A contact net safety inspection system based on binocular detection

Abstract

The application discloses a kind of integral side-by-side inspection contact network safety inspection system based on binocular detection, belong to computer vision technical field;Including integrated box and its inside integrated binocular image acquisition unit, edge processing unit and display unit. The processor unit inside edge processing unit is constructed with asynchronous double clock first-in-first-out queue and ping-pong frame buffer module, to decouple the timing between acquisition frame period and pre-processing time consumption;Edge computing processor executes defect detection thread and stereo matching thread in parallel, and uses intra-frame local difference model to calculate the local relative deformation variable of contact network components relative to its own fixed base. The application realizes real-time quantitative detection of contact network component deformation under the condition of train high-speed running and vibration, solves the problem of bandwidth bottleneck caused by dependence on cloud in prior art, monocular vision cannot obtain three-dimensional deformation information and the problem of insufficient reliability of split design.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, specifically a binocular detection-based integrated contact network safety inspection system that combines simultaneous imaging and inspection. Background Technology

[0002] As a core component of the train traction power supply system, the overhead contact line is constantly exposed to a complex and ever-changing natural environment. Affected by factors such as wind, snow, and corrosion, it is highly susceptible to defects such as broken droppers, misaligned positioners, and damaged insulators, seriously threatening the safe operation of trains. Traditional manual pole climbing or track vehicle inspection methods are limited by low efficiency, high safety risks, constraints imposed by maintenance windows, and high costs, making them unsuitable for the large-scale maintenance needs of modern high-speed rail networks. With the advancement of digital image processing technology, data acquisition systems based on digital cameras are gradually becoming mainstream, and automated inspection methods are beginning to partially replace manual visual inspection.

[0003] Furthermore, several automated improvement solutions for overhead contact line inspection have been publicly disclosed within the industry. Specifically, Chinese patent application CN112002031A discloses an overhead contact line inspection service base station and an overhead contact line inspection system. This system includes a service base station, a backend server, and a drone. The control server controls the drone's operation and transmits the images captured by the drone to the control server. In addition, the control server integrates an OpenCV-based image processing module, employing deep learning model algorithms to intelligently identify faulty parts in the overhead contact line. On the other hand, to obtain richer spatial geometric information, 3D vision technology is gradually being introduced into the railway inspection field. For example, Chinese patent application CN109377530A discloses a binocular depth estimation method based on deep neural networks. This method uses only the left and right viewpoint images acquired by a binocular camera as network input, constructing a multi-scale network model containing multiple convolutional layers, residual connections, and multi-scale pooling connections, and outputs the corresponding depth map through unsupervised learning.

[0004] However, the aforementioned existing technologies still face several significant technical bottlenecks in the high-speed operation environment of actual high-speed railway lines: Similar to the front-end data acquisition and back-end server processing paradigm described in CN112002031A, this approach is rigidly dependent on wireless communication bandwidth. When a train is traveling at a high speed of 300 kilometers per hour, high-frequency acquisition can generate tens of gigabytes of raw image data within minutes. However, wireless network coverage along railway lines (especially in mountainous areas and tunnels) generally suffers from blind spots or unstable signal quality. This directly leads to data transmission delays, packet loss, and even interruptions, resulting in significant delays in back-end identification results and an inability to provide timely safety warnings to train drivers or dispatchers. Furthermore, ordinary two-dimensional planar judgment has inherent limitations when dealing with scenarios requiring precise quantitative spatial information, such as locator offset measurement.

[0005] If a deep neural network as described in CN109377530A is directly introduced for binocular stereo matching and depth estimation, the multi-scale network model involves complex structures such as multi-layer convolution and multi-scale pooling, resulting in a massive size and computationally cumbersome process. When deployed simultaneously with the defect detection network on a vehicle-mounted computing chip with limited computing resources, it is highly susceptible to insufficient system data throughput due to computing power contention and thread scheduling conflicts, leading to data flow blockage or high-frequency frame drop crashes.

[0006] Traditional inspection devices often employ a separate structure where the acquisition, processing, and communication units are independent, with these functional units connected in series over long distances via physical cables. Under the intense mechanical vibrations of 5Hz to 200Hz generated by high-speed train operation, the physical connection points of this separate system are highly susceptible to loosening or poor contact. This not only significantly reduces the long-term operational reliability of the entire hardware system but also causes micrometer-level relative displacement of the physical baseline of the binocular camera, directly leading to parameter failure in the mathematical formulas for spatial geometric measurements.

[0007] In summary, existing overhead contact line inspection technologies suffer from systemic technical bottlenecks in areas such as data processing paradigms, concurrent scheduling of multiple models at the end points, three-dimensional quantitative measurement, and structural stability under strong vibration conditions, making it difficult to meet the growing demand for intelligent inspection of overhead contact lines in high-speed railways. Summary of the Invention

[0008] The purpose of this invention is to provide an integrated contact network safety inspection system based on binocular detection, which solves the problems in the existing contact network inspection systems mentioned in the background art, such as bandwidth bottlenecks caused by cloud-based data processing architecture, inability of monocular vision to acquire depth information, insufficient reliability of split design, and poor real-time performance of data analysis.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: An integrated contact network safety inspection system based on binocular detection, comprising an integrated housing, which integrates a binocular image acquisition unit, an edge processing unit, and a display unit; the binocular image acquisition unit includes a first industrial camera and a second industrial camera for simultaneously acquiring contact network images; the edge processing unit includes a processor unit and an edge computing processor; characterized in that... The processor unit is internally equipped with an asynchronous dual-clock first-in-first-out queue and a ping-pong frame buffer module. The write-side clock of the asynchronous dual-clock first-in-first-out queue is synchronized with the pixel clock of the industrial camera, and the read-side clock is synchronized with the internal system clock of the processor unit. The ping-pong frame buffer module is used to temporarily store image data to decouple the timing between the acquisition frame period and the preprocessing time. The edge computing processor is configured to execute a defect detection thread and a stereo matching thread in parallel. The defect detection thread is used to detect defects in the contact wire components of the image and output the defect category and a two-dimensional bounding box. The stereo matching thread is used to perform stereo matching on the image to generate a disparity map. Based on the disparity value of the defect region in the disparity map and the two-dimensional bounding box output by the defect detection thread, the local relative deformation of the contact wire component relative to its own fixed base is calculated using the intra-frame local difference model. The intra-frame local difference model calculates the local relative deformation displacement of the component by the difference between the pixel coordinates of the defect target point and the pixel coordinates of its fixed base, combined with the physical baseline distance of the binocular image acquisition unit and the corresponding disparity value.

[0010] According to the above technical solution, the intra-frame local difference model calculates the lateral local relative deformation using the following formula:

[0011] In the formula, This refers to the lateral local relative deformation displacement. This represents the x-coordinate pixel value of the defect target point in the image coordinate system. To correspond to the x-coordinate pixel value of the fixed base in the image coordinate system, The distance between the first industrial camera and the second industrial camera is the physical baseline distance. The disparity value is the value corresponding to the defect target point.

[0012] According to the above technical solution, the hardware depth of the asynchronous dual-clock first-in-first-out queue inside the processor unit satisfies the following constraints:

[0013] In the formula, The minimum hardware depth of the first-in-first-out queue. For the number of ping-pong buffer groups, and These are the pixel width and pixel height of a single frame image, respectively. This refers to the number of bytes per pixel.

[0014] According to the above technical solution, the processor unit is further configured to perform a preprocessing pipeline on the original image acquired by the binocular image acquisition unit. The preprocessing pipeline includes, in sequence: black level correction based on factory calibration parameters, defect point compensation using a pre-stored noise mapping table, nonlocal mean denoising based on spatiotemporal domain joint, distortion correction based on calibration parameters, and iterative automatic exposure adjustment.

[0015] According to the above technical solution, the defect detection thread adopts a two-stage detection framework. The first stage uses a lightweight backbone network to generate candidate regions containing defect targets. The second stage classifies the candidate regions to output defect categories and performs boundary regression to output the two-dimensional bounding box. The stereo matching thread adopts a twin network architecture, extracts feature maps from the images acquired by the first and second industrial cameras respectively through a weight-sharing feature extractor, and generates the disparity map from the feature maps through a transposed convolutional decoder with skip connections.

[0016] According to the above technical solution, the intra-frame local difference model is further configured to calculate the vertical distance between the longitudinal local relative deformation and the defect target; the longitudinal local relative deformation is calculated based on the difference in the ordinate of the defect target point and the fixed base, the physical baseline distance, and the disparity value; the vertical distance is calculated based on the physical baseline distance, the pixel focal length of the first industrial camera, and the disparity value.

[0017] According to the above technical solution, the edge processing unit also integrates an image compression module. The image compression module adopts an end-to-end compression network based on deep learning. The loss function of the compression network superimposes a perceptual loss term constrained by the defect detection network on the basis of the reconstruction quality loss, so that the compressed image has an adjustable compression ratio while retaining the features required for defect detection.

[0018] According to the above technical solution, the enclosure also integrates a communication unit and a power management module; The communication unit is configured to: selectively operate in real-time transmission mode or local storage mode based on the comparison results of real-time monitored wireless channel quality parameters and preset conditions; in real-time transmission mode, upload defect detection results and compressed image data through the network; in local storage mode, store the data to be transmitted in the local storage module, detect the network status at fixed intervals, and perform historical data retransmission after the network recovers. The power management module has a dual-range input adaptive topology circuit, which locks the first safe operating range when the input voltage is detected to be the first nominal voltage, and automatically switches the transformer tap topology to adapt to the second safe operating range when the input voltage is detected to be the second nominal voltage through a solid-state electronic switch.

[0019] According to the above technical solution, the inspection system supports a multi-machine collaborative working mode. In the multi-machine collaborative mode, one set of inspection devices serves as the master control unit, and the rest serve as slave units. The master control unit and the slave units use a precise time protocol for time synchronization. The master control unit summarizes the detection results of each slave unit and performs cross-view fusion.

[0020] According to the above technical solution, it also includes a cloud server, which is configured to: receive defect detection results and image data uploaded by inspection terminals through a distributed message middleware; store structured data in a columnar distributed database and store image data in an object storage system with erasure coding fault tolerance; perform spatiotemporal consistency alignment of multi-source inspection data using a catenary geographic information system and generate a health status assessment report of the catenary facilities; and incrementally train the defect detection model using accumulated inspection data and remotely push the updated model to the edge processing unit via a wireless network.

[0021] Compared with the prior art, the present invention has the following beneficial effects: In this invention, by constructing an asynchronous dual-clock FIFO queue and a ping-pong frame buffer module within the processor unit, the burst data writing from the camera pixel clock domain and the pipelined reading from the system clock domain are decoupled in the time domain, fundamentally solving the timing mismatch problem between the high-speed acquisition frame period and the preprocessing time. Simultaneously, by using an edge computing processor to execute the defect detection thread and the stereo matching thread in parallel, and replacing the traditional serial overlay mode with a heterogeneous parallel scheduling strategy, the end-to-end processing latency is significantly reduced under limited edge computing power, ensuring real-time processing without frame loss at high-frequency acquisition rates. Regarding measurement accuracy, by constructing an intra-frame local difference model, the local relative deformation is calculated using the pixel coordinate difference between the defect target point and the corresponding fixed base, combined with the baseline distance and disparity value. This mathematically offsets the impact of transient drift of camera intrinsic parameters caused by train vibration on measurement accuracy, achieving accurate quantitative measurement of the deformation of catenary components under strong vibration conditions, filling the technological gap where monocular vision systems cannot acquire three-dimensional deformation information. In terms of system architecture, an integrated enclosure design is adopted for the binocular image acquisition unit and edge processing unit, integrating all data processing tasks within a single inspection terminal. This avoids the interconnection failure points and installation complexity associated with split designs, significantly improving system reliability in vehicle environments. Regarding data transmission, the communication unit automatically switches between real-time transmission mode and local storage mode based on real-time channel quality, and performs historical data retransmission after network recovery. This eliminates the rigid dependence of inspection tasks on wireless network quality, ensuring the integrity of inspection data in signal blind spots. Furthermore, a multi-layered security protection mechanism across the entire link ensures the confidentiality and integrity of inspection data. Incremental training of accumulated inspection data and remote model updates via a cloud server enable continuous iterative optimization of the defect detection algorithm's performance. Attached Figure Description

[0022] Figure 1 This is a flowchart of the security inspection system of the present invention; Figure 2 This is a schematic diagram of the composition of the binocular image acquisition unit of the present invention; Figure 3 This is a schematic diagram of the composition of the edge processing unit of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1

[0024] like Figure 1As shown, the contact network safety inspection system based on binocular detection described in this invention adopts an integrated enclosure design for its core hardware architecture, with all functional modules integrated within a unified protective enclosure. Specifically, the enclosure of the edge computing inspection terminal is integrally formed from high-strength aluminum alloy material, with external dimensions of 400mm × 300mm × 150mm. The surface of the enclosure undergoes anodizing treatment and anti-corrosion coating spraying, achieving an IP54 protection level, effectively preventing dust intrusion and liquid splashing. Furthermore, a shock-absorbing mounting bracket is fixedly installed at the bottom of the enclosure. This bracket uses a composite shock-absorbing structure composed of rubber shock-absorbing pads and metal springs, with its natural frequency designed between 15Hz and 25Hz. This effectively attenuates the vibration energy generated during train operation within the frequency range of 5Hz to 200Hz, preventing micron-level deformation of the physical baseline of the binocular camera due to high-frequency, high-amplitude vibrations, thereby ensuring the stability of spatial geometric measurements.

[0025] Specifically, such as Figure 2 As shown, the binocular image acquisition unit is located on the front panel of the enclosure. It includes two industrial cameras: a first industrial camera and a second industrial camera. The two industrial cameras are arranged side by side with their optical axes at a predetermined angle to achieve binocular field of view coverage of the same scene. Furthermore, both the first and second industrial cameras are global shutter type industrial cameras. The image sensors are CMOS photosensitive elements with a photosensitive surface size of 2 / 3 inch and an effective pixel count of 2448×2048. The raw image data size of a single frame is approximately 6MB. The maximum acquisition frame rate of the two industrial cameras is no less than 17 frames per second. Frame synchronization is achieved through a hardware trigger signal, which is generated by the processor unit within the edge processing unit. The trigger frequency is precisely matched with the camera's maximum frame rate. Preferably, the two industrial cameras are each equipped with a corresponding wide-angle lens. The focal length of the wide-angle lens is 8mm, the field of view is not less than 90 degrees, and the lens distortion rate is controlled within 3% at the image edges. The distance between the viewing axes of the two wide-angle lenses forms the baseline distance, which can be selected from 120mm to 200mm. In this embodiment, the baseline distance is selected as 160mm. The selection of the baseline distance is based on the depth of field requirements and measurement accuracy requirements of the catenary inspection scenario. Specifically, when the train speed is 300 km / h and the acquisition frame rate is 17 frames per second, the physical displacement distance between two adjacent frames is calculated to be approximately 4.9 meters by multiplying the time step by the train's physical speed. This displacement distance ensures that there is a field of view overlap rate of not less than 20% between adjacent frames in spatial distribution, thereby fundamentally avoiding the omission of key components of the catenary under high-speed travel conditions.

[0026] Furthermore, such as Figure 3As shown, the edge processing unit is located inside the enclosure and includes an edge computing processor, a processor unit, a DDR4 memory module, and a local storage module. Specifically, the edge computing processor uses a multi-core ARM architecture processor with a main frequency of 2.2GHz and 8 cores, integrating a neural network acceleration unit, achieving a peak computing power of 8 TOPS. The processor unit uses a large-scale programmable logic device, whose internal logic circuits implement functions such as camera trigger signal generation, image preprocessing pipeline, and feature extraction acceleration. The DDR4 memory module has a capacity of 16GB and a data transfer rate of 3200MT / s, used to support large-scale data caching during edge computing. The local storage module uses an NVMe interface solid-state drive with a capacity of 2TB, employing a cyclic overwrite strategy to store inspection data. Furthermore, the edge processing unit also integrates a catenary defect detection algorithm module, an image compression module, and a data encryption module. The catenary defect detection algorithm module is based on a deep convolutional neural network, with a lightweight backbone network architecture, a model parameter size of 6.5MB, and an inference time of no more than 35ms / frame. The input image size is 2448×2048 pixels, and the output results include defect category labels, defect region bounding box coordinates, and confidence scores. The image compression module uses a learning-based end-to-end compression algorithm, with a compression ratio adjustable from 1:5 to 1:10, and the peak signal-to-noise ratio of the compressed image is no less than 35dB. The data encryption module uses the AES-256-GCM symmetric encryption algorithm, with the encryption key stored inside a security chip that meets EAL5+ security level certification requirements.

[0027] Preferably, the communication unit is located inside the enclosure and includes a 4G / 5G communication module, a Wi-Fi module, and a train control system communication interface. The 4G / 5G communication module supports NSA and SA dual-mode networking, with a downlink peak rate of 1Gbps and an uplink peak rate of 200Mbps. The module integrates a multi-band antenna array with an antenna gain of 4dBi. The Wi-Fi module supports the 802.11ax protocol, operates on dual-band frequencies of 2.4GHz and 5GHz, and has a maximum transmission rate of 1.2Gbps. The train control system communication interface uses a CAN bus interface or an Ethernet interface, with the interface protocol conforming to the IEC61375 standard specification and a communication baud rate of 1Mbps. Furthermore, the display unit employs a 15.6-inch LCD screen with a resolution of 1920×1080 pixels, a brightness of 1200 nits, a contrast ratio of 1500:1, and a viewing angle of 178 degrees. The screen is equipped with a capacitive touch panel supporting multi-touch operation with a touch response time of 20ms. The display unit also integrates a graphics processing unit to drive the screen for real-time video preview, defect annotation overlay, inspection parameter display, and user interface rendering. Specifically, the enclosure is equipped with an active cooling system, including a cooling fan and a heat pipe cooling module. The cooling fan uses PWM speed control, and its speed is adaptively adjusted based on temperature sensor feedback from the edge processing unit. This cooling system ensures the edge processing unit operates stably and continuously in an ambient temperature not exceeding 50℃.

[0028] The data processing method of this invention mainly includes six core steps. In order to achieve real-time detection without frame loss in low computing power and high speed environment at the edge, this scheme constructs heterogeneous parallel control logic and proprietary mathematical model in the time domain and spatial domain.

[0029] Specifically, in step one: data acquisition and asynchronous spatial preprocessing, the two industrial cameras of the binocular image acquisition unit simultaneously acquire contact network images at a frame rate of 17 frames per second under the control of the frame synchronization trigger signal generated by the processor unit of the edge processing unit. The continuous data stream throughput generated by the two cameras is 204 megabytes per second. The acquired raw data is transmitted to the edge processing unit through dual independent GigEVision protocol network interfaces. The available bandwidth of the physical layer on the bus is reserved at no less than 240MB / s to ensure unblocked data flow. Furthermore, in order to solve the problem of data bursts and timing mismatches caused by short single-frame preprocessing time and long input frame period, this technical solution abandons the traditional serial bus waiting mechanism and innovatively constructs an asynchronous dual-clock FIFO queue and ping-pong frame buffer mechanism inside the processor unit. The data writing side clock is synchronized with the camera pixel clock to continuously stream burst pixel streams, and the data reading side clock uses the high-frequency system clock inside the processor unit. Preferably, in order to solve the problem of physical bus overflow caused by sudden influx of high-speed and high-volume traffic at the hardware level, this invention derives and proposes an asynchronous storage boundary topology constraint model by modeling the time-domain decoupling cross-clock behavior:

[0030] In the formula, The minimum hardware depth space of a circular queue of dedicated static random access memory configured within a processor unit, in bytes. ); The number of ping-pong cache groups participating in alternating streaming writes is limited in this embodiment. ; This represents the pixel width of a single frame image from a stereo camera, with a value of 2448 pixels. This represents the pixel height of a single frame image from a stereo camera, with a value of 2048 pixels. This represents the number of bytes for single-pixel bit-depth conversion; for the system's 8-bit grayscale image, the value is... The boundary constraint model calculates that the minimum hardware depth space required by the system is 10.02MB, thus providing quantization boundary support at the physical layer. Specifically, the preprocessing cascade pipeline sequentially performs black level correction, defect point compensation, denoising, distortion correction, and automatic exposure adjustment under the drive of the system's high-frequency clock. Further, in the black level correction stage, the black level correction module integrated within the processor unit compensates for and corrects the dark areas of each frame image according to the black level parameters calibrated at the camera's factory settings; in the defect point compensation stage, a fixed-pattern noise mapping table pre-embedded in flash memory is used, employing a neighborhood median adaptive operator to dynamically replace sensor dead pixels at the sub-pixel level; in the denoising stage, a nonlocal mean filtering algorithm based on spatiotemporal domain joint processing is used, establishing an 11×11 pixel search window in the spatial domain and introducing a circular queue in the temporal domain to buffer the three frames before and after the current frame for temporal joint weighted filtering. The filtering intensity parameter of its Gaussian weighted kernel function is defined as... In the distortion correction stage, real-time pixel coordinate mapping is performed using Brown-Conrady model camera calibration parameters, which include radial and tangential distortion coefficients. After correction, the absolute geometric distortion rate of the image edges is controlled within 0.5%. In the automatic exposure adjustment stage, an iterative exposure mapping operator is used to statistically analyze the histogram distribution of the current frame image in real time. Iterative feedback control of gain adjustment stabilizes the average image brightness. Within the specified range. The total preprocessing time for a single frame image is no more than 5ms. After processing, the data is transferred to the shared area of ​​the DDR4 memory module via the AXI bus.

[0031] Furthermore, in step two: edge defect identification with spatiotemporal overlap, in order to prevent the risk of frame loss caused by the total time approaching the frame period limit due to the serial execution of each operator, the edge computing processor runs a multi-threaded asynchronous control architecture based on spatiotemporal overlap pipeline. The preprocessed binocular image data is directly written to the shared memory area of ​​DDR4. The neural network acceleration unit is divided into different computing power execution clusters at the hardware layer. The defect detection thread and the stereo matching thread are synchronously launched in a parallel heterogeneous split mode. Specifically, the defect detection algorithm employs a two-stage detection framework. The first stage is a region proposal network, responsible for generating candidate regions that may contain defects on the feature map. This network uses a lightweight ResNet-18 as the underlying feature extraction backbone network. The input is a preprocessed image from the left camera, and the output is a 153×128 pixel feature map. It generates 2000 multi-scale adaptive anchor points on the feature map to predict the target probability value and coarse bounding box. The inference time of this stage is 13ms. Preferably, the second stage is a classification and regression network, responsible for fine classification and boundary regression of the candidate regions output by the region proposal network. The classification head outputs confidence scores for eight types of defects, including broken droppers, slack droppers, locator offset, insulator damage, insulator contamination, broken clues, loose clues, and corroded support structures. The regression head outputs the sub-pixel level precise bounding box coordinates of each candidate region in the two-dimensional image coordinate system. The inference time of this stage is 15ms. Furthermore, the edge processing unit performs post-processing on the detection results. The post-processing includes three stages: non-maximum suppression, confidence threshold filtering, and size filtering. The IoU threshold for non-maximum suppression is set to 0.5, the confidence filtering threshold is set to 0.7, and the size filtering is used to exclude positional noise regions with a size smaller than 32×32 pixels or larger than 512×512 pixels. The end-to-end inference time of the entire defect detection process is controlled within 30ms.

[0032] Specifically, in step three: stereo matching and quantitative reconstruction of 3D space, the stereo matching thread and the defect detection network operate in parallel on the underlying computing cluster without interference, and directly read the original preprocessed image from the DDR4 shared memory area. The stereo matching algorithm uses a Siamese network architecture to achieve end-to-end disparity estimation. The backbone network uses an improved ResNet-34 structure as a dual-path weight-sharing feature extractor and outputs a 32-channel downsampled feature map. The decoder part progressively upsamples and outputs the disparity map through four transposed convolutions with skip connections. The disparity inference time per frame is 12ms. Furthermore, the matching results are filtered for confidence using a left-right consistency check. When the absolute difference between the predicted disparity of a point in the left camera image and the inverse predicted disparity of the corresponding point in the right camera is greater than 1 pixel, it is determined to be an invalid disparity point. For invalid disparity regions, a local spatial plane fitting algorithm is used for second-order interpolation filling to ensure that the effective completeness of the final disparity map is not less than 99.5%. Furthermore, the vertical absolute depth of field distance of the defect target relative to the physical plane of the optical center of the first industrial camera is in principle inversely proportional to the pixel disparity value output by the twin estimation network, and is linearly scaled using the product of the calibrated horizontal pixel focal length of the camera and the physical baseline distance as the scaling operator. Its relative measurement error is constrained to within two percent of the measurement distance by the differential propagation model.

[0033] Preferably, in order to solve the core technical problems of camera intrinsic parameter physical drift caused by strong mechanical vibrations of 5Hz to 200Hz generated by trains traveling at high speeds (300km / h), the conflict between the magnitude of absolute depth-of-field positioning accuracy and the inability to directly obtain the geometric values ​​of microscopic deformation of catenary components, this invention, based on the forward-looking orientation characteristics of the optical axis of the vehicle-mounted binocular camera, abandons the conventional conversion method of directly subtracting absolute spatial coordinates, and instead proposes a vibration-resistant intra-frame local differential spatial quantization microscopic solution model based on the binocular differential geometric topological conjugate structure:

[0034] In the formula, It is the absolute quantity of the local relative deformation displacement of the key components of the overhead contact line under test in the lateral physical space, which has a clear physical unit in the millimeter range. ); The x-coordinate pixel value of the center point of the defect target edge output from step two in the pixel coordinate system of the first industrial camera, in pixels ( ); The x-coordinate pixel value of the center of the steady-state structure fixed base of the component, locked via semantic prior, in the pixel coordinate system of the first industrial camera, is given in pixels. ); The physical baseline distance between the two industrial cameras is fixed at 160mm in this embodiment; This is the local disparity value corresponding to the center pixel of the detection region, output by the stereo matching network, in pixels ( ). ).

[0035] Furthermore, as can be seen from the aforementioned original mathematical model, by performing conjugate cancellation of the pixel focal length term of the first industrial camera in the numerator and denominator, the relative deformation calculation system constructed in this invention fundamentally eliminates the strong dependence on the absolute working distance of the camera and the high-precision calibration parameters of the camera's intrinsic parameters. Specifically, when a train experiences axial or radial bumps under strong vibration conditions, causing a micrometer-level transient drift in the center of the camera's optical intrinsic parameters or the principal focal length, due to the difference operator... Within the same temporal physical frame, the system uses this model to achieve automatic offsetting and hardware-level cancellation of common-mode vibration noise. Furthermore, this ensures at the physical layer that even at a distance of ten meters with an absolute offset of approximately 20cm in absolute measurement depth of field, the measurement error of local deformation of components such as locator offset and suspension slack within a working distance of four to five meters at the near end can still be locked within 5cm. Thus, at the underlying mathematical mechanism, it guarantees the coordinated unity of large-scale safety defense positioning and local fine deformation measurement.

[0036] Furthermore, in step four: deep-sensing data compression and encryption, the edge processing unit compresses the defective images and structured data to be transmitted. Specifically, image compression employs an end-to-end compression neural network model based on deep learning. Its encoder consists of multi-level strided convolutions to extract compact latent feature vectors, and the decoder reconstructs a high-fidelity image from the latent feature vectors through a deconvolution structure. The network's loss function, based on the mean squared error, jointly introduces an adversarial perception loss term constrained by the defect feature detection network. The single-frame compression inference time is 8ms. In this embodiment, a typical compression ratio of 1:8 is used to ensure that the single-frame image size is no greater than 750KB and the peak signal-to-noise ratio of the reconstructed image is no less than 35dB. Preferably, metadata compression adopts a scheme combining dictionary encoding and entropy encoding. The structured text and the calculated three-dimensional spatial physical coordinates are efficiently dictionary encoded and entropy compressed using a nested arithmetic encoder. Furthermore, the compressed mixed binary data stream is sent to the data encryption module, which uses the AES-256-GCM algorithm for active encryption. The module calls the HKDF key derivation function to derive the temporary encryption working key and authentication key of the current frame from the built-in hardware master key and performs stream encryption. Finally, it is encapsulated into a standard data packet containing a frame header, encrypted payload and authentication tag.

[0037] Specifically, in step five: network state adaptive transmission and timing storage, the communication unit performs a fully closed-loop adaptive switching based on the current wireless channel's channel quality state machine. Further, when in real-time transmission mode, the communication unit continuously monitors the received signal strength indication and reference signal received power reported by the wireless module at 100ms intervals. When satisfied Furthermore, when the available uplink bandwidth of the baseband physical link assessment is greater than 25Mbps, the encrypted data packet is pushed to the cloud server via UDP protocol with breakpoint resumption control logic, and the end-to-end latency of the critical defect alarm is no greater than 500ms. Preferably, when a fault is detected... If the wireless link packet loss rate exceeds 15%, the communication unit automatically switches to local storage mode within 2 seconds, sequentially appending data packets to the dynamic inspection storage area of ​​the NVMe solid-state drive. This area is configured with clock-back-based cyclic overlay logic and retains complete inspection data for the most recent 72 hours by default. Simultaneously, the system activates a timed network detection suspension task with a period of 30 seconds. Once the detected signal power recovers to normal, the task will resume operation. If the duration exceeds 5 seconds, the system immediately initiates a background historical data adaptive retransmission process. Specifically, the train control system communication interface transmits key high-risk defect structured alarm information, with a length not exceeding 128 bytes, to the display terminal in the train driver's cab in real time via a backbone network conforming to the IEC61375 standard, with a local network transmission latency not exceeding 50ms.

[0038] Furthermore, in step six: Cloud-based Multi-Source Data Fusion and Incremental Evolution Analysis, the cloud server receives mixed encrypted data uploaded from multiple on-line inspection terminals. Specifically, the data receiving module adopts a distributed message middleware Kafka cluster architecture, with a single node physical throughput designed to be... The system processes messages per second, firstly by decrypting and verifying the authentication tags of the encapsulated data packets and extracting high-precision timestamps for time-series rearrangement. Furthermore, the data storage module employs a combination of distributed database and object storage. Structured defect records and 3D positional geometric parameters are stored in the distributed database ClickHouse for efficient spatial correlation queries and multi-dimensional micro-aggregation statistical analysis. Meanwhile, original images and keyframe images are stored in the object storage system MinIO. MinIO utilizes a physical erasure coding fault-tolerant encoding algorithm at its underlying layer to provide high data reliability, ensuring extremely high data loss prevention for the system. Preferably, the data fusion module introduces a digital spatial association algorithm based on the geographic information system of the entire catenary network. According to the spatial coordinates and physical numbers of characteristic towers, the catenary component data of the same section are spatiotemporally aligned and a first-order Markov chain state transition matrix is ​​established to generate a global health status assessment report. Furthermore, the model training module uses the accumulated inspection data to incrementally fine-tune the defect detection algorithm model on the edge side. The updated model uses the differential generation algorithm to output the model incremental patch file and pushes it to each vehicle edge processing unit through OTA remote upgrade, so as to realize the continuous iterative optimization of the algorithm model.

[0039] The inspection system of this invention is equipped with self-test and diagnostic function modules. These modules perform a comprehensive self-test process upon system power-up, including camera imaging quality detection, edge computing unit function detection, communication module signal detection, storage module read / write detection, and display module display detection. Specifically, during camera imaging quality detection, the edge processing unit controls the binocular image acquisition unit to acquire images of a standard test target. This standard test target uses an internationally standard 1920×1080 resolution transmissive target, with a target pattern including a checkerboard, dot matrix, and wedge-shaped bars. The edge processing unit analyzes the sharpness, distortion, and contrast parameters of the acquired images to determine if the camera is functioning correctly. Furthermore, the self-test results and diagnostic information during operation can be viewed through the display unit or exported and analyzed via a ground maintenance terminal after the train enters the depot. The diagnostic information includes the operating status of each module, temperature data, storage utilization rate, and communication quality indicators.

[0040] Preferably, the inspection system of the present invention adopts a layered architecture design at the software level, consisting of a hardware abstraction layer, a driver layer, an algorithm layer, a business logic layer, and an application layer, from bottom to top. Specifically, the hardware abstraction layer provides a unified hardware resource access interface, shielding the differences between different hardware vendors and encapsulating underlying functions; the driver layer includes software implementations of camera drivers, storage drivers, communication drivers, and display drivers, where the camera driver implements the GigEVision protocol stack, the storage driver implements the NVMe protocol driver stack, the communication driver implements network connection management, and the display driver implements the OpenGLES graphics interface; the algorithm layer encapsulates core algorithm modules such as defect detection, stereo matching, image compression, and data encryption, and the input and output data of each algorithm module adopt a unified Tensor format; the business logic layer implements logical functions such as inspection task management, data flow control, state machine scheduling, and exception handling, and maintains a state machine including initialization state, inspection state, standby state, and fault state; the application layer provides functions such as human-computer interaction interface, alarm display, and parameter configuration, and uses the QT framework to implement a cross-platform graphical interface. Furthermore, the layers communicate with each other through a standardized message bus. The message bus uses the D-Bus protocol to realize inter-process communication. When the system is upgraded, a differential update strategy can be used to realize hot updates of the algorithm layer or business logic layer without affecting the operation of other layers.

[0041] The inspection system of this invention supports a multi-machine collaborative working mode. When a train is equipped with multiple inspection devices simultaneously, the different inspection devices synchronize data and time through the train's internal network. The multi-machine collaborative mode can achieve multi-angle coverage and acquisition of the same overhead contact line section. Specifically, in the multi-machine collaborative mode, one inspection device acts as the master control unit, responsible for coordinating the working sequence of each subordinate unit. The time synchronization between the master control unit and the subordinate units adopts the PTP precise time protocol with a time synchronization accuracy of no more than 1 millisecond. The master control unit summarizes the detection results of each subordinate unit and performs cross-view result fusion.

[0042] Furthermore, the inspection system of this invention employs multi-layered security measures for data security. Specifically, the first layer is data acquisition security, which prevents data tampering and frame loss through a frame counting verification mechanism during camera data transmission; the second layer is data processing security, where the operating system of the edge processing unit adopts a secure boot mechanism and verifies the digital signatures at each boot stage, and the encryption key is stored in a hardware security module that meets the security level requirements; the third layer is data transmission security, where the communication link uses the TLS 1.3 protocol for end-to-end encrypted transmission, and the transmitted data is enhanced with an application-layer message authentication code; the fourth layer is data storage security, where the local storage module adopts a full-disk encryption scheme, using AES-256-XTS mode for encryption, and the encryption key is bound to the hardware security module.

[0043] The inspection system of this invention includes a power management module integrated into the edge processing unit. The power management module receives DC power from the train and converts it into the operating voltage required by each module of the system. Specifically, to accommodate the physical differences in the power systems of different train models, the power management module employs a dual-range input adaptive topology circuit at the hardware level. When the internal digital comparator triggers the high-range logic and identifies the input as standard DC 110V, the locked safe operating voltage range is set to DC 99V to DC 121V. Voltages below or above this range will trigger hardware-limited overvoltage protection. Furthermore, when the system automatically identifies the input power as DC 72V from urban rail or older train models, the system automatically switches the transformer tap topology via a solid-state electronic switch, adaptively adjusting the corresponding operating monitoring and safety protection range to DC 64.8V to DC 79.2V. This eliminates logical conflicts in the input voltage compatibility definition from the ground up. Preferably, the power management module has a built-in power monitoring circuit that monitors voltage stability and load conditions in real time. The power management module also supports a wake-up function, which can automatically power on when the train starts by receiving a wake-up command via the CAN bus, and automatically enter a low-power standby state after the train stops and the speed signal is 0 for more than 30 seconds.

[0044] This embodiment also provides a specific implementation method. This embodiment selects a section of the overhead contact system of a high-speed railway under the jurisdiction of a railway bureau for actual inspection testing. The test section is a 120-kilometer-long overhead contact system in both directions, with a train speed of 300 kilometers per hour. The inspection system used in the test is configured with an edge computing processor using an 8-core ARM processor and an integrated neural network acceleration unit with a computing power of 8 TOPS, an industrial camera using a 2448×2048 pixel global shutter CMOS camera with a frame rate of 17 frames per second, a baseline distance of 160mm, and a wide-angle lens with a focal length of 8mm. Specifically, the inspection system continuously inspected the 120-kilometer section and acquired a total of 1,468,800 frames. The edge computing processor successfully processed 1,465,200 frames without frame loss, and the system's global timing non-blocking frame loss rate was less than 0.24%. Furthermore, during the inspection, 23 slack suspension wire defects, 8 positioner misalignment defects, 15 insulator contamination defects, and 6 support structure corrosion defects were detected. All detected defects have been uploaded to a cloud server. Preferably, when the 4G / 5G signal is good, the average latency in real-time transmission mode is 380ms. When entering areas with poor signal coverage, such as tunnels, the system automatically switches to local storage mode within 2 seconds. After passing through the tunnel, the network recovers and historical data is adaptively retransmitted within 30 seconds. Specifically, the system's average power consumption is 85W in continuous operation and only 4.2W in standby low-power mode. Furthermore, the active cooling system keeps the junction temperature of the edge processor stable below 78℃ during load testing in an environment with an extreme external temperature of 50℃.

[0045] The embodiment also provides a specific implementation method. This comparative example introduces a monocular vision inspection system with identical hardware and software except for the camera configuration for parallel comparison testing. Due to the lack of physical baseline constraints, the monocular system only uses a conventional target detection algorithm to output the two-dimensional pixel coordinate boxes of the defect image. Specifically, the comparison results show that the binocular detection system of this invention improves the number of defects detected by approximately 35% compared to the monocular detection system.

[0046] Furthermore, regarding the slack cable defect, the binocular integrated system of this invention successfully detected 23 defects, while the comparative monocular inspection system detected only 18. For the positioner offset defect, the binocular integrated system of this invention detected 8 defects, while the comparative monocular inspection system detected only 6. Preferably, when detecting component deformation at a working distance of 4 to 5 meters, the binocular integrated system of this invention, due to the introduction of an anti-vibration local differential spatial quantization microscopic solution model, achieves a measured local relative deformation error of no more than 5 centimeters. In contrast, the comparative monocular inspection system, lacking depth and baseline transformation derivation information, cannot directly measure the relative displacement in physical units. Specifically, when performing large-scale spatial positioning at a working distance of 10 meters, the binocular integrated system of this invention, based on the physical depth-of-field inverse conversion, measures an absolute depth of field with an absolute error of approximately 20 centimeters, perfectly corresponding to the simulation performance of the aforementioned differential propagation model. The comparative monocular inspection system, similarly, cannot directly perform high-precision absolute depth of field measurement. Test results show that the present invention has successfully achieved a fundamental leap from simple qualitative image recognition of defects in contact wire components to high-precision three-dimensional quantitative geometric measurement.

[0047] The embodiment also provides a specific implementation method. This example compares the data transmission performance of different compression schemes, including the deep learning-based end-to-end compression scheme used in this invention, the traditional JPEG compression scheme, and the WebP compression scheme. Specifically, the test conditions involved acquiring 1000 frames of contact wire images containing typical defects and performing compression tests using the three compression schemes respectively. Further, the test data shows that, under the same compression ratio of 8:1 and a single frame data size of 750KB, the deep learning-based end-to-end compression scheme used in this invention achieves a peak signal-to-noise ratio of 38.2dB and a defect detection accuracy of 97.8%; while the WebP compression scheme achieves a peak signal-to-noise ratio of 35.6dB and a defect detection accuracy of 96.2%; and the JPEG compression scheme achieves a peak signal-to-noise ratio of 34.1dB and a defect detection accuracy of 95.4%. Preferably, the comparative data clearly shows that the deep learning end-to-end compression scheme adopted in this invention has a significantly higher peak signal-to-noise ratio than traditional compression schemes at the same compression ratio and better image quality. During the training process, it takes perceptual loss as one of the optimization objectives, thereby retaining more image features that are important for defect detection, so that the impact of the compression scheme on the accuracy of defect detection is minimized.

[0048] The embodiment also provides a specific implementation method. This example compares the system performance of edge computing processing mode and pure cloud processing mode. The test scenario is real-time inspection at a running speed of 300 kilometers per hour, with the system acquiring 17 frames of images per second. Specifically, the test results show that in edge computing processing mode, the total latency from image acquisition to defect detection result output does not exceed 35ms (typically 30ms), and the defect detection result can be displayed in real time on the local display unit with a display latency of no more than 100ms. Further, in cloud processing mode, the raw image data is uploaded to the cloud server via the network for defect detection. Since the raw throughput generated by the high-frequency acquisition of the two cameras is no less than 204 megabytes per second, while the actual available bandwidth of the wireless network is usually no more than 50Mbps, it cannot meet the real-time upload requirements of the raw images. Preferably, if the image resolution is reduced to 640×480 pixels to forcibly adapt to the bandwidth, although the cloud processing mode can be achieved, the defect detection accuracy drops significantly to 89.3% after reducing the resolution, which is 8.8 percentage points lower than the edge computing mode of the present invention. The above comparative data shows that the edge computing architecture of this invention can efficiently complete data processing and solve the three-dimensional spatial coordinates of defects on the edge, reducing the dependence on communication bandwidth and ensuring the accuracy of defect detection. Example 2

[0049] Building upon the aforementioned edge processing unit, this embodiment further explores and constructs a full-time-domain adaptive energy efficiency scheduling and control mechanism to address the challenges of computing power contention, localized overheating, and rooftop power supply fluctuations caused by high concurrency of multiple models under complex high-speed rail operating conditions. Specifically, the neural network acceleration unit integrated within the edge computing processor is decoupled into multiple independent and controllable computing power clusters at the hardware layer. By capturing the transient changes in train speed and the complexity of the current image scene in real time, the operating frequency and task load weight of each computing power cluster are dynamically adjusted. Preferably, when the train decelerates in complex areas such as line hubs and stations, the required temporal sampling rate decreases accordingly. The system thus establishes a collaborative optimization mathematical model based on frame period limits and computing power resource allocation.

[0050] Specifically, the total computing power resources of the neural network acceleration unit are defined as follows: The unit is TOPS. The system dynamically allocates computing resources into computing power allocation factors responsible for defect detection. The computing power allocation factor responsible for stereo matching And the computing power allocation factor responsible for image compression And each allocation factor satisfies the normalization constraint. Furthermore, the total computational cost required for defect detection, stereo matching, and end-to-end compression in a single frame image is defined as follows: , and The unit is Terra-Operations. Therefore, the total system inference delay... The physical mathematical model is specifically expressed as follows:

[0051] Furthermore, in order to ensure at any train speed Without timing blockages or frame drops in the system, the total inference latency must be limited by the maximum allowable time step between two adjacent frames. The rigid boundary constraints. Preferably, combining the inter-frame displacement and field-of-view overlap constraints derived in step one, the temporal bounded adaptive control model is defined as:

[0052] In the formula, The actual longitudinal physical field of view boundary of the binocular camera in the direction of train travel, in meters ( ); To ensure that no component is missed during inspection, the minimum field-of-view overlap rate of adjacent frames is set to 20% in this embodiment.

[0053] Specifically, the scheduling and monitoring module inside the edge processing unit dynamically acquires the speed signal from the onboard control system using a 100ms periodic clock. When the train's speed is affected by shunting at a station... When decreasing, the time step limit To broaden the scope, the scheduling module dynamically reduces the global clock frequency of the neural network acceleration unit by solving the duty cycle solution space of the aforementioned inequality model. While ensuring operational timing safety, it also shuts down some idle computing clusters to achieve hardware-level gating power saving. Furthermore, this adaptive energy-efficient scheduling control mechanism significantly reduces the instantaneous power consumption of the edge processing unit, effectively attenuating the temperature rise rate of local chips and extending the mean time between failures (MTBF) of the system within its enclosed integrated aluminum alloy enclosure. Example 3

[0054] Building upon the aforementioned binocular stereo matching and 3D quantitative reconstruction, this embodiment further explores and constructs a cross-frame spatiotemporal consistency false alarm suppression mechanism based on rigid body motion projection transformation to address the challenges of occasional flickering false alarms and local disparity matching edge overflow under strong train vibration conditions. Specifically, this scheme, based on the geometric characteristics of the forward optical axis of the binocular camera, transforms the train's forward motion into a dynamic transition along the depth direction of the optical axis in the camera coordinate system. It utilizes known physical displacements as geometric priors to establish spatiotemporal filtering constraints, thereby eliminating discontinuous transient false alarms.

[0055] Specifically, let's assume that in the current... In the frame image, steps two and three jointly output the center point of a suspected defective target, whose calculated coordinates in the three-dimensional physical space coordinate system of the first industrial camera are: In the formula The horizontal physical distance. The vertical physical distance. This represents the absolute depth of field physical distance. Furthermore, the time step between adjacent frames is known. Inside, the train, equipped with an inspection terminal, moves forward along the track, and the physical displacement of the relative rigid body translation in the direction of the optical axis depth of field is: Preferably, in order to accurately predict the perspective scaling caused by movement, this technical solution constructs a dynamic depth-of-field compensated cross-frame perspective projection model to predict the defective target in the next time frame (i.e., the first frame). Theoretical projection position in the pixel coordinate system of the frame image :

[0056]

[0057] In the formula, For the predicted first The theoretical projection value of the horizontal coordinate in the pixel coordinate system of the frame image, in pixels ( ); For the predicted first The theoretical projection value of the ordinate in the pixel coordinate system of the frame image, in pixels ( ); and These are the horizontal and vertical pixel focal lengths determined by the calibration of the first industrial camera, in pixels. ); and These are the pixel coordinates of the principal optical center point of the first industrial camera, in pixels. ).

[0058] Specifically, when the system collects high-frequency data and executes the data up to the [number]th [time]... At frame rate, the edge defect detection thread produces the actual detection bounding boxes in the new frame image. The scheduling module predicts pixel positions theoretically. With the The center point of the defect area actually detected by the frame Euclidean geometric distance between The Euclidean geometric distance satisfies the calculation formula:

[0059] In the formula, The pixel geometric distance between the predicted point and the actual detection point across frames, in pixels ( ); and The first The x and y coordinates of the center point of the actual detection bounding box in the frame are measured in pixels. Furthermore, the system sets a cross-frame spatiotemporal geometric intersection and union distance threshold. In the formula The maximum permissible pixel deviation limit for rigid body constraint consistency is determined if and only if the inequality constraint is satisfied. When the system determines that the target satisfies the rigid body motion consistency characteristics in both the time and spatial domains, it is confirmed as a real physical defect and assigned a positive state tracking weight to the target. If the geometric distance is severely discrete and does not satisfy this inequality, it is determined to be a transient false alarm point, and the system automatically performs hardware-level interception and removes it. Furthermore, this model establishes a physical filtering barrier for dynamic noise at the algorithm's underlying layer, greatly improving the accuracy and robustness of security inspection alarms.

[0060] Specifically, based on the technical features and test performance of the above embodiments, the key parameters and technical specifications of the system of the present invention are specifically defined as follows: The effective resolution of the industrial camera in the binocular image acquisition unit is not less than 2448×2048 pixels, its acquisition frame rate is constant at 17 frames / second, the baseline distance between the two cameras is selectable between 120mm and 200mm, the equipped wide-angle lens has a focal length of 8mm and a physical field of view of not less than 90 degrees; the processor frequency of the edge computing processor in the edge processing unit is not less than 2.0GHz, the number of physical cores of the processor is not less than 8, and its internally integrated neural network acceleration unit provides a peak computing power of not less than 8 TOPS, thereby ensuring that the end-to-end inference time of a single high-dimensional network model is not greater than 35ms / frame; the memory capacity of the DDR4 memory module is not less than 1... The local storage module uses an NVMe solid-state drive with a local storage capacity of no less than 2TB; the image compression module has an adjustable compression ratio range of 1:5 to 1:10, and the lowest peak signal-to-noise ratio of the compressed image is no less than 35dB; the downlink peak rate of the 4G / 5G communication module in the communication unit is no less than 1Gbps, and the uplink peak rate is no less than 200Mbps; the physical display size of the LCD screen in the display unit is 15.6 inches, and the display resolution is no less than 1920×1080 pixels; the enclosure protection level of the edge computing inspection terminal is no less than IP54, and its internal active cooling system can ensure continuous and stable operation of the system in an operating environment temperature not exceeding 50℃, and the effective vibration attenuation frequency range of the shock-absorbing mounting bracket at the bottom of the enclosure is 5Hz to 200Hz. Furthermore, the aforementioned continuous and logical parameter constraints together constitute the technical closed loop of the present invention, effectively solving the technical problems existing in the current catenary inspection system, such as bandwidth bottlenecks caused by cloud-based data processing architecture, inability of monocular vision to acquire depth and accurate three-dimensional positioning information, insufficient reliability of split design, and poor real-time performance of data analysis.

[0061] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0062] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An integrated contact network safety inspection system based on binocular detection, comprising an integrated housing, wherein the housing integrates a binocular image acquisition unit, an edge processing unit, and a display unit; the binocular image acquisition unit includes a first industrial camera and a second industrial camera for simultaneously acquiring contact network images; the edge processing unit includes a processor unit and an edge computing processor; characterized in that: The processor unit is internally equipped with an asynchronous dual-clock first-in-first-out queue and a ping-pong frame buffer module. The write-side clock of the asynchronous dual-clock first-in-first-out queue is synchronized with the pixel clock of the industrial camera, and the read-side clock is synchronized with the internal system clock of the processor unit. The ping-pong frame buffer module is used to temporarily store image data to decouple the timing between the acquisition frame period and the preprocessing time. The edge computing processor is configured to execute a defect detection thread and a stereo matching thread in parallel. The defect detection thread is used to detect defects in the contact wire components of the image and output the defect category and a two-dimensional bounding box. The stereo matching thread is used to perform stereo matching on the image to generate a disparity map. Based on the disparity value of the defect region in the disparity map and the two-dimensional bounding box output by the defect detection thread, the local relative deformation of the contact wire component relative to its own fixed base is calculated using the intra-frame local difference model. The intra-frame local difference model calculates the local relative deformation displacement of the component by the difference between the pixel coordinates of the defect target point and the pixel coordinates of its fixed base, combined with the physical baseline distance of the binocular image acquisition unit and the corresponding disparity value.

2. The integrated contact wire safety inspection system based on binocular detection according to claim 1, characterized in that: The intra-frame local difference model calculates the lateral local relative deformation using the following formula: In the formula, This refers to the lateral local relative deformation displacement. This represents the x-coordinate pixel value of the defect target point in the image coordinate system. To correspond to the x-coordinate pixel value of the fixed base in the image coordinate system, The distance between the first industrial camera and the second industrial camera is the physical baseline distance. The disparity value is the value corresponding to the defect target point.

3. The integrated contact wire safety inspection system based on binocular detection according to claim 2, characterized in that: The hardware depth of the asynchronous dual-clock first-in-first-out queue within the processor unit satisfies the following constraints: In the formula, The minimum hardware depth of the first-in-first-out queue. For the number of ping-pong buffer groups, and These are the pixel width and pixel height of a single frame image, respectively. This refers to the number of bytes per pixel.

4. The integrated contact wire safety inspection system based on binocular detection according to claim 1, characterized in that: The processor unit is further configured to perform a preprocessing pipeline on the raw images acquired by the binocular image acquisition unit. The preprocessing pipeline includes, in sequence: black level correction based on factory calibration parameters, defect point compensation using a pre-stored noise mapping table, nonlocal mean denoising based on spatiotemporal domain joint, distortion correction based on calibration parameters, and iterative automatic exposure adjustment.

5. The integrated contact wire safety inspection system based on binocular detection according to claim 1, characterized in that: The defect detection thread adopts a two-stage detection framework. The first stage uses a lightweight backbone network to generate candidate regions containing defect targets. The second stage classifies the candidate regions to output defect categories and performs boundary regression to output the two-dimensional bounding boxes. The stereo matching thread adopts a twin network architecture, extracts feature maps from the images acquired by the first and second industrial cameras respectively through a weight-sharing feature extractor, and generates the disparity map from the feature maps through a transposed convolutional decoder with skip connections.

6. The integrated contact wire safety inspection system based on binocular detection according to claim 1, characterized in that: The intra-frame local difference model is further configured to calculate the longitudinal local relative deformation and the vertical distance between the defect target and the target; the longitudinal local relative deformation is calculated based on the difference in the ordinate of the defect target point and the fixed base, the physical baseline distance, and the disparity value; the vertical distance is calculated based on the physical baseline distance, the pixel focal length of the first industrial camera, and the disparity value.

7. The integrated contact wire safety inspection system based on binocular detection according to claim 1, characterized in that: The edge processing unit also integrates an image compression module, which adopts an end-to-end compression network based on deep learning. The loss function of this compression network superimposes a perceptual loss term constrained by the defect detection network on the basis of reconstruction quality loss, so that the compressed image has an adjustable compression ratio while retaining the features required for defect detection.

8. The integrated contact wire safety inspection system based on binocular detection according to claim 1, characterized in that: The enclosure also integrates a communication unit and a power management module; The communication unit is configured to selectively operate in real-time transmission mode or local storage mode based on the comparison results of real-time monitored wireless channel quality parameters and preset conditions. In real-time transmission mode, the defect detection results and compressed image data are uploaded via the network; In local storage mode, the data to be transmitted is stored in the local storage module, and the network status is checked at fixed intervals. After the network is restored, the historical data is retransmitted. The power management module has a dual-range input adaptive topology circuit, which locks the first safe operating range when the input voltage is detected to be the first nominal voltage, and automatically switches the transformer tap topology to adapt to the second safe operating range when the input voltage is detected to be the second nominal voltage through a solid-state electronic switch.

9. The integrated contact wire safety inspection system based on binocular detection according to claim 1, characterized in that: The inspection system supports a multi-machine collaborative working mode. In this mode, one inspection device serves as the master control unit, while the others serve as slave units. The master control unit and the slave units synchronize their time using a precise time protocol. The master control unit aggregates the detection results from each slave unit and performs cross-view fusion.

10. A binocular detection-based integrated contact wire safety inspection system with simultaneous shooting and inspection according to any one of claims 1 to 9, characterized in that: It also includes a cloud server, which is configured to: receive defect detection results and image data uploaded by inspection terminals through a distributed message middleware; store structured data in a columnar distributed database and store image data in an object storage system with erasure coding fault tolerance; and use the catenary geographic information system to perform spatiotemporal consistency alignment of multi-source inspection data and generate a health status assessment report of the catenary facilities. Furthermore, the accumulated inspection data is used to incrementally train the defect detection model, and the updated model is remotely pushed to the edge processing unit via a wireless network.

Images