased by 30 times, ensuring that the core inspection feature stream data is not lost during the disconnection phase. This solves the problem of rigid overflow of the edge cache queue in a short time and the resulting loss of key features under the previous disconnection conditions. Attached Figure Description
[0018] Figure 1This is a schematic diagram of the multi-source feature dimensionality reduction and temporal flow control scheduling processing module of the present invention; Figure 2 This is a diagram illustrating the communication network disconnection and recovery status under high-speed mobile conditions, as presented in this invention. Detailed Implementation
[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.
[0020] An edge-cloud collaborative intelligent municipal maintenance inspection data processing system includes a central server, a communication network, and vehicle-mounted edge computing nodes. The vehicle-mounted edge computing node includes an inspection and perception module, a data acquisition module, a data dimensionality reduction and mapping module, a runtime flow control module, and a topology adaptive scheduling module; The data acquisition module is used to acquire multi-source high-dimensional geometric disease feature data collected by the patrol perception module during its movement. The data dimensionality reduction mapping module is used to determine the variability of spatial geometric features by introducing a pre-dimensionality reduction mapping process and based on the safety threshold value of the geometric variation coefficient. The runtime flow control module is used to monitor the channel bandwidth and calculate the cascaded flow control response coefficient. Based on the cascaded flow control response coefficient, it generates rewriting instructions and injects them into the front-end dimensionality reduction image process to rewrite the geometric variation coefficient safety threshold value upward according to the proportion of the cascaded flow control response coefficient, thus linking the front-end dimensionality reduction and the back-end flow control. The topology adaptive scheduling module is used to cut off data transfer and activate the local cyclic overlay stack when the communication network is disconnected, based on at least one of the following: the cascaded flow control response coefficient exceeds the cascaded flow control threshold and the available bandwidth residual returns to zero. The local cyclic overlay stack is used to degrade high-dimensional image features into 1-dimensional scalar text features, discard redundant background data, and retain the location coordinates and category codes of multi-source high-dimensional geometric disease feature data.
[0021] Preferably, the vehicle-mounted edge computing node also includes a local storage module and a local clock protocol stack; when the communication network is disconnected and the local cyclic overlay stack is activated, the topology adaptive scheduling module starts a two-stage qualification stripping mechanism and cuts off the real-time network parameter input, reads the spatiotemporal overlap factor of the edge node accumulated before the disconnection from the local storage module, and uses the historical gradient decreasing trend of the spatiotemporal overlap factor of the edge node as a control benchmark to perform a step-by-step fixed-step correction on the logical rack bounce of the local clock protocol stack.
[0022] Preferably, the inspection and perception module includes a laser scanning module and an image acquisition module; when the inspection and perception module is in a high-speed moving state, the data acquisition module acquires multi-source high-dimensional geometric defect feature data synchronously collected by the laser scanning module and the image acquisition module; the multi-source high-dimensional geometric defect feature data includes the three-dimensional point cloud coordinates of pavement cracks and the pixel matrix of apparent defect images.
[0023] Preferably, the local loop overlay stack includes a text feature conversion module, which is used to degrade high-dimensional image features into 1-dimensional scalar text features, so that the local storage space of the vehicle edge computing node can additionally carry 30 times the number of 1-dimensional scalar text features.
[0024] Preferably, the runtime flow control module calculates the cascaded flow control response coefficient by extracting the available bandwidth residual and delay fluctuation index of the channel bandwidth in real time; when the cascaded flow control response coefficient generates a rewrite instruction, the distributed network topology and computing power allocation parameters are adjusted.
[0025] Preferably, when the data dimensionality reduction mapping module rewrites the geometric coefficient of variation safety threshold upwards, it reduces the number of data frame sequences to be sent based on the geometric coefficient of variation safety threshold to reduce the input load.
[0026] Preferably, when the topology adaptive scheduling module reads the spatiotemporal overlap factor of the edge nodes accumulated before the disconnection from the local storage module, it extracts the historical gradient change rate of the spatiotemporal overlap factor of the edge nodes.
[0027] Preferably, the topology adaptive scheduling module uses a stepped fixed-step correction to control the logic rack bounce of the local clock protocol stack, limiting the logic rack bounce to a preset range in microseconds.
[0028] Preferably, when the communication network connection is restored, the central server receives the 1D scalar text features carried from the local circular overlay stack by the vehicle-mounted edge computing node and updates the global consistent state topology.
[0029] Example 1: The operation mode of a current edge-cloud collaborative municipal maintenance intelligent inspection data processing system under vehicle-mounted mobile conditions is as follows: When the vehicle-mounted edge computing node travels with the inspection equipment at a speed of 60km / h through urban underground tunnels and continuous elevated bridge areas with high electromagnetic shielding, the inspection perception module collects the three-dimensional point cloud coordinates of road surface cracks and the pixel matrix of apparent defects. This causes the data acquisition module to instantly accumulate spatiotemporal tensor flow. At this time, the wireless communication network experiences severe jitter due to the multipath fading effect, causing the available bandwidth residual to suddenly drop from 20Mbps to 4.2Mbps. At the same time, the network latency fluctuation index rises to 45ms. This non-ideal environment caused by the high-frequency compression of data flow and the deep fading of the channel leads to high-frequency jumps in the data flow gating boundary and increased computing power overhead in the decoupled scheduling architecture, resulting in linear data flow divergence and loss of feature information in the core municipal defects.
[0030] In response to the aforementioned deep fading and data backlog in communication network links, the runtime flow control module extracts the available bandwidth residual and latency fluctuation index of the communication network channel bandwidth in real time, calculates the control variables characterizing the degradation degree of the transmission path, and periodically reads the current sampled value and historical cache queue length in the local storage module. The difference between the two is divided by the preset single sampling time interval to obtain the local queue backlog gradient parameter characterizing the saturation evolution trend of the edge storage space. The runtime flow control module uses the local queue backlog gradient parameter as the numerator and performs an algebraic ratio operation with the flow control adjustment gain factor issued by the central server and temporarily stored in the local configuration register. It outputs the cascade flow control response coefficient that fluctuates dynamically with the communication link. When the cascade flow control response coefficient exceeds the cascade flow control threshold temporarily stored in the read-only memory, the system changes the runtime sequence of the sliding window attenuation flow control gating, shortening the feature slice period based on the fixed time window in situ to 1 / 2 of the current running period.
[0031] The runtime flow control module generates in-situ rewriting instructions based on the cascaded flow control response coefficients and injects these instructions into the preceding data dimensionality reduction mapping process. This allows the data dimensionality reduction mapping module to determine the safety threshold value of the geometric variation coefficient of spatial geometric feature variability and rewrite upwards according to the linear positive correlation ratio of the cascaded flow control response coefficients. This changes the filtering level of the preceding feature dimensionality reduction, linking the front-end dimensionality reduction with the back-end flow control, reducing the input load when sending data frame sequences. This enables the overall system to operate under a bidirectional dependent and cooperative architecture of forward operators and backward timing. Simultaneously, the topology adaptive scheduling module receives discrete feature slice streams with priority tags. When it detects high-frequency oscillations in the channel transmission rate between 3.8Mbps and 4.5Mbps, it runs adaptive topology path matrix arbitration to redirect the data stream to a backup logical topology channel with a low packet loss rate, thereby eliminating parasitic delay resistance caused by high-frequency network disturbances.
[0032] When the multi-node data throughput balance between the front-end and back-end changes, i.e. when the cascade flow control response coefficient continuously exceeds the cascade flow control threshold, or when the available bandwidth residual returns to zero due to electromagnetic shielding, thus determining that the communication network has been disconnected, the topology adaptive scheduling module cuts off the feature data transfer between multiple nodes and activates the local cyclic overlay stack. The text feature conversion module inside the local cyclic overlay stack converts the image features into 1D scalar text features, discards redundant background data, and retains the 3D point cloud coordinates of road cracks and the spatial geometric absolute position coordinates and category codes of the apparent defects in the image. This allows the local storage space of the vehicle edge computing node to carry an additional 30 times the number of 1D scalar text features. The core inspection feature stream data is completely retained during the long-term disconnection without overflow or loss. During the conversion process, the text feature conversion module downsamples the length and width resolution of the input 2D image pixel matrix and adjusts it to a 256-pixel multiplied 256-pixel baseline feature map.
[0033] Next, the module uses a preset shallow convolution feature extraction operator to perform edge skeletonization extraction on the baseline feature map, identifying the continuous pixel skeleton representing the crack direction. Then, the module traverses the discrete pixels of the skeleton, extracts the centroid x-coordinate pixel value and y-coordinate pixel value of the crack lesion patch, and uses these two pixel values as absolute position coordinates. Finally, based on the geometric aspect ratio and contour area of the lesion feature, the corresponding lesion category code is retrieved from the preset classification index table. The absolute position coordinates and category code are concatenated into a line of pure string containing 5 discrete values according to a fixed text format. The first 3 values represent the absolute metric coordinates of the 3D point cloud in the spatial coordinate system, and the last 2 values represent the plane center pixel coordinates and category integer code detected by the image, respectively.
[0034] Thus, the original feature stream containing high-dimensional matrix information of brightness and darkness of hundreds of thousands of pixels is compressed and degenerated in situ into a 1-dimensional scalar text feature string occupying only 24 bytes, achieving a significant reduction in data dimensionality. At the same time, the topology adaptive scheduling module initiates a two-stage qualification stripping mechanism and cuts off the real-time network parameter input. Instead, it reads the spatiotemporal overlap factor of the edge nodes accumulated in the last effective sampling period before the disconnection from the local storage module. Using the historical first-order gradient decreasing trend of the spatiotemporal overlap factor of the edge nodes as the control benchmark, and based on the continuous decay state of the spatiotemporal overlap factor, the logical rack bounce of the local clock protocol stack is corrected in a step-by-step fixed step manner. Under the condition of loss of network feedback, the motion inertia of the vehicle's physical displacement is used to maintain the time axis alignment of the local operator, and the logical rack bounce is controlled within the microsecond preset range to avoid the risk of scheduling misalignment and logical dead loop caused by runtime timing lag.
[0035] When the vehicle leaves the electromagnetically shielded area such as the underground tunnel and the communication network is restored, the central server receives the 1D scalar text features transferred from the local cyclic overlay stack by the vehicle-mounted edge computing node. When the integrity factor of the received data frame meets the preset convergence condition, the newly sent feature data is merged into the global historical state topology. At the same time, based on the current central concurrent queue pressure parameters, the latest flow control adjustment gain factor is calculated and injected into the runtime flow control module of the vehicle-mounted edge computing node. This reduces the average transient latency of global data transfer from 850ms to 35ms. Under the condition that the wireless communication link bandwidth is limited to 70%, the integrity of feature data transmission for municipal disease feature data remains above 99.4%. This distributed collaborative architecture, which relies on the spatiotemporal transmission characteristics of a single network channel, keeps the computing load and communication overhead stable in the dynamic second-order differential interlocking closed loop of forward dimensionality reduction and backward flow control, maintaining the balance of data flow throughput and information protection of the multi-node high-concurrency distributed system in the mobile patrol environment.
[0036] Example 2: When the system faces the inspection conditions of high-frequency acquisition by vehicle-mounted mobile computing nodes and deep fading of transmission channels due to spatial physical shielding, the experiment is carried out on an edge cloud collaborative hardware-in-the-loop simulation platform containing multi-level edge processing nodes and a distributed central server. The experimental input data comes from the standard municipal maintenance laser point cloud and disease appearance image verification dataset, and is mixed with the on-site measured physical data streams obtained by a 3D LiDAR with a scanning frequency of 100Hz and a high-definition perception camera with a resolution of 3840 x 2160 pixels. At the same time, in order to reproduce the complex electromagnetic interference in underground tunnels and continuous viaduct areas, the wireless communication channel simulation unit actively superimposes a multipath Rayleigh fading model and Gaussian white noise with a signal-to-noise ratio of 20dB on the data bus transport path, and introduces power frequency harmonic disturbances with a frequency of 50Hz by the baseband interference generator, thereby constructing a non-ideal engineering environment in which the baseline network packet loss rate randomly operates between 4.5% and 6.2%.
[0037] In the initial configuration of the above simulation test platform, the determination of the single sampling time interval on which the runtime flow control module depends includes closed programmatic decision-making logic. The range of the single sampling time interval in this invention is 50ms to 200ms. In this embodiment, the impact of different single sampling time intervals on system performance is examined. The data shows that when the single sampling time interval is set to 50ms, the hardware interrupt frequency of the local chip is too high, which leads to an increase in the task drop rate due to frequent context switching of the processor core computing resources. When the single sampling time interval is set to 200ms, the update of the edge storage queue status is delayed, which leads to failure to respond in time when there is a sudden backlog of high-dimensional tensor flow and data overflow and drop. When the single sampling time interval is set to 100ms, the feature data throughput time and processor computing overhead are kept in a stable balance. Therefore, under the condition that the vehicle speed is 60km / h and the feature concurrent throughput is 150 frames per second, the single sampling time interval is selected as 100ms.
[0038] The system includes an experimental group consisting of a complete technical solution, a control group that selectively removes the cascaded flow control feedback rewriting process while keeping the dimensionality reduction threshold static, a first control group with the cascaded flow control threshold set below the lower limit of the normal threshold, and a second control group with the cascaded flow control threshold set above the upper limit of the normal threshold. The system operates in a problem intensity gradient that progressively increases from low-level to high-level channel fading. In a high-level channel fading scenario where the available bandwidth residual decreases from 15.0 Mbps to 8.0 Mbps and then suddenly drops to 4.2 Mbps, the buffer queue length generated by the data acquisition module increases from the initial 12 frames. At this point, the local queue backlog gradient parameter calculated by the runtime flow control module in the test group reaches 85.4 frames per second, and the cascaded flow control response coefficient derived by differential interlocking increases to 2.46, triggering a high-frequency event-driven state, shortening the feature slicing period to 15.2ms, and at the same time, the generated in-situ rewrite instruction drives the geometric variation coefficient safety threshold value in the data dimensionality reduction mapping module to be rewritten from the initial 0.15 to 0.38, thereby increasing the background pixel filtering rate in the apparent lesion image pixel matrix from 32.4% to 78.6%, reducing the input load when sending data frame sequences, and the output feature data stream volume converges to the low bandwidth channel.
[0039] When the cascaded flow control response coefficient exceeds the cascaded flow control threshold, and the available bandwidth residual is determined to be disconnected due to shielding reaching zero, the topology adaptive scheduling module cuts off the feature data transfer between the vehicle-mounted edge computing node and the central server, activating the local cyclic overlay stack. The text feature conversion module within this stack converts image features into 1D scalar text features, discards redundant background data, and retains the 3D point cloud coordinates of road cracks and the spatial geometric absolute location coordinates and category codes of apparent defects in the image. This allows the local storage space of the vehicle-mounted edge computing node to hold 30.4 times the number of 1D scalar text features, maintaining connectivity during the communication disconnection. The core inspection feature stream data is completely retained without overflow or loss. The theoretical basis for this 30-fold increase in carrying capacity comes from the comparison of space storage occupancy between high-dimensional image tensors and plain text documents: Under normal operating conditions, even after basic compression, the physical volume of a single frame of single-channel road surface defect pixel matrix data acquired by the image acquisition module reaches 720 kilobytes; while the 1-dimensional scalar text features generated by the text feature conversion module, due to the complete removal of redundant non-feature pixel information such as background color, texture, and environmental noise, retain only pure numeric strings representing geometric defect location coordinates and category codes, the physical size of a single text string is limited to 24 bytes.
[0040] By dividing the original image volume of 720 kilobytes by the text volume of 24 bytes, the data compression ratio of single-frame features before and after transformation is as high as 30,000 times. After deducting the fixed hardware overhead occupied by the local cyclic overlay stack management and control protocol, the local storage space of the vehicle edge computing node can safely carry more than 30 times the total number of discrete disease feature text records under the same hardware capacity, thus theoretically proving the technical effect. At the same time, the topology adaptive scheduling module starts the two-stage qualification stripping process, cuts off the network parameter input, and reads the spatiotemporal position overlap factor of the edge node recorded in the last sampling period before the disconnection from the local storage module. The step-by-step control process of the two-stage qualification stripping mechanism is immediately awakened when the communication network is determined to be disconnected: the first stage is the physical stripping of the network parameter input: when the system detects that the available bandwidth residual has been continuously zero for 200 milliseconds, the underlying task scheduler forcibly suspends the network socket polling thread responsible for receiving wireless baseband signals, stops the collection of external dynamic fading factors, and sets the network handshake timeout judgment flag to invalid.
[0041] The second stage involves preempting and stripping edge node computing priorities: Within 10 milliseconds immediately following the aforementioned network socket suspension, the local CPU reclaims and strips all non-core computing threads currently concurrently executing tasks such as central synchronization state maintenance and global historical state tree updates, which rely on cloud interaction. These threads are then converted to a local inertial timing calculation state with the highest interrupt priority driven by a hardware timer. This frees up core chip resources for the exclusive operation of the local loop overlay stack and local clock protocol stack. Utilizing the inherent historical first-order gradient decrease trend of edge node spatiotemporal overlap as a control benchmark, and based on the spatiotemporal overlap... The continuous decay state of the stacking factor and the step-by-step fixed-step correction of the logic rack bounce of the local clock protocol stack maintain the time axis alignment of the local operator by relying on the motion inertia of the vehicle's physical displacement in the absence of network feedback. The logic rack bounce is controlled within a preset range of 5.4 microseconds to 12.6 microseconds to eliminate scheduling misalignment caused by runtime lag. In the gradient scenarios of low-degree channel fading, medium-degree channel fading and high-degree channel fading, as the degree of communication degradation increases, the data compression ratio of the text feature conversion module adaptively evolves from small to large in a step-by-step manner, and the corresponding data retention time shows a gradient extension law.
[0042] When the communication network is restored, the central server receives the 1D scalar text features transferred from the local cyclic overlay stack by the vehicle-mounted edge computing node. When the received data frame integrity factor meets the preset convergence condition, the newly uploaded feature data is merged into the global historical state topology. Simultaneously, based on the current central concurrent queue pressure parameters, the latest flow control adjustment gain factor is calculated and injected into the runtime flow control module of the vehicle-mounted edge computing node. At this point, the average transient latency of the system's global data transfer converges stepwise from 852.4ms to 35.6ms. Under conditions where the wireless communication link bandwidth is limited to 70%, the feature data transmission integrity for municipal disease feature data remains at 99.43%, and this integrity is maintained even under low-level, medium-level, and high-level channel fading. In the channel fading gradient control test, the feature data transmission integrity of the experimental groups were 99.94%, 99.76%, and 99.43%, respectively, with corresponding average transient delays of 12.4ms, 22.8ms, and 35.6ms. In contrast, the feature data transmission integrity of the partially missing first control group dropped to 52.17% under high channel fading conditions. The feature data transmission integrity of the out-of-range first control group and the out-of-range second control group was limited to 71.34% and 64.82%, respectively. These results conclusively confirm that the computing load and communication overhead maintain a balance in the dynamic second-order differential interlocking closed loop of forward dimensionality reduction and backward flow control. The multi-node high-concurrency distributed system maintains stable data flow throughput and complete inspection feature information under severe channel dynamic disturbances.
[0043] Example 3: This example combines Figures 1 to 2This document describes a smart data processing system for municipal maintenance inspections that integrates edge computing and cloud computing. Figure 1 As shown, the inspection and perception module collects high-dimensional features and outputs them unidirectionally to the data acquisition module to obtain multi-source features. The data acquisition module obtains multi-source features and transmits them to the data dimensionality reduction and mapping module to determine the feature variability. The data is then transferred to the topology adaptive scheduling module to be converted into one-dimensional scalar text and input into the communication network channel bandwidth for transmission. The communication network channel bandwidth transmission sends the data to the central server to update the status topology. At the same time, it connects to the runtime flow control module via the monitoring channel bandwidth to calculate the cascade response coefficient. The runtime flow control module calculates the cascade response coefficient and sends it back to the data dimensionality reduction and mapping module via reverse injection rewrite instructions to determine the feature variability.
[0044] like Figure 2 As shown, when in a high-speed moving state, the response coefficient exceeds the threshold and triggers a severe jitter. At the node where severe jitter occurs, the system either resets and loops to the backup channel, or when the available bandwidth residual returns to zero, it flows unidirectionally to determine if the communication network is disconnected. During the period when the communication network is disconnected, the system maintains its loop by activating the local overlay stack. When leaving the electromagnetic shielding area, the system enters the communication network to restore the connection. When the communication network is restored and the preset convergence condition is met, the system returns to the high-speed moving state.
[0045] Example 4: When the system faces the inspection situation where the vehicle-mounted edge computing node collects data and the transmission channel suffers deep fading due to spatial physical shielding, the data acquisition module, with the local high-speed cache queue capacity limit set to 128 frames and the processor clock frequency of the vehicle-mounted edge computing node at 2.0 GHz, collects multi-source high-dimensional geometric disease feature data composed of three-dimensional point cloud coordinates and apparent disease image pixel matrix at a rate of 150 frames per second. This causes the data backlog in the local storage module to approach the hardware overflow boundary during the passage through a 1.2-kilometer urban dense building area. At this time, the available bandwidth residual is reduced to 3.5 Mbps due to the multipath obstruction of the wireless signal. Since the data input rate is greater than the channel transmission rate, the traditional decoupling algorithm causes buffer memory out-of-bounds and multi-node state machine logic deadlock within the preset memory sliding window, causing data loss in the inspection feature data chain during the cross-node transfer process.
[0046] The runtime flow control module is used to calculate the spatiotemporal topology consistency differentiation index. The spatiotemporal topology consistency differentiation index is determined by using the spatial physical displacement increment of the vehicle edge computing node within a preset sliding window period as the denominator and the absolute value of the local queue backlog gradient parameter generated by the local storage module within the same sliding window period as the numerator. The ratio reflects the topological distribution state of heterogeneous feature flows under the constraints of the time axis and the spatial axis. When the spatial physical displacement increment of the patrol equipment is 15.6 meters and the local queue backlog gradient parameter is 24 frames per second, the runtime flow control module outputs a spatiotemporal topology consistency differentiation index of 1.54. In order to eliminate the dimensional mismatch between the macroscopic physical motion scale and the memory storage queue frame rate, a normalized dimensional elimination mapping is performed before calculating the index: Specifically, the module divides the measured spatial physical displacement increment by the system's preset standard movement distance constant of 10 meters, eliminates the length dimension, and converts it into a dimensionless displacement scaling coefficient of 1.56.
[0047] Simultaneously, the absolute value of the measured local queue backlog gradient parameter is divided by the baseline throughput rate constant of 10 frames per second, eliminating the time and frame dimensions and converting it into a dimensionless backlog scale coefficient of 2.40. Then, the timing flow control module uses the converted backlog scale coefficient as the numerator and the displacement scaling coefficient as the denominator to perform an algebraic ratio operation, thereby achieving logical alignment of cross-dimensional parameters at the same dimensionless level. Finally, it outputs a dimensionless heterogeneity index value of 1.54, representing the bidirectional topological distribution characteristics of time and space, safely and self-consistently. During runtime, the timing flow control module retrieves the discrete step damping rule in the read-only memory. According to the trend that for every 0.1 increase in the spatiotemporal topological consistency heterogeneity index, the flow control adjustment gain factor is reduced by 5% in the opposite direction, the flow control adjustment gain factor temporarily stored in the control register is corrected from 1.20 to 0.73. The corrected flow control adjustment gain factor is then input into the differential calculation loop of the cascaded flow control response coefficient, driving the operator timing trigger point to work in the convergence direction of the sampling window and adjusting the load of the edge storage space.
[0048] The system is equipped with a timeliness reconstruction component for stabilizing the clock protocol stack. This component eliminates outdated historical data through a time sliding window covering 48 consecutive inspection cycles. When the system detects that the clock deviation variance between the vehicle edge computing nodes in the multi-node collaborative network exceeds the safety deviation variance threshold for 10 consecutive sampling points, the central server sends a network topology recalibration command. Each vehicle edge computing node switches to the backup control path and uses the physical acceleration integral result output by the local inertial navigation system to replace the wireless network timing signal. The specific replacement control algorithm is as follows: when the component detects that the wireless timing signal has been lost and entered the blind zone, the backup control path locks the reference clock crystal oscillator frequency that was successfully synchronized in the last period before the disconnection.
[0049] The algorithm reads the discrete three-axis physical acceleration data output by the local inertial navigation system in real time with a period of 10 milliseconds. It performs a first-order integral over time to calculate the instantaneous speed of the vehicle in the current period, and performs a second-order integral over time to calculate the cumulative spatial displacement distance relative to the disconnection cut-off point. Then, it divides the cumulative spatial displacement distance by the speed of light constant to calculate the theoretical time delay deviation of electromagnetic wave propagation caused by the change in position. Finally, the local protocol stack achieves autonomous fine-tuning of the local system clock main axis by algebraically superimposing the calculated theoretical time delay deviation on the reference timestamp generated by each hardware timer interrupt. This completely eliminates the physical dependence on the timing pulse signal of the external wireless base station network. Under the condition of disconnection for 300 seconds when the vehicle passes through the shielded blind zone, the bounce of the logic rack inside the vehicle edge computing node is stable within the range of 5.4 microseconds to 12.6 microseconds. This underlying transmission chain, which performs step correction of microsecond-level protocol stack clock jumps based on macroscopic physical displacement, is realized by using the evolution law of spatial propagation time delay of radio waves as a physical intermediate medium.
[0050] When a vehicle moves at high speed within a disconnection blind zone, its movement causes a deterministic physical rate of change in the spatial distance between it and the previously associated tunnel base station. Since the propagation speed of electromagnetic waves in air is a constant, this increase in spatial physical displacement directly leads to a linear cumulative lag or advance in the wireless signal air interface transmission delay. Before the system disconnection, the historical decreasing trend of the spatiotemporal overlap factor of edge nodes stored in the local storage module is physically equivalent to the macroscopic motion Doppler evolution trend of the vehicle moving away from the communication multipath overlap center. The system maps this historical first-order gradient decreasing trend into a stepwise increase in the expected air interface delay. The predicted quantity is used to directly compensate the queuing delay of the local network protocol stack logic rack's receiving buffer queue in microsecond increments, thereby offsetting the communication network transmission lag caused by physical displacement. This maintains the time axis alignment of the local operator in the absence of external time synchronization. Under the operating conditions of distributed computing with high concurrency tensor throughput of 150 frames per second and channel bandwidth limited to 70%, the integrity of the end-to-end transmission of municipal disease characteristic data is maintained at 99.43%, and the average transient latency of data transfer is controlled at 35.6ms. The bidirectional correlation of spatiotemporal parameters is used to maintain the data throughput balance of the distributed collaborative system.
[0051] Example 5: When the system faces the pre-calibration condition of on-site deployment of vehicle-mounted edge computing nodes, the on-site calibration process is started in the reference environment where the vehicle is stationary. The data acquisition module loads simulated data frames into the local memory in a loop, increasing from 20 frames per second to 200 frames per second, driving the processor to undergo a stepped load test. During runtime, the flow control module reads the reference transport delay of the local storage module through the kernel counter, and calculates the ratio of the first-order causal difference rate of the reference transport delay to the core throughput rate to obtain the computing power dissipation coefficient. When the computing power dissipation coefficient exceeds the preset safety index, that is, when the data bus generates a nonlinear dissipation bottleneck, the system uses the measured value at this inflection point as the boundary mapping input to calculate and establish the initial value of the cascade flow control threshold.
[0052] To complete the data filling for the flow control parameters, the on-site calibration process simulates the channel fading change as the available bandwidth residual decreases from 20Mbps to 4.2Mbps. During runtime, the flow control module performs a differential hierarchical concatenation operation on the measured transient tensor flow and the initial value to obtain the gradient calibration table of the flow control adjustment gain factor and writes it into the local configuration register. This gradient calibration table establishes a quantitative interval mapping relationship between the computing power dissipation coefficient and the flow control adjustment gain factor: Specifically, when the computing power dissipation coefficient is in the first preset interval of 1.0 to 1.2, it indicates that there is no nonlinear dissipation bottleneck on the bus, and the flow control adjustment gain factor output by the calibration table is fixed at 1.20. At this time, the front end performs normal data retention.
[0053] When the computing power dissipation coefficient jumps to the second preset range of 1.2 to 1.5, it indicates that the bus begins to generate slight dissipation. The flow control adjustment gain factor corresponding to the calibration table linearly and negatively decays to 0.95, thus limiting the flow rate. When the computing power dissipation coefficient further deteriorates to the third preset range of 1.5 to 2.0, the flow control adjustment gain factor in the calibration table is forced to step-shrink to 0.73, driving the operator timing convergence. When the computing power dissipation coefficient exceeds the limit boundary of 2.0, the calibration table directly outputs a minimum protection gain value of 0.45. The segmented tiered parameter matrix is stored in a local non-volatile configuration register in discrete form, realizing precise negative feedback adjustment of traffic gain based on resource consumption intensity. Based on the update of this benchmark parameter, the topology adaptive scheduling module can automatically rewrite the geometric variation coefficient safety threshold value according to the dynamic transition of the computing power dissipation coefficient in actual vehicle-mounted inspection scenarios, control the integrity of feature data transmission to remain above 99.4%, and ensure that the data throughput of the distributed municipal inspection system operates stably under the closed-loop mechanism of automatic parameter calibration.
[0054] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.
Claims
1. A municipal maintenance intelligent inspection data processing system with edge-cloud collaboration, characterized in that, This includes a central server, a communication network, and in-vehicle edge computing nodes. The vehicle-mounted edge computing node includes an inspection and perception module, a data acquisition module, a data dimensionality reduction and mapping module, a runtime flow control module, and a topology adaptive scheduling module; The data acquisition module is used to acquire multi-source high-dimensional geometric disease feature data collected by the patrol perception module during its movement. The data dimensionality reduction mapping module is used to determine the variability of spatial geometric features by introducing a pre-dimensionality reduction mapping process and based on the safety threshold value of the geometric variation coefficient. The runtime flow control module is used to monitor the channel bandwidth and calculate the cascaded flow control response coefficient. Based on the cascaded flow control response coefficient, it generates rewriting instructions and injects them into the front-end dimensionality reduction image process to rewrite the geometric variation coefficient safety threshold value upward according to the proportion of the cascaded flow control response coefficient, thus linking the front-end dimensionality reduction and the back-end flow control. The topology adaptive scheduling module is used to cut off data transfer and activate the local cyclic overlay stack when the communication network is disconnected, based on at least one of the following: the cascaded flow control response coefficient exceeds the cascaded flow control threshold and the available bandwidth residual returns to zero. The local cyclic overlay stack is used to degrade high-dimensional image features into 1-dimensional scalar text features, discard redundant background data, and retain the location coordinates and category codes of multi-source high-dimensional geometric disease feature data.
2. The edge-cloud collaborative intelligent municipal maintenance inspection data processing system according to claim 1, characterized in that, The vehicle-mounted edge computing node also includes a local storage module and a local clock protocol stack. When the communication network is disconnected and the local cyclic overlay stack is activated, the topology adaptive scheduling module starts a two-stage qualification stripping mechanism and cuts off the real-time network parameter input. It reads the spatiotemporal overlap factor of the edge node accumulated before the disconnection from the local storage module, and uses the historical gradient decreasing trend of the spatiotemporal overlap factor of the edge node as a control benchmark to perform a step-by-step fixed-step correction on the logical rack bounce of the local clock protocol stack.
3. The edge-cloud collaborative intelligent municipal maintenance inspection data processing system according to claim 1, characterized in that, The inspection and perception module includes a laser scanning module and an image acquisition module; when the inspection and perception module is in a high-speed moving state, the data acquisition module acquires multi-source high-dimensional geometric defect feature data synchronously collected by the laser scanning module and the image acquisition module; the multi-source high-dimensional geometric defect feature data includes the three-dimensional point cloud coordinates of pavement cracks and the pixel matrix of apparent defect images.
4. The edge-cloud collaborative intelligent municipal maintenance inspection data processing system according to claim 1, characterized in that, The local loop overlay stack contains a text feature transformation module, which is used to degenerate high-dimensional image features into 1-dimensional scalar text features, enabling the local storage space of the vehicle edge computing node to carry an additional 30 times the number of 1-dimensional scalar text features.
5. The edge-cloud collaborative intelligent municipal maintenance inspection data processing system according to claim 1, characterized in that, The runtime flow control module calculates the cascaded flow control response coefficient by extracting the available bandwidth residual and delay fluctuation index of the channel bandwidth in real time; when the cascaded flow control response coefficient generates a rewrite instruction, the distributed network topology and computing power allocation parameters are adjusted.
6. The edge-cloud collaborative intelligent municipal maintenance inspection data processing system according to claim 1, characterized in that, When the data dimensionality reduction mapping module rewrites upwards based on the geometric coefficient of variation safety threshold, it reduces the number of data frame sequences to be sent to reduce the input load.
7. The edge-cloud collaborative intelligent municipal maintenance inspection data processing system according to claim 2, characterized in that, When the topology adaptive scheduling module reads the spatiotemporal overlap factor of the edge nodes accumulated before the disconnection from the local storage module, it extracts the historical gradient change rate of the spatiotemporal overlap factor of the edge nodes.
8. The edge-cloud collaborative intelligent municipal maintenance inspection data processing system according to claim 2, characterized in that, The topology adaptive scheduling module uses a stepped fixed-step correction to control the logic rack bounce of the local clock protocol stack, limiting the logic rack bounce to a preset range in microseconds.
9. The edge-cloud collaborative intelligent municipal maintenance inspection data processing system according to claim 1, characterized in that, When the communication network is restored, the central server receives the 1D scalar text features carried from the local circular overlay stack by the vehicle-mounted edge computing node and updates the global consistent state topology.
Citation Information
Patent Citations
Storage battery online monitoring and fault prediction system based on BSCA-SPCA collaborative optimization
CN121256196A
