An LED screen fault diagnosis method and system based on edge computing

Abstract

This invention relates to the field of LED screen fault diagnosis technology, specifically to an LED screen fault diagnosis method and system based on edge computing. The method includes: collecting low-level cable data at the edge node of the receiving card to generate a diagnostic dataset; pre-setting a fault fingerprint database containing faults such as oxidation, loosening, and copper foil micro-cracks; extracting feature parameters, clustering them, and mapping them to the fault fingerprint database to obtain fault type identification results; based on a preset reverse inference rule for the fault type, comparing it with the feature data to obtain the reverse inference fit, selecting fault types that meet the threshold as candidates, and extracting feature data from adjacent modules and comparing it with the propagation rule to confirm the target fault; constructing a fingerprint association topology map based on the cascade order of the screen modules, reverse inferring the theoretical values ​​of the feature parameters and comparing them with the actual values, and determining the interface with the smallest deviation as the fault point. This invention collects low-level cable data at the edge node, combines a fingerprint database with reverse tracing to identify cable faults, and uses topology positioning to achieve interface-level fault location.

Description

Technical Field

[0001] This invention relates to the field of LED screen fault diagnosis technology, and more specifically, to an LED screen fault diagnosis method and system based on edge computing. Background Technology

[0002] An LED screen is composed of multiple modules connected by cascading ribbon cables. The ribbon cables are responsible for data transmission between hardware units at each level, and their connection reliability directly affects the display effect.

[0003] In actual operation, cable connection faults such as poor contact, oxidation, loosening, and signal attenuation are common and difficult-to-diagnose types of faults. These faults manifest as localized speckles, flickering lines, or image tearing on the display screen, and often exhibit intermittent characteristics.

[0004] Image anomalies caused by cable malfunctions are visually very similar to those caused by driver chip malfunctions or power fluctuations, making it difficult to accurately identify the root cause of the problem.

[0005] To address the aforementioned issues, existing technologies typically employ cloud-based diagnostic methods based on image analysis. This involves installing a camera outside the screen to capture the displayed images, uploading the image data to a cloud server, using a deep learning model to identify the fault type, and locating the fault area using heatmap technology.

[0006] However, the above methods have certain limitations: the diagnostic model relies solely on the displayed screen and cannot obtain the underlying state data of the ribbon cable transmission link, including CRC bit error rate, LVDS differential signal voltage, signal rise time, ribbon cable characteristic impedance, etc. It can only identify the surface symptoms of the fault and cannot distinguish the root causes such as poor ribbon cable contact, driver chip failure, power fluctuations, etc., resulting in ambiguous diagnostic results. Maintenance personnel still need to check each cable one by one, which cannot quickly and effectively locate the fault; especially for large-screen LED screens, where there are many ribbon cables and the fault points are scattered, relying solely on screen diagnosis is insufficient to meet the operation and maintenance needs.

[0007] Therefore, there is an urgent need for a diagnostic method for LED screens that can monitor the status of the ribbon cable link in real time and effectively identify ribbon cable connection faults. Summary of the Invention

[0008] One objective of this invention is to provide a fault diagnosis method for LED screens based on edge computing. This method involves deploying edge nodes at the screen's receiving card to collect underlying status data from the cable transmission, establishing a fault fingerprint database, identifying cable connection faults through clustering, confirming the root cause of the fault by comparing it with data from adjacent modules using inverse reasoning, and finally locating the fault interface by combining the module topology map. This solves the problems mentioned in the background section, namely: Image anomalies caused by LED screen cable connection failures are visually similar to other faults. Furthermore, existing cloud-based diagnostic methods cannot obtain the underlying status data of the cable transmission, make it difficult to distinguish the root cause of the fault, and cannot quickly and accurately locate the fault, thus failing to meet the maintenance needs of large-screen LED screens.

[0009] To achieve the above objective, the method includes the following steps: S1. Deploy acquisition terminals at the edge nodes corresponding to each receiving card of the LED screen to collect physical layer data transmitted via ribbon cable, module operating parameter data and lightweight visual feature data. After processing, generate the edge end raw diagnostic dataset containing feature data of each module. S2. Pre-set a fault fingerprint database at the edge nodes, which includes three types of faults: oxidation faults, loosening faults, and copper foil micro-crack faults. Feature parameters were extracted from the original diagnostic dataset at the edge and clustered into normal fingerprint clusters and suspected faulty fingerprint clusters. The feature data of suspected fault fingerprint clusters are mapped to the fault fingerprint database to obtain the fault type identification results and confidence level; S3. Based on the fault type, preset the fault back-inference rule and back-inference fit threshold, compare the fault back-inference rule with the feature data of the suspected fault fingerprint cluster to obtain the back-inference fit, and take the fault type that meets the back-inference fit threshold as the candidate fault. The feature data of adjacent modules are extracted from the original diagnostic dataset at the edge and compared with the propagation pattern of candidate faults. When the degree of fit of the propagation pattern reaches the preset verification threshold, the target fault is confirmed. S4. Construct a fingerprint association topology diagram based on the cascading order of the screen modules; The fingerprint clustering results are bound to the module positions in the topology map. The starting module is located by traversing the topology map according to the target fault. The theoretical values ​​of the feature parameters of the fault point are inferred from the propagation law of the target fault. The values ​​are compared with the feature parameters in the original diagnostic dataset at the edge. The interface with the smallest deviation is determined as the fault point, and the fault location information is output.

[0010] In the above technical solution, diagnostic capabilities are extended to the edge nodes of the receiving card, directly collecting the underlying state data of the physical layer of the ribbon cable transmission. This is because the essence of ribbon cable connection failure is abnormal signal transmission at the physical link layer. Only by obtaining this data from the hardware layer can the three types of failures—oxidation, loosening, and copper foil micro-cracks—be fundamentally distinguished. After initially identifying the failure type by constructing a fault fingerprint database and using clustering algorithms, a reverse tracing mechanism is further designed. Based on the physical mechanism of the failure, the characteristic change law is deduced in reverse and compared with actual data. Fault types with a reverse deduction matching degree reaching a threshold are selected as candidates, solving the problem of not being able to confirm the root cause of the failure when the confidence of the clustering results is insufficient. The feature data comparison of adjacent modules is introduced. Taking advantage of the difference in the propagation law between ribbon cable failures, which are limited to a single interface, and power supply failures, which affect multiple modules, the propagation law matching degree is used to verify and exclude other hardware failures, thereby confirming the target failure. Finally, a fingerprint association topology map is constructed based on the cascade relationship of the screen modules. By deducing the theoretical values ​​of feature parameters in reverse and comparing them with the actual values, the landing point of the interface with the smallest deviation is located. Without collecting underlying data, it is impossible to distinguish between cable faults and other faults; without reverse tracing, it is impossible to confirm fuzzy fingerprints; without verification from adjacent modules, the fault type may be misjudged; and without topology positioning, it is impossible to provide a specific repair location.

[0011] Based on this, an improved lightweight K-means clustering algorithm is used to cluster the feature parameters extracted from the original diagnostic dataset at the edge. The similarity threshold during the clustering process is dynamically adjusted according to the degree to which the characteristic impedance of the cabling in the original diagnostic dataset at the edge deviates from the standard value, and the clustering response time does not exceed 3 milliseconds.

[0012] In another technical solution, the feature parameters extracted from the original diagnostic dataset at the edge include CRC bit error rate, LVDS differential signal voltage, ribbon cable characteristic impedance, module port temperature, pixel deviation rate, and number of line synchronization signal loss. The feature parameters are compressed to 4 dimensions through a feature dimensionality reduction algorithm. The core features retained after dimensionality reduction include CRC bit error rate, LVDS differential signal voltage, ribbon cable characteristic impedance, and module port temperature. The compressed feature parameters participate in clustering operations.

[0013] This technical solution addresses the limited computing power of edge nodes by optimizing the clustering algorithm. An improved lightweight K-means clustering algorithm is employed, with the similarity threshold dynamically adjusted based on the deviation of the cable characteristic impedance from the standard value. This is because the cable characteristic impedance is the most direct parameter reflecting connection reliability; the greater the deviation, the more lenient the clustering conditions should be to avoid missed detections due to parameter shifts. Furthermore, a feature dimensionality reduction algorithm compresses the 8-dimensional feature parameters to 4 dimensions, retaining four core features: CRC bit error rate, LVDS differential signal voltage, cable characteristic impedance, and module port temperature. This is because CRC bit error rate and LVDS differential signal voltage directly reflect signal transmission quality, cable characteristic impedance reflects physical connection status, and module port temperature reflects heat generation caused by changes in contact resistance. These four features are sufficient to distinguish between three types of faults: oxidation, loosening, and copper foil micro-cracks. Pixel deviation rate and the number of lost line synchronization signals are used as auxiliary features and can be temporarily processed for dimensionality reduction during the clustering stage. Without dynamic threshold adjustment, fixed thresholds are prone to missed detection when parameters deviate significantly; without feature dimensionality reduction, direct clustering of 8-dimensional features will cause the number of parameters to exceed the computing power of edge nodes, resulting in excessively long response times that cannot meet the requirements of real-time diagnosis.

[0014] The second objective of this invention is to provide an LED screen fault diagnosis system based on edge computing, including an edge data acquisition module, a fault fingerprint clustering module, a dynamic source tracing and reverse inference module, a topology positioning module, a self-learning optimization module, and a cloud collaboration module. The edge data acquisition module is deployed at the edge node corresponding to each receiving card of the LED screen to collect physical layer data transmitted by the ribbon cable, module operating parameter data and lightweight visual feature data. After filtering, denoising and normalization, the original edge diagnostic dataset is generated. The fault fingerprint clustering module is used to pre-set a fault fingerprint database containing oxidation faults, loosening faults and copper foil micro-crack faults in the local memory of the edge node, extract feature parameters from the original diagnostic dataset at the edge end and cluster them into normal fingerprint clusters and suspected fault fingerprint clusters, and map the feature data of the suspected fault fingerprint clusters to the fault fingerprint database to obtain the fault type identification result and confidence level. The dynamic source tracing and reverse inference module is used to preset the fault reverse inference rule and the reverse inference fit threshold according to the fault type. It compares the fault reverse inference rule with the feature data of the suspected fault fingerprint cluster to obtain the reverse inference fit. The fault type that meets the reverse inference fit threshold is used as the candidate fault. It extracts the feature data of the adjacent module from the original diagnostic dataset at the edge end and compares it with the propagation rule of the candidate fault. When the propagation rule fit reaches the preset verification threshold, the target fault is confirmed. The topology localization module is used to construct a fingerprint association topology map according to the cascading order of the screen modules, bind the fingerprint clustering results with the module positions in the topology map, locate the starting module according to the target fault traversal topology map, back-calculate the theoretical value of the feature parameters of the fault point according to the propagation law of the target fault, compare it with the feature parameters in the original diagnostic dataset at the edge end, determine the interface with the smallest deviation as the fault point, and output the fault location information. The self-learning optimization module is used to store the entire process data as fault cases. When the number of cases of the same fault type reaches a preset number, common features are extracted to update the fault fingerprint database, and the parameters of fault back-inference and propagation are calibrated in combination with operation and maintenance feedback. The cloud collaboration module is used to report optimized parameters and new fault cases to the cloud platform when the network is connected, and to cache the reported data when the network is interrupted and upload it asynchronously after the network is restored.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention extends diagnostic capabilities to the edge nodes of the LED screen receiver card, directly collecting low-level state data such as CRC error rate, signal voltage, and characteristic impedance of the ribbon cable transmission physical layer. Through a fault fingerprint database and clustering algorithm, it achieves preliminary identification of three types of ribbon cable connection faults: oxidation, loosening, and micro-cracks in the copper foil. Building upon this, a reverse tracing mechanism is further designed. Based on the physical mechanism of the fault, the characteristic change patterns are deduced in reverse and compared with actual data. This solves the problem of not being able to confirm the root cause of the fault when the confidence level of the clustering results is insufficient. Simultaneously, it introduces the comparison of feature data from adjacent modules. Utilizing the difference in propagation patterns between ribbon cable faults, which are limited to a single interface, and power supply faults, which affect multiple modules, the similarity in propagation patterns is used to verify and exclude other hardware faults, thus achieving effective differentiation between ribbon cable connection faults and other hardware faults.

[0016] 2. This invention constructs a fingerprint association topology map based on the cascade relationship of screen modules, binds the fingerprint clustering results to the topology nodes, and achieves the location of the interface with the smallest deviation by reverse-engineering the theoretical values ​​of feature parameters and comparing them with the actual values, thereby improving the fault location accuracy from the module level of traditional methods to the interface level. Simultaneously, addressing the limited computing power of edge nodes, a feature dimensionality reduction algorithm is used to compress 8-dimensional features into 4-dimensional features while retaining core features. A lightweight clustering algorithm is designed to dynamically adjust the similarity threshold according to the deviation of the cable characteristic impedance, keeping the clustering response time within 3 milliseconds. This ensures the feasibility of the diagnostic method running in real time on edge nodes, solves the diagnostic blind spot problem caused by network interruptions in existing cloud-based diagnostic solutions, and meets the autonomous operation and maintenance needs of large-screen LED screens in complex network environments. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall structure of the LED screen fault diagnosis method based on edge computing of the present invention. Figure 2 This is a flowchart of step S2 of the present invention; Figure 3 This is a flowchart of step S3 of the present invention. Detailed Implementation

[0018] 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.

[0019] Here are some explanations of technical terms: CRC, or Cyclic Redundancy Check, is a technique used to detect errors in data transmission. It generates a check code by performing a polynomial calculation on the data at the sending end, and then recalculates the check code at the receiving end and compares it with the sending end's check code. If they do not match, it is determined that an error has occurred in the data transmission process. LVDS, or Low Voltage Differential Signaling, is an interface technology for high-speed data transmission. It transmits data through the voltage difference between two signal lines and has the advantages of strong anti-interference ability and low power consumption. It is widely used in LED displays for transmitting video signals via ribbon cables. Sobel edge detection is a gradient-based edge detection algorithm that uses two convolutional kernels in the horizontal and vertical directions to calculate the gradient magnitude of the image, thereby extracting the edge features of the image. Local Binary Pattern (LBP) is an algorithm used to describe local texture features of an image. It generates a binary code by comparing the gray values ​​of the center pixel with those of its neighboring pixels, and then converts the binary sequence into a decimal number as the texture feature value of that pixel.

[0020] Currently, image anomalies caused by LED screen cable connection failures are visually similar to other faults. Furthermore, existing cloud-based diagnostic methods cannot obtain the underlying status data of the cable transmission, making it difficult to distinguish the root cause of the fault and accurately locate it, thus failing to meet the maintenance needs of large-screen LED screens. This invention provides an LED screen fault diagnosis method based on edge computing. (See [link to relevant documentation]). Figure 1 As shown, it includes the following steps: S1. Deploy acquisition terminals at the edge nodes corresponding to each receiving card of the LED screen to collect physical layer data transmitted via ribbon cable, module operating parameter data and lightweight visual feature data. After processing, generate the edge end raw diagnostic dataset containing feature data of each module. S2. Pre-set a fault fingerprint database at the edge nodes, which includes three types of faults: oxidation faults, loosening faults, and copper foil micro-crack faults. Feature parameters were extracted from the original diagnostic dataset at the edge and clustered into normal fingerprint clusters and suspected faulty fingerprint clusters. The feature data of suspected fault fingerprint clusters are mapped to the fault fingerprint database to obtain the fault type identification results and confidence level; S3. Based on the fault type, preset the fault back-inference rule and back-inference fit threshold, compare the fault back-inference rule with the feature data of the suspected fault fingerprint cluster to obtain the back-inference fit, and take the fault type that meets the back-inference fit threshold as the candidate fault. The feature data of adjacent modules are extracted from the original diagnostic dataset at the edge and compared with the propagation pattern of candidate faults. When the degree of fit of the propagation pattern reaches the preset verification threshold, the target fault is confirmed. S4. Construct a fingerprint association topology diagram based on the cascading order of the screen modules; The fingerprint clustering results are bound to the module positions in the topology map. The starting module is located by traversing the topology map according to the target fault. The theoretical values ​​of the feature parameters of the fault point are inferred from the propagation law of the target fault. The values ​​are compared with the feature parameters in the original diagnostic dataset at the edge. The interface with the smallest deviation is determined as the fault point, and the fault location information is output.

[0021] By using fault fingerprint clustering, dynamic source tracing and fingerprint association positioning logic, combined with multi-source data collection at the edge and full-link self-learning optimization, we can achieve rapid and accurate identification of LED screen cable faults, interface-level positioning and self-optimization of diagnostic methods, effectively improving diagnostic efficiency. At the same time, we can accurately distinguish the root cause of cable faults from other faults, ensuring the stable and efficient operation of LED screens.

[0022] Step S1 of this invention involves multi-source data acquisition and standardization processing at the edge, which forms the data foundation for subsequent diagnostic steps. First, a data acquisition terminal is deployed at the edge node corresponding to each receiver card on the LED screen. This edge node integrates an edge data acquisition module and an edge computing unit.

[0023] The deployment density of the acquisition terminals matches the cascading density of the screens, meaning one acquisition terminal is deployed for each receiver card, and each acquisition terminal covers a maximum of six cabinet modules. This avoids signal attenuation issues caused by excessively long transmission distances, which could affect data acquisition accuracy. The acquisition terminals collect three types of core data in real time. All collected raw data undergoes standardization processing to generate a raw diagnostic dataset for the edge device that can be directly used in subsequent steps. Specifically, the following steps are included: Ⅰ: The acquisition terminal uses its built-in signal sampling circuit to sample at a frequency f p =1kHz real-time acquisition of physical layer status data of the cabling transmission link. The acquired data includes the number of CRC check failures N. crc LVDS differential signal for voltage V LVDS signal rise time t rand the characteristic impedance Z of the ribbon cable cable .

[0024] Among them, the CRC check failure count records the number of check failures during data transmission within a unit time Δt, in units of times / second; the standard threshold for the voltage of the LVDS differential signal is V. std =3.3v The standard threshold for signal rise time is: The standard threshold for the characteristic impedance of the ribbon cable is: .

[0025] The collected raw data undergoes amplitude limiting filtering to remove obvious abnormal jump values, followed by wavelet denoising to eliminate high-frequency noise interference, ultimately generating a standardized physical layer data set A for cabling transmission. The CRC error rate a1 is defined as the ratio of the number of failed CRC checks to the total amount of transmitted data, calculated using the following formula: ; In the formula, R is the data transmission rate; Δt is the sampling time window, and we take Δt = 1s.

[0026] The LVDS differential signal voltage a2 is averaged after being amplitude-limited and filtered, i.e.: ; In the formula, n=f p ·Δt=1000 is the number of sampling points, f p =1kHz is the sampling frequency; V LVDS (i) represents the differential signal voltage value at the i-th sampling point.

[0027] The signal rise time a3 is taken as the maximum value within the sampling window, that is: ; In the formula, t r (i) represents the rise time value of the signal at the i-th sampling point.

[0028] The characteristic impedance a4 of the ribbon cable was measured using the time-domain reflection method, and the result was output after wavelet denoising. The calculation formula is as follows: ; In the formula, Z cable (i) represents the characteristic impedance measurement value of the i-th sampling point.

[0029] Ultimately, the physical layer data set transmitted via ribbon cable is represented as follows: .

[0030] II: The data acquisition terminal uses the temperature sensing chip and voltage sensor built into the module to sample at a frequency f. p=1Hz real-time acquisition of module operating parameter data. The acquired data includes module port temperature T port and module power supply voltage V supply The monitoring range for module port temperature is T. min = to The standard threshold for the module's power supply voltage is: .

[0031] Furthermore, the collected raw data is normalized to map data of different dimensions to the [0, 1] interval. The formula for calculating the normalized value b1 of the module port temperature is: ; The formula for calculating the normalized value b2 of the module power supply voltage is: ; This is the lower limit of the voltage standard. This is the lower limit of the voltage standard.

[0032] When V supply When the value is outside this range, b2 takes the value 0 or 1 respectively.

[0033] Ultimately, the standardized operating parameter data set is represented as ,in .

[0034] III: The data acquisition terminal uses a miniature camera built into the module to sample at a frequency f. v This method acquires local area images at 1Hz. Unlike traditional methods that acquire full-screen images, this method only acquires local area images to reduce computational consumption at edge nodes. The acquired image data is initially processed by a hybrid algorithm of Sobel edge detection and LBP local binary mode to extract two feature parameters: pixel deviation rate c1 and the number of line synchronization signal loss c2.

[0035] Pixel deviation rate c1 is defined as the proportion of pixel values ​​deviating from the standard value in a local area, and the calculation formula is: ; In the formula, M·N is the total number of pixels in the local image region; p ij Let (i,j) be the grayscale value of pixel (i,j). p std This is the theoretical standard grayscale value for this region; δ is the deviation threshold; This is an indicator function that takes the value 1 when the condition is true and 0 otherwise.

[0036] The number of horizontal synchronization signal loss, c2, directly records the number of times the horizontal synchronization signal is interrupted within a unit time Δt, in units of times / second. The acquired raw data is processed by a hybrid algorithm combining Sobel edge detection and LBP local binary mode to output a lightweight visual feature dataset. .

[0037] IV: The acquisition terminal integrates the standardized data set generated in the above steps to form the edge-end raw diagnostic dataset D, represented as: ; All data is encapsulated in JSON format and stored in the local storage of the edge nodes. The local storage period is set to 7 days to ensure that sufficient historical data can be retained for fault backtracking even in the event of a network outage.

[0038] Through the above step S1, the present invention realizes real-time acquisition and standardized processing of multi-source data on the status of the ribbon cable transmission link, module operating parameters and local visual features at the edge of the LED screen, providing a precise and uniformly formatted data foundation for subsequent fault diagnosis.

[0039] like Figure 2 As shown, step S2 of this invention involves edge-end fault fingerprint clustering and feature mapping. Specifically, this step receives the original diagnostic dataset from the edge node. A fault fingerprint database is pre-stored in the local memory of the edge node. This database contains standard fingerprint templates for three types of core cable faults, corresponding to oxidation faults, loosening faults, and copper foil micro-crack faults, respectively. This step uses an improved lightweight K-means clustering algorithm to cluster the collected feature data into normal fingerprint clusters and suspected fault fingerprint clusters. Then, it maps these clusters with the standard templates in the fault fingerprint database, calculates the mapping confidence, and outputs the fault type identification result and confidence for subsequent steps.

[0040] Ⅰ: Pre-configure a fault fingerprint database in the local storage of the edge node. This fault fingerprint database is built based on historical fault data and contains standard fingerprint templates for three types of core cable faults. Let the set of fault types be... These correspond to oxidation faults, loosening faults, and copper foil micro-crack faults, respectively. Each type of fault has a Tfault rating. k Corresponding to a standard fingerprint template Φ k Defined as the ideal range of values ​​for each characteristic parameter under this fault type: ; In the formula, φ k,j Indicates fault type T k Corresponding to the j-th feature parameter d j The standard value range of j, where j corresponds to the index of the 8 feature parameters in the original diagnostic dataset D at the edge end. .

[0041] The fault fingerprint database supports dynamic updates, which are triggered by a self-learning mechanism in subsequent steps, requiring no manual maintenance.

[0042] II: This sub-step employs an improved lightweight K-means clustering algorithm to cluster the feature parameters in the dataset D output from step S1 into normal fingerprint clusters and suspected faulty fingerprint clusters. First, the feature vector x participating in the clustering is extracted from D, and seven key feature parameters are selected: a1, a2, a3, a4, b1, c1, and c2, forming a 7-dimensional feature vector. ; The clustering algorithm sets the number of clusters K=2, corresponding to normal fingerprint clusters C. normal And suspected faulty fingerprint cluster C suspect The algorithm iteratively optimizes the cluster centers, minimizing the sum of squared distances from samples within a cluster to the cluster center. The objective function is calculated as follows: ; In the formula, μ i Let be the center vector of the i-th cluster.

[0043] During clustering, the intra-cluster similarity threshold θ is dynamically calculated. sim The deviation value of the characteristic impedance a4 in the threshold linkage step S1 is calculated using the following formula: ; In the formula, θ base =0.85 is the baseline similarity threshold; λ=0.15 is the linkage coefficient; This is the standard value of characteristic impedance.

[0044] The greater the deviation of a4 from the standard value, the more θ sim The lower the value, the more lenient the clustering conditions, thus avoiding missed detections due to parameter offsets.

[0045] After clustering is completed, the cluster category to which the sample belongs and the coordinates of the cluster center are output.

[0046] III: The fingerprint will be classified into the suspected faulty fingerprint cluster C. suspect Feature vector x and fault fingerprint database Standard fingerprint template Φ k Perform a mapping and calculate the mapping confidence level K. The mapping logic uses feature bias weighted comparison, assigning different weights to core features and auxiliary features.

[0047] Specifically, define the feature weight vector. The weights of the core features a1, a2, and a4 are: The weights of auxiliary features b1, c1, c2, and a3 are: b2 does not participate in the mapping in this step. The weights satisfy the normalization condition: ; In practical applications, normalization is required. Let the normalization factor W = 1.3, then the weights of each feature are adjusted as follows: ,satisfy .

[0048] Characteristic deviation Δ k,j Defined as actual eigenvalue d j With fault type T k Standard range φ k,j The normalized bias is calculated using the following formula: ; In the formula, μ k,j Fault type T k Lower feature d j The standard value; σ k,j This represents the standard deviation tolerance.

[0049] Mapping confidence K k The calculation formula is: ; In the formula, Δ max =3 is the maximum permissible deviation multiple. When the deviation exceeds 3 times the standard deviation, it is considered a complete deviation, and the confidence contribution drops to 0.

[0050] Furthermore, iterate through all fault types T k The confidence set is calculated. Take the maximum value As a matching result, a judgment is made based on the confidence threshold: ①If K max Output the target fault T that matches the output. k* and mapping report, among which ; ②If The fingerprint is labeled as a fuzzy fingerprint, and a set of candidate fault types is output. And the corresponding confidence level, proceed to step S3 for tracing back to the source; ③If If the system is determined to be operating normally, return to step S1 to collect data again.

[0051] IV: To adapt to the computing power limitations of edge nodes, this step optimizes the clustering algorithm with lightweight features. First, the 7-dimensional feature vector x is compressed to 4 dimensions using a feature dimensionality reduction algorithm, retaining the core features a1, a2, a4, and b1. The feature vector after dimensionality reduction is: ; The dimensionality-reduced features retained the highest correlation with the original 7-dimensional features, with a correlation coefficient of [missing value]. The formula for calculating the number of parameters in a clustering algorithm is: ; In the formula, K=2 represents the number of clusters; d dim =4 represents the feature dimension after dimensionality reduction; A byte is the storage size for a single-precision floating-point number; Calculated byte.

[0052] In the actual algorithm implementation, with the addition of auxiliary data structures such as cluster center initialization and iterative calculation, the total number of parameters is compressed to less than 0.8MB, and the clustering response time is reduced. It is compatible with computing power conditions where the MCU main frequency is ≤800MHz.

[0053] This step uses an improved lightweight K-means clustering algorithm to cluster the collected feature data into normal fingerprint clusters and suspected fault fingerprint clusters. Then, it maps the clusters to standard templates in the fault fingerprint database, calculates the mapping confidence, and finally outputs the fault type identification result and confidence for subsequent steps.

[0054] like Figure 3 As shown, step S3 of this invention is dynamic source tracing and fault confirmation, which is the confirmation step in fault diagnosis. Specifically, this step receives the fault type identification result and confidence level output from step S2, and the original edge-end diagnostic dataset output from step S1. When the mapping confidence level K output from step S2 satisfies 0.6≤K<0.8, this step initiates the source tracing mechanism, confirming the fuzzy fingerprint through the dual logic of reverse tracing and bidirectional verification, and finally outputting the confirmed fault type and root cause information. Specifically, it includes the following steps: I: First, assume three types of core cable faults as potential root causes, corresponding to oxidation faults, loosening faults, and copper foil micro-crack faults, respectively. For each type of root cause, deduce the corresponding characteristic parameter variation law based on its physical mechanism.

[0055] Specifically, the reverse pattern of oxidation faults is as follows: the LVDS differential signal voltage continues to decrease, with a decrease of more than 0.3 volts; the CRC bit error rate increases intermittently; and the module port temperature continues to rise, with the normalized temperature value increasing by more than 0.1 volts.

[0056] The reverse pattern of loosening faults is as follows: the LVDS differential signal voltage fluctuates, with a fluctuation amplitude of more than 0.3 volts; the CRC bit error rate increases intermittently; and the signal rise time fluctuates, with a fluctuation amplitude of more than 2 nanoseconds.

[0057] The inverse relationship of copper foil microcrack faults is as follows: the characteristic impedance value of the ribbon cable continues to decrease, with a decrease of more than 5 ohms; the LVDS differential signal voltage is unstable, with a fluctuation of more than 0.2 volts; the number of horizontal synchronization signal loss increases, with an increase of more than 2 times per second.

[0058] The actual changes in the collected feature parameters are compared item by item with the above-mentioned inverse inference rules to calculate the inverse inference fit η for each type of fault. k Let J be the set of feature indices involved in the back-inference. For fault type T... k The formula for calculating the compatibility degree is as follows: ; In the formula, This is an indicator function, which takes the value 1 when the actual characteristic change matches the inverse law, and 0 otherwise; Fault type T k Lower feature d j The inverse reasoning pattern; If η k If the value is ≥0.8, then the root cause of the fault, T, is considered. k Output as a candidate fault.

[0059] II: For the candidate fault root causes output in sub-step I, this step adopts a two-way verification mechanism of historical data and adjacent data to further verify the accuracy of fault judgment.

[0060] For historical data verification, historical data stored locally on the edge node is accessed to extract the feature parameter sequence of the cable interface for the previous 30 seconds. Where t0 is the current time. The degree of fit between historical data and the inverse relationship of candidate faults, η. hist The calculation formula is: ; In the formula, N hist This represents the number of historical sampling points. d j (t i ) represents a historical moment t i The characteristic parameter values; Fault type T k Lower feature d j The inverse reasoning pattern; Regarding adjacent data verification, assuming the current module number is m, extract the current feature data of adjacent modules m-1 and m+1. The feature difference between the adjacent modules and the current module is then calculated. The calculation formula is: ; Based on the propagation rules of candidate faults, a propagation rule fit degree is defined. for: ; In the formula, This is a set of feature indexes used to verify the propagation patterns. Fault type T k The expected threshold for the difference of features.

[0061] Furthermore, by combining historical fit with the fit of dissemination patterns, a comprehensive verification fit is calculated. : ; In the formula, α = 0.6 represents the weight of historical data; 1-α=0.4 is the weight of adjacent data.

[0062] Next, based on comprehensive verification of the fit Make a judgment: ①If Then the target fault and its root cause are confirmed, and the mapping confidence is updated. ; ②If If the clustering fails, return to step S2 for re-clustering, returning at most once. ③If If the result is not a cable fault, the result will be output directly, and subsequent cable-related diagnostic steps will be terminated.

[0063] Based on the verification results from the above steps, output fault confirmation information. This information includes the target fault, a description of the fault's root cause, the updated confidence level, and the feature data used for verification. The output format is structured data, which can be used in subsequent steps.

[0064] Through the above step S3, the present invention realizes reverse tracing and bidirectional verification of the root cause of fuzzy fingerprints, and improves the traditional forward multi-level judgment logic into a confirmation mechanism that combines reverse tracing and bidirectional verification, effectively improving the accuracy of fault diagnosis.

[0065] Step S4 of this invention is fingerprint-associative topology localization, which is the localization step in fault diagnosis. Specifically, this step receives fault confirmation information and achieves interface-level localization of the fault location through fingerprint-associative topology modeling and feature point back-inference mechanism. This step breaks through the limitations of gradient calculation and unidirectional localization in traditional topology localization, and adopts a dual localization logic combining fingerprint-associative topology and feature point back-inference to output the specific fault location and deviation information. Specifically, it includes the following steps: Part I: First, a fingerprint association topology graph is constructed based on the cascaded information of the screen modules. The fault fingerprint of each module is bound to a topology node, forming a three-in-one topology structure of node-fingerprint-feature. Assume there are M modules in the screen, numbered 1, 2, ..., M according to the cascade order. For each module, its topology node is defined. The node attributes include module number, cascade position, fault fingerprint label, and feature parameter vector. The fault fingerprint label comes from the clustering results in step S2, and the feature parameter vector comes from the original diagnostic dataset at the edge end in step S1.

[0066] In the fingerprint association topology graph, the node set contains the topology nodes corresponding to all modules, and the edge set represents the cascaded connection relationship between modules, that is, there are connecting edges between adjacent modules. Each node is associated with its fault fingerprint label, forming a three-in-one structure of node-fingerprint-feature, which does not require separate modeling and can directly reuse the output data of steps S1 and S2.

[0067] II: Based on the root cause of the fault identified in step S3, trace the propagation path of the fault fingerprint. Different root causes correspond to different fingerprint propagation patterns; this step selects the appropriate propagation tracing strategy based on the target fault.

[0068] For oxidation faults, the fingerprint propagates along the high-temperature region, meaning the fault's impact is limited to a single interface, with no significant changes in the characteristic parameters of adjacent nodes; the propagation path is a single-point localization. For loosening faults, the fingerprint propagates along the vibration region, potentially affecting multiple related interfaces; the changes in characteristic parameters gradually decrease along the cascading direction; the propagation path is an attenuation chain. For copper foil microcrack faults, the fingerprint propagates along the signal transmission direction; the changes in characteristic parameters are abrupt, with significant jumps in characteristic parameters before and after the fault point.

[0069] Fingerprint propagation path tracing involves traversing the nodes of the topology graph, calculating the matching degree between the feature parameters of each node and the fault fingerprint template, and locating the starting node of the propagation path. Let node V... m fingerprint matching degree ρ m The calculation formula is: ; In the formula, A set of feature indexes for participating in propagation tracking; For the target fault Tk* Lower feature d j The standard value; Standard deviation tolerance; Traverse all nodes and calculate the matching degree sequence. The node whose matching degree first drops below the threshold is identified as the fault initiation node.

[0070] III: The feature-based back-inference logic is adopted. Based on the fault fingerprint propagation path, the theoretical value of the feature parameters of the fault point is inferred, and then compared with the actual collected value. The interface with the smallest deviation is the fault point. At the same time, the characteristic impedance data is combined for accuracy calibration.

[0071] First, based on the target fault T k* Based on the propagation pattern, we can deduce the theoretical values ​​of the characteristic parameters of the fault point m*. For single-point location faults such as oxidation faults, the theoretical value is taken as the average of the characteristic parameters of adjacent nodes: ; For attenuation chain faults such as loosening faults, the theoretical value is calculated using linear interpolation: ; In the formula, λ is the propagation attenuation coefficient, which is determined by fitting historical data.

[0072] Secondly, the feature parameter deviation ε of each candidate interface is calculated. m : ; In the formula, For the feature index set used in localization, feature parameters sensitive to cable faults, such as a2, a4, and c1, are selected. For the target fault T k* Lower feature d j Standard deviation tolerance.

[0073] Furthermore, the interface with the smallest deviation is identified as the fault point: ; In the formula, This is a set of candidate interfaces, typically including nodes with abnormal matching in the propagation path tracing and their neighboring nodes.

[0074] Finally, accuracy calibration is performed using the characteristic impedance data. If the characteristic impedance of the fault point m* is... Deviation from standard value If the accuracy exceeds 10%, the positioning results will be further confirmed, and the positioning accuracy will be calibrated to the interface level. .

[0075] Based on the above location results, output the fault location information. The location information includes the receiver card number, cable interface number, module number, fault root cause type, characteristic parameter deviation value, and location accuracy level. The output format is structured data, which can be directly used by maintenance personnel without additional troubleshooting.

[0076] Through the above step S4, the present invention realizes a localization mechanism that combines fault fingerprint association topology modeling and landing point reverse inference, improving the traditional gradient calculation and unidirectional localization into a dual localization logic of fingerprint association and bidirectional reverse inference, effectively improving the accuracy and reliability of fault localization.

[0077] Step S5 of this invention is fingerprint self-learning and closed-loop optimization, which is a dynamic optimization step in fault diagnosis. Specifically, this step receives the original edge-end diagnostic dataset output from step S1, the clustering results and mapping confidence from step S2, the target fault confirmation results from step S3, and the location information from step S4. Through fault case fingerprint storage, a self-learning update mechanism, and end-to-end adaptive calibration, the diagnostic method is continuously optimized, forming a complete closed-loop optimization process from data acquisition to diagnosis and parameter calibration. Specifically, it includes the following steps: Ⅰ: The entire process data generated in steps S1 to S4 is structured to form a complete fault case record. Each fault case is associated with a unique fault identifier, and the stored content includes the edge-end original diagnostic dataset of step S1, i.e., the complete values ​​of the 8 feature parameters collected when the fault occurred; the clustering results of step S2, including the fingerprint cluster category to which the fault belongs and the mapping confidence; the target fault confirmation results of step S3, including the confirmed fault type, fault root cause description, and two-way corroboration data; and the location information of step S4, including the receiver card number, cable interface number, module number, and location accuracy level where the fault is located.

[0078] Simultaneously, this step extracts the actual fault fingerprint from the aforementioned full-process data. The actual fault fingerprint differs from the preset standard fingerprint template; it is a combination of real fault features generated based on the actually collected feature parameters and the confirmed fault type. The actual fault fingerprint includes a fault type label and the actual value range of the corresponding 8 feature parameters. All fault cases are stored in the local storage of the edge node, with a storage period set to 180 days for subsequent self-learning updates.

[0079] II: Statistical analysis is performed on the stored fault cases. When the cumulative number of actual fault fingerprint cases of the same type reaches a preset threshold, the fingerprint self-learning mechanism is automatically triggered. The preset threshold is set to 20, meaning that after 20 actual fault cases of the same fault type are stored, the self-learning update process is started.

[0080] The self-learning update process extracts common features from all actual cases of this type of fault, calculates the statistical distribution of each feature parameter, including the mean, standard deviation, and confidence interval. The statistical results are compared with the standard fingerprint template for this fault type in the fault fingerprint database. If the deviation between the statistical mean of the actual cases and the standard template exceeds a preset adjustment threshold, the standard fingerprint template is automatically updated.

[0081] For example, the deviation threshold for the LVDS differential signal voltage in the standard template for oxidation faults was originally set to 8%. If actual case statistics show that the average voltage deviation when this type of fault occurs reaches 7.5% and the standard deviation is small, then the deviation threshold of the standard template will be adjusted to 7.5%. The updated standard fingerprint template will take effect immediately and will be used in subsequent fault diagnosis processes.

[0082] The self-learning update simultaneously optimizes the clustering similarity threshold and feature weight parameters in step S2. Through backtracking analysis of real-world cases, the discriminative power of each feature parameter in different fault types is calculated. Features with high discriminative power are assigned higher weights, while features with low discriminative power are assigned lower weights, thereby improving the accuracy of subsequent fingerprint clustering.

[0083] III: Based on the actual fault causes and handling effects entered by maintenance personnel, the entire link parameters of the diagnostic process are adaptively calibrated. After completing fault repair, maintenance personnel enter the actual fault causes, repair measures, and handling effects feedback through the maintenance terminal. The system compares and analyzes this feedback data with the source tracing results of step S3 and the location results of step S4.

[0084] If the actual cause of the fault is consistent with the diagnostic conclusion and the handling effect is good, then the weight of relevant parameters in the current diagnostic process will be strengthened, including the threshold of the consistency of the reverse inference pattern of this type of fault in the source tracing and the parameters of the propagation pattern of this type of fault in the localization algorithm.

[0085] If there is a discrepancy between the actual cause of the fault and the diagnostic conclusion, the relevant parameters are adjusted according to the discrepancy. For example, if multiple actual faults are copper foil microcracks, but the back-inference result in step S3 has a large deviation, the weights of the back-inference features of the copper foil microcrack fault are adjusted, increasing the weight coefficients of the characteristic impedance of the ribbon cable and the number of times the horizontal synchronization signal is lost, and decreasing the weight coefficients of other features to improve the accuracy of subsequent back-inference.

[0086] The end-to-end adaptive calibration simultaneously optimizes the back-calculation parameters of step S4, including the propagation attenuation coefficients and deviation tolerance thresholds for various faults. By comparing actual positioning results with manual troubleshooting results, these parameters are dynamically adjusted to continuously improve positioning accuracy.

[0087] IV: After self-learning and calibration are completed, the optimized parameters are applied to the next diagnostic process. The edge node stores the updated fault fingerprint database, clustering parameters, source tracing and inference patterns, and location parameters in its local storage as the initial configuration for subsequent diagnoses. Simultaneously, step S1 continues to collect the original diagnostic dataset from the edge in real time, triggering a new round of fault diagnosis, enabling the entire system to continuously optimize itself during operation.

[0088] The closed-loop optimization in this step is completed entirely locally on the edge node, without relying on a cloud server. Even if the edge node is disconnected from the cloud network, the self-learning and calibration mechanisms continue to operate normally, ensuring continuous optimization of the diagnostic methods. At the same time, the edge node periodically reports the optimized parameters and new fault cases to the cloud platform for cross-screen statistical analysis and process improvement, but the reporting does not affect the independent operation of the edge node locally.

[0089] Through the above steps, this invention achieves self-learning update of fault fingerprints and adaptive calibration of full-link parameters, improving the traditional one-way fixed diagnostic logic into a dynamic closed-loop optimization mechanism, enabling the diagnostic method to continuously evolve in practical applications and constantly improve the accuracy of fault identification and the precision of fault location.

[0090] The second objective of this invention is to provide an LED screen fault diagnosis system based on edge computing. This system, corresponding to the aforementioned method, is used to achieve real-time diagnosis and effective location of LED screen faults. The system includes an edge data acquisition module, a fault fingerprint clustering module, a dynamic source tracing and reverse engineering module, a topology localization module, a self-learning optimization module, and a cloud collaboration module. The specific functions of each module are as follows: The edge data acquisition module is deployed at the edge node corresponding to each receiving card of the LED screen, and integrates a data acquisition unit and an edge computing unit.

[0091] The data acquisition unit collects physical layer data of the ribbon cable transmission in real time at a sampling frequency of 1kHz, including the number of CRC check failures, LVDS differential signal voltage, signal rise time, and ribbon cable characteristic impedance; it also collects LED module operating parameter data in real time at a sampling frequency of 1Hz, including module port temperature and module power supply voltage; and it collects local area images through the miniature camera built into the module at a sampling frequency of 1Hz, extracts pixel deviation rate and number of horizontal synchronization signal loss, and generates lightweight visual feature data.

[0092] The edge computing unit performs amplitude limiting filtering, wavelet denoising, and normalization on the collected raw data to generate a raw diagnostic dataset at the edge, which is stored in the local memory of the edge node for a period of 7 days. The standardized data output by this module serves as the data foundation for the entire diagnostic system and is used by subsequent modules.

[0093] The fault fingerprint clustering module is connected to the edge data acquisition module, receives the raw diagnostic dataset from the edge, and runs in the computing unit of the edge node. This module has a built-in fault fingerprint library containing standard fingerprint templates for three core cable tray faults: oxidation faults, loosening faults, and copper foil micro-crack faults.

[0094] The fault fingerprint clustering module employs an improved lightweight K-means clustering algorithm to extract 7-dimensional feature vectors from the original diagnostic dataset at the edge, clustering them into normal fingerprint clusters and suspected fault fingerprint clusters. During the clustering process, the similarity threshold is dynamically adjusted in conjunction with the deviation value of the characteristic impedance of the cabling.

[0095] The fault fingerprint clustering module maps the feature data of suspected fault fingerprint clusters to standard templates in the fault fingerprint database, calculates the mapping confidence, and outputs the fault type identification result and confidence. This module has been lightweight and optimized, with a clustering response time of no more than 3 milliseconds and the number of parameters compressed to less than 0.8MB, making it suitable for edge node computing power conditions.

[0096] The dynamic source tracing and reverse inference module is connected to the fault fingerprint clustering module to receive fault type identification results and confidence levels. It is also connected to the edge data acquisition module to receive the raw diagnostic dataset from the edge. When the mapping confidence level is between 0.6 and 0.8, this module activates the source tracing and reverse inference mechanism.

[0097] The dynamic source tracing and reverse deduction module first assumes oxidation faults, loosening faults, and copper foil microcrack faults as potential fault root causes. Based on the physical mechanism of each fault, it reverse deduces the variation law of characteristic parameters, calculates the degree of fit between the actual characteristic changes and the reverse deduction law, and selects fault root causes with a degree of fit of more than 80% as candidate faults.

[0098] The dynamic source tracing and reverse engineering module further utilizes the historical data stored locally on the edge nodes for the previous 30 seconds to calculate the degree of fit between the historical data and the reverse engineering patterns of candidate faults. Simultaneously, it extracts feature data from adjacent modules, calculates the feature differences between adjacent modules and the current module, and calculates the degree of fit based on the propagation patterns of candidate faults. This module combines the historical fit and the propagation pattern fit; when the overall fit reaches 85% or higher, the target fault is confirmed, and the mapping confidence is increased by 0.1, but not exceeding 0.9. This module outputs the confirmed target fault type and root cause information.

[0099] The topology localization module is connected to the dynamic source tracing and reverse inference module to receive target fault information. It is also connected to the edge data acquisition module to receive the original diagnostic dataset from the edge and to the fault fingerprint clustering module to receive the fingerprint clustering results.

[0100] The topology localization module constructs a fingerprint-associated topology graph based on the cascaded information of the screen modules, binding the fault fingerprint of each module to a topology node to form a three-in-one topology structure of node-fingerprint-feature. This module selects an appropriate propagation and tracing strategy according to the type of the target fault, traverses the nodes of the topology graph, calculates the matching degree between the feature parameters of each node and the target fault fingerprint template, and locates the starting node of the propagation path.

[0101] The topology positioning module employs a feature-based fault location logic. Based on the propagation patterns of the target fault, it infers the theoretical values ​​of the fault's characteristic parameters, calculates the characteristic parameter deviations of each candidate interface, and identifies the interface with the smallest deviation as the fault point. Accuracy calibration is then performed using characteristic impedance data. This module outputs fault location information, including the receiver card number, cable interface number, module number, fault root cause type, characteristic parameter deviation value, and positioning accuracy level.

[0102] The self-learning optimization module is connected to the edge data acquisition module, fault fingerprint clustering module, dynamic source tracing and reverse engineering module, and topology positioning module, respectively, and receives the full-process data output by each module. This module performs structured processing of the complete data from each diagnosis, generates fault cases, associates them with unique fault identifiers, and stores them in the local storage of the edge node for a storage period of 180 days.

[0103] When the cumulative number of actual fault fingerprint cases of the same type reaches 20, the self-learning optimization module automatically triggers the self-learning mechanism to extract common features of similar cases, update the standard fingerprint template in the fault fingerprint database, and optimize the clustering similarity threshold and feature weight parameters. This module also adaptively calibrates the source tracing and location parameters based on the actual fault causes and handling effects entered by maintenance personnel, and stores the optimized parameters in local storage for subsequent diagnostic processes.

[0104] The cloud collaboration module connects with the self-learning optimization module, responsible for data synchronization between edge nodes and the cloud platform. When the network connection is normal, this module reports the optimized parameters and newly added fault cases generated by the self-learning optimization module to the cloud platform for cross-screen statistical analysis and process improvement. When the network is interrupted, this module caches the reported data and uploads it asynchronously after the network is restored. The reporting operation of this module does not affect the independent diagnostic operation of the edge nodes locally, ensuring that the system can still work normally in unstable network environments.

[0105] Through the coordinated work of the above six modules, this invention realizes a complete diagnostic closed loop from data acquisition, fault identification, source tracing and reverse inference, effective localization to self-learning optimization. It effectively solves the technical problems in the prior art where, when facing LED screens spliced ​​together from hundreds or thousands of modules, the diagnostic system is unable to obtain the underlying status data of the ribbon cable transmission link and is unable to distinguish ribbon cable connection faults from other hardware faults.

[0106] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for diagnosing a fault of an LED screen based on edge computing, characterized in that, Includes the following steps: S1. Deploy acquisition terminals at the edge nodes corresponding to each receiving card of the LED screen to collect physical layer data transmitted via ribbon cable, module operating parameter data and lightweight visual feature data. After processing, generate the edge end raw diagnostic dataset containing feature data of each module. S2. Pre-set a fault fingerprint database at the edge nodes, which includes three types of faults: oxidation faults, loosening faults, and copper foil micro-crack faults. Feature parameters were extracted from the original diagnostic dataset at the edge and clustered into normal fingerprint clusters and suspected faulty fingerprint clusters. The feature data of suspected fault fingerprint clusters are mapped to the fault fingerprint database to obtain the fault type identification results and confidence level; S3. Based on the fault type, preset the fault back-inference rule and back-inference fit threshold, compare the fault back-inference rule with the feature data of the suspected fault fingerprint cluster to obtain the back-inference fit, and take the fault type that meets the back-inference fit threshold as the candidate fault. The feature data of adjacent modules are extracted from the original diagnostic dataset at the edge and compared with the propagation pattern of candidate faults. When the degree of fit of the propagation pattern reaches the preset verification threshold, the target fault is confirmed. S4. Construct a fingerprint association topology diagram based on the cascading order of the screen modules; The fingerprint clustering results are bound to the module positions in the topology map. The starting module is located by traversing the topology map according to the target fault. The theoretical values ​​of the feature parameters of the fault point are inferred from the propagation law of the target fault. The values ​​are compared with the feature parameters in the original diagnostic dataset at the edge. The interface with the smallest deviation is determined as the fault point, and the fault location information is output.

2. The LED screen fault diagnosis method based on edge computing according to claim 1, characterized in that: The physical layer data transmitted via the ribbon cable includes CRC bit error rate, LVDS differential signal voltage, signal rise time, and ribbon cable characteristic impedance. The module operating parameter data includes module port temperature and module power supply voltage. The lightweight visual feature data includes pixel deviation rate and number of line synchronization signal loss.

3. The LED screen fault diagnosis method based on edge computing according to claim 1, characterized in that: The processing of the physical layer data, module operating parameter data, and lightweight visual feature data transmitted via the cabling includes amplitude limiting filtering and wavelet denoising, and the processed data is normalized to generate the original diagnostic dataset at the edge.

4. The LED screen fault diagnosis method based on edge computing according to claim 1, characterized in that: An improved lightweight K-means clustering algorithm is used to cluster the feature parameters extracted from the original diagnostic dataset at the edge. The similarity threshold during the clustering process is dynamically adjusted according to the degree to which the characteristic impedance of the cabling in the original diagnostic dataset at the edge deviates from the standard value. The clustering response time does not exceed 3 milliseconds.

5. The LED screen fault diagnosis method based on edge computing according to claim 2, characterized in that: The feature parameters extracted from the original diagnostic dataset at the edge include CRC bit error rate, LVDS differential signal voltage, ribbon cable characteristic impedance, module port temperature, pixel deviation rate, and number of line synchronization signal loss. The feature parameters are compressed to 4 dimensions using a feature dimensionality reduction algorithm. The core features retained after dimensionality reduction include CRC bit error rate, LVDS differential signal voltage, ribbon cable characteristic impedance, and module port temperature. The compressed feature parameters participate in clustering operations.

6. The LED screen fault diagnosis method based on edge computing according to claim 1, characterized in that: The fault regression patterns include regression patterns for oxidation faults, loosening faults, and copper foil microcrack faults. The regression patterns for oxidation faults are: LVDS differential signal voltage decrease, CRC bit error rate increase, and module port temperature increase. The regression patterns for loosening faults are: LVDS differential signal voltage fluctuation and signal rise time fluctuation. The regression patterns for copper foil microcrack faults are: decrease in ribbon cable characteristic impedance and increase in the number of horizontal synchronization signal loss.

7. The LED screen fault diagnosis method based on edge computing according to claim 1, characterized in that: The propagation law fit is calculated as follows: extract feature data of adjacent modules from the original diagnostic dataset at the edge, calculate the feature difference between the adjacent modules and the current module, and determine whether the feature difference meets expectations based on the propagation law of the candidate fault to obtain the propagation law fit.

8. The LED screen fault diagnosis method based on edge computing according to claim 1, characterized in that: After identifying the interface with the smallest deviation as the fault point, the accuracy is calibrated by combining the characteristic impedance of the fault point's cable. If the characteristic impedance deviates from the standard value by more than a preset percentage, the positioning accuracy is at the interface level; otherwise, it is at the module level.

9. The LED screen fault diagnosis method based on edge computing according to claim 1, characterized in that: It also includes storing the entire process data as fault cases. When the number of cases of the same fault type reaches a preset number, common features are extracted to update the fault type standard in the fault fingerprint database, and the parameters of fault back-inference and propagation are calibrated in combination with operation and maintenance feedback.

10. An LED screen fault diagnosis system based on edge computing, used to execute the LED screen fault diagnosis method based on edge computing as described in any one of claims 1-9, characterized in that, It includes an edge data acquisition module, a fault fingerprint clustering module, a dynamic source tracing and reverse inference module, a topology positioning module, a self-learning optimization module, and a cloud collaboration module; The edge data acquisition module is deployed at the edge node corresponding to each receiving card of the LED screen to collect physical layer data transmitted by the ribbon cable, module operating parameter data and lightweight visual feature data. After filtering, denoising and normalization, the original edge diagnostic dataset is generated. The fault fingerprint clustering module is used to pre-set a fault fingerprint database containing oxidation faults, loosening faults and copper foil micro-crack faults in the local memory of the edge node, extract feature parameters from the original diagnostic dataset at the edge end and cluster them into normal fingerprint clusters and suspected fault fingerprint clusters, and map the feature data of the suspected fault fingerprint clusters to the fault fingerprint database to obtain the fault type identification result and confidence level. The dynamic source tracing and reverse inference module is used to preset the fault reverse inference rule and the reverse inference fit threshold according to the fault type. It compares the fault reverse inference rule with the feature data of the suspected fault fingerprint cluster to obtain the reverse inference fit. The fault type that meets the reverse inference fit threshold is used as the candidate fault. It extracts the feature data of the adjacent module from the original diagnostic dataset at the edge end and compares it with the propagation rule of the candidate fault. When the propagation rule fit reaches the preset verification threshold, the target fault is confirmed. The topology localization module is used to construct a fingerprint association topology map according to the cascading order of the screen modules, bind the fingerprint clustering results with the module positions in the topology map, locate the starting module according to the target fault traversal topology map, back-calculate the theoretical value of the feature parameters of the fault point according to the propagation law of the target fault, compare it with the feature parameters in the original diagnostic dataset at the edge end, determine the interface with the smallest deviation as the fault point, and output the fault location information. The self-learning optimization module is used to store the entire process data as fault cases. When the number of cases of the same fault type reaches a preset number, common features are extracted to update the fault fingerprint database, and the fault back-inference and propagation parameters are calibrated in combination with operation and maintenance feedback. The cloud collaboration module is used to report optimized parameters and new fault cases to the cloud platform when the network is connected, and to cache the reported data when the network is interrupted and upload it asynchronously after the network is restored.

Images