Electronic splicing screen fault automatic diagnosis and repair method

Abstract

The invention relates to the technical field of fault diagnosis and repair of electronic equipment, and discloses an automatic fault diagnosis and repair method for an electronic splicing screen, which comprises the following steps of: in the operation process of the electronic splicing screen, acquiring multi-source data in real time through various sensors and monitoring equipment deployed in a splicing screen system; carrying out preprocessing and feature extraction on the collected multi-source data to obtain an image feature vector, an electrical feature vector and a communication feature vector; fusing the extracted image feature vector, the extracted electrical feature vector and the extracted communication feature vector to form a comprehensive feature vector, inputting the comprehensive feature vector into a fault diagnosis model for processing, and outputting a fault diagnosis result; a repair strategy is determined according to a fault diagnosis result, repair operation is automatically executed, after repair is completed, data are collected again for verification, and if the fault is still not eliminated, the repair strategy is optimized; according to the invention, the accuracy and reliability of fault diagnosis are improved, and the maintenance cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic equipment fault diagnosis and repair, and in particular to a method for automatic fault diagnosis and repair of an electronic splicing screen. Background Art

[0002] With the widespread application of electronic splicing screens in commercial displays, monitoring and command, conference systems and other fields, their stability and reliability are of vital importance. However, electronic splicing screens are composed of numerous electronic components, display units and connecting lines. During long-term operation, they are prone to various problems such as display anomalies (flowered screen, black screen, color cast, etc.), signal transmission failures, power supply failures, etc. Traditional fault diagnosis methods mainly rely on manual inspections and experience-based judgment. This method is inefficient, inaccurate, and difficult to quickly locate the root cause of complex faults. Summary of the Invention

[0003] The purpose of the present invention is to solve the above problems and to design a method for automatic diagnosis and repair of electronic splicing screen faults.

[0004] The present invention provides a method for automatically diagnosing and repairing faults of an electronic splicing screen, the method comprising the following steps: S1. During the operation of the electronic splicing screen, multiple sensors and monitoring devices deployed in the splicing screen system are used to collect multi-source data in real time; S2. Preprocessing and feature extraction are performed on the collected multi-source data to obtain image feature vectors, electrical feature vectors, and communication feature vectors; S3, fusing the extracted image feature vector, electrical feature vector, and communication feature vector to form a comprehensive feature vector, inputting the comprehensive feature vector into a fault diagnosis model for processing, and outputting a fault diagnosis result; S4. Determine the repair strategy based on the fault diagnosis results and automatically perform the repair operation. After the repair is completed, collect data again for verification. If the fault is still not eliminated, optimize the repair strategy.

[0005] Optionally, in a first implementation of the present invention, step S1 includes the following steps: S11, using a high-definition camera to shoot the display screen of the spliced screen in real time to obtain display image data; S12. Collect the working current, voltage and temperature data of each module of the splicing screen through the current sensor, voltage sensor and temperature sensor to obtain electrical parameter data; S13, monitoring the communication signals between the splicing screen and the signal source and the control host, recording the communication protocol data, data transmission rate and number of error frames, and obtaining the communication data.

[0006] Optionally, in a second implementation of the present invention, step S2 includes the following steps: S21. For image data from multi-source data, a GAN network is used to perform denoising, and an image registration algorithm based on feature point matching is used for image registration. An improved convolutional neural network algorithm is used to extract features from the pre-processed image data to obtain image feature vectors. S22. For electrical parameter data in multi-source data, the isolation forest algorithm is used to detect electrical parameter anomalies. The dynamic time warping algorithm is combined to perform pattern matching on the changing trends of the electrical parameters. The ReliefF algorithm is used to calculate feature weights. The electrical parameter features are screened from the preprocessed electrical parameter data, and the electrical feature vector is obtained through principal component analysis and dimensionality reduction. S23. For the communication data in the multi-source data, perform semantic analysis on the communication data and convert it into a word vector sequence. Then input it into the LSTM network to capture the temporal dependency in the communication data, and combine it with the wavelet transform algorithm to extract the communication feature vector.

[0007] Optionally, in a third implementation of the present invention, step S21 includes the following steps: Input the image data into the GAN generator for denoising to generate a denoised image; The denoised image is layered using a Gaussian pyramid. The FAST corner detector is used to detect feature points in each layer. The corner points are filtered using the Harris response function to obtain the top N feature points with the highest response values. For each feature point, its main direction is calculated and a rotation-invariant BRIEF descriptor is generated. Use bidirectional kd-tree for fast feature point matching, calculate the Euclidean distance of matching point pairs and use ratio test to eliminate false matches; Bilinear interpolation is used for sub-pixel resampling. A local adaptive window is introduced during the interpolation process to complete image registration and obtain preprocessed image data. Construct a deep separable convolutional neural network architecture, which includes 5 residual modules and 3 attention modules. Each attention module consists of channel attention and spatial attention. The preprocessed image data is input into a depthwise separable convolutional neural network. The channel attention adopts the squeeze excitation mechanism of SENet, the spatial attention generates a 2D attention map through convolution, and the image feature vector is output from the global average pooling layer.

[0008] Optionally, in a fourth implementation of the present invention, step S22 includes the following steps: The electrical parameter data is divided into sliding windows, and the BiLSTM network is used to predict and fill in the missing data of the electrical parameter data; Extract a sample subset from the electrical parameter data after missing value processing, recursively randomly select features and split points in the feature space to generate an isolation tree, and calculate the path length of each sample point in all isolation trees to determine the abnormal point; For the detected abnormal points, the dynamic time warping algorithm is used to perform pattern matching of the change trend, find the optimal alignment path, and obtain the pre-processed electrical parameter data; Apply the ReliefF algorithm to calculate the weights of features in the preprocessed electrical parameter data and filter out features whose weights are higher than a preset threshold; The filtered electrical parameter features are subjected to principal component analysis and dimensionality reduction to obtain the electrical feature vector.

[0009] Optionally, in a fifth implementation of the present invention, step S23 includes the following steps: Parse communication data layer by layer, extract protocol header fields and payload content, and perform data cleaning to identify and filter invalid data packets and noise data; The Word2Vec algorithm is used to convert the denoised data into a word vector sequence, which is then input into the LSTM network. The long-distance temporal dependencies are captured through the dynamic adjustment of the forget gate, input gate, and output gate. The time series features output by LSTM are decomposed by wavelet transform, and the time series features extracted by LSTM and the frequency domain features obtained by wavelet transform are spliced to form a communication feature vector.

[0010] Optionally, in a sixth implementation of the present invention, step S3 includes the following steps: S31, for the image feature vector, the electrical feature vector, and the communication feature vector, using a normalization method to map all feature values to the same interval; S32. Use a Transformer-based feature fusion network to deeply interact and fuse the processed image feature vector, electrical feature vector, and communication feature vector under the action of self-attention and cross-attention, and output the fused feature vector; S33. Input the fused feature vector into the fault diagnosis model composed of CNN network, RNN network and GNN network, and integrate it through the fully connected layer to output the fault diagnosis result of the electronic splicing screen.

[0011] Optionally, in a seventh implementation of the present invention, step S33 includes the following steps: The CNN network extracts and abstracts images through convolutional layers and pooling layers; The RNN network uses a cyclic structure to process the temporal characteristics of the data and capture the changing patterns of the data over time; The GNN network constructs a graph structure based on the connection relationship between electronic splicing screen devices and analyzes the topological relationship and fault propagation path between devices.

[0012] Optionally, in an eighth implementation of the present invention, the fault diagnosis result includes at least the fault type, fault location and fault cause.

[0013] Optionally, in a ninth implementation of the present invention, a system for implementing the aforementioned method for automatically diagnosing and repairing electronic splicing screen faults includes: Real-time acquisition module, used to collect multi-source data in real time during the operation of the electronic splicing screen through various sensors and monitoring equipment deployed in the splicing screen system; Feature extraction module, used to preprocess and extract features from the collected multi-source data to obtain image feature vectors, electrical feature vectors, and communication feature vectors; The fault diagnosis module is used to fuse the extracted image feature vector, electrical feature vector and communication feature vector to form a comprehensive feature vector, input the comprehensive feature vector into the fault diagnosis model for processing, and output the fault diagnosis result; The automatic repair module is used to determine the repair strategy based on the fault diagnosis results and automatically perform the repair operation. After the repair is completed, data is collected again for verification. If the fault is still not eliminated, the repair strategy is optimized.

[0014] In the technical solution provided by the present invention, during the operation of the electronic splicing screen, multiple sensors and monitoring equipment deployed in the splicing screen system are used to collect multi-source data in real time; the collected multi-source data are pre-processed and feature extracted to obtain image feature vectors, electrical feature vectors and communication feature vectors; the extracted image feature vectors, electrical feature vectors and communication feature vectors are fused to form a comprehensive feature vector, which is input into the fault diagnosis model for processing and the fault diagnosis result is output; the repair strategy is determined according to the fault diagnosis result, and the repair operation is automatically executed. After the repair is completed, data is collected again for verification. If the fault is still not found, the fault is detected. If it is excluded, the repair strategy is optimized; the present invention collects multi-source data during the operation of the electronic splicing screen, comprehensively obtains the equipment operation status information, and combines with advanced feature extraction algorithms to accurately extract key features reflecting the fault, providing rich and accurate data support for fault diagnosis. The fault diagnosis model is adopted to improve the accuracy and reliability of fault diagnosis, and can effectively identify a variety of complex fault types and causes; it realizes the automatic repair function of the fault, can quickly handle the faults that can be automatically repaired, reduces manual intervention, improves the fault handling efficiency and operation stability of the electronic splicing screen, reduces maintenance costs, and has good application prospects and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Various other advantages and benefits will become apparent to those skilled in the art by reading the detailed description of the preferred embodiment below. The accompanying drawings are only for the purpose of illustrating the preferred embodiment and are not to be considered as limiting the present invention; Figure 1 A schematic diagram of a first embodiment of a method for automatically diagnosing and repairing electronic splicing screen faults provided by an embodiment of the present invention; Figure 2 A schematic diagram of a second embodiment of the method for automatically diagnosing and repairing electronic splicing screen faults provided by an embodiment of the present invention; Figure 3 A schematic diagram of a third embodiment of the method for automatic fault diagnosis and repair of an electronic splicing screen provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The terms "first," "second," "third," "fourth," and so forth (if any) in the description and claims of the present invention and in the accompanying drawings are used to distinguish similar items and are not necessarily used to describe a particular order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that shown or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, apparatus, product, or device that includes a series of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, product, or device.

[0017] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 A schematic diagram of a first embodiment of a method for automatically diagnosing and repairing electronic splicing screen faults provided by an embodiment of the present invention, the method specifically comprising the following steps: S1. During the operation of the electronic splicing screen, multiple sensors and monitoring devices deployed in the splicing screen system are used to collect multi-source data in real time; In this embodiment, a high-definition camera is used to capture the display screen of the splicing screen in real time to obtain display image data; the operating current, voltage and temperature data of each module of the splicing screen are collected through current sensors, voltage sensors and temperature sensors to obtain electrical parameter data; the communication signals between the splicing screen and the signal source and control host are monitored, and the communication protocol data, data transmission rate and number of error frames are recorded to obtain communication data.

[0018] In this embodiment, image data: a high-definition camera is used to shoot the display screen of the splicing screen in real time to obtain display image data. The image data acquisition frequency is not less than 25 frames per second to capture subtle changes in the picture. Electrical parameter data: the operating current, voltage and temperature data of each module of the splicing screen are collected through current sensors, voltage sensors and temperature sensors. The acquisition cycle is once every 100 milliseconds to ensure that the dynamic changes of electrical parameters can be reflected in a timely manner. Communication data: the communication signals between the splicing screen and external devices (such as signal sources, control hosts, etc.) are monitored, and communication-related information such as communication protocol data, data transmission rate, number of error frames, etc. are recorded to discover potential problems in the communication link.

[0019] S2. Preprocessing and feature extraction are performed on the collected multi-source data to obtain image feature vectors, electrical feature vectors, and communication feature vectors; S3, fusing the extracted image feature vector, electrical feature vector, and communication feature vector to form a comprehensive feature vector, inputting the comprehensive feature vector into a fault diagnosis model for processing, and outputting a fault diagnosis result; S4. Determine the repair strategy based on the fault diagnosis results and automatically perform the repair operation. After the repair is completed, collect data again for verification. If the fault is still not eliminated, optimize the repair strategy.

[0020] In this embodiment, a knowledge base of the correspondence between electronic splicing screen fault types and repair strategies is established; historical fault cases, expert experience and repair methods in equipment manuals are collected, and each known fault type (such as display screen distortion, power module overheating, communication protocol error, etc.) is structured and stored with corresponding repair steps, operating instructions, required tools or replacement parts, etc.; for complex faults, it is decomposed into multiple sub-faults, and the repair sequence and correlation relationship of each sub-fault are recorded; at the same time, the applicable conditions and priority are marked for each repair strategy for subsequent rapid matching; the fault type and fault cause output by the fault diagnosis model are used as search keywords and matched in the fault-repair strategy mapping knowledge base; a forward reasoning algorithm is used to obtain the corresponding information from the knowledge base. Search for a repair strategy that fully matches the diagnosis result in the knowledge base; if there are multiple possible matching strategies, calculate the comprehensive score of each strategy based on the severity, frequency, repair cost and other factors of the fault, combined with the preset weights, and select the repair strategy with the highest score as the initial implementation plan; if no fully matching strategy is found, enable the similarity-based retrieval mechanism, calculate the similarity between the diagnosis result and similar faults in the knowledge base, select the repair strategy corresponding to the fault with the highest similarity, and make appropriate adjustments based on the differences; generate a detailed sequence of operation instructions based on the selected repair strategy; for software-related repairs (such as configuration file modification, program restart), automatically send instructions to execute the operation through the system interface; for hardware faults (such as replacing a damaged circuit board), , adjust the connection cables), send operation instructions to the maintenance robot or smart terminal, including disassembly steps, component model, installation location and other information; during the execution process, monitor the operation progress and status in real time, and record the execution time, operation results and other information of each step; if a step that cannot be executed automatically is encountered (such as manual operation of a specific switch), the automatic process is paused and a prompt is issued to the operator, waiting for manual confirmation to continue to execute the subsequent steps; after the repair operation is completed, a new round of data collection process is started; the data collection method in the fault diagnosis stage is used, and the image data, electrical parameter data and communication data of the electronic splicing screen are collected again in real time through the deployed sensors and monitoring equipment; ensure the collection time range, data frequency and number The data type is kept consistent with that before fault diagnosis to facilitate effective comparative analysis. The collected raw data is preprocessed as necessary, including data cleaning (removing noise and invalid data) and format conversion, to prepare for subsequent verification analysis. The data collected after the repair is input into the fault diagnosis model and the fault diagnosis is performed again. The diagnosis results are compared with those before the repair to determine whether the fault has been eliminated. If the diagnosis result shows no fault, the repair is confirmed to be successful, and the relevant information of the repair process (including fault type, repair strategy, execution steps, repair time, etc.) is recorded in the knowledge base for subsequent reference and optimization. If the diagnosis result still shows that a fault exists or a new fault prompt appears, the repair strategy optimization phase is entered.Reinforcement learning algorithms are used to optimize repair strategies. The current fault state, executed repair strategies, and repair results are used as state inputs, and a reward mechanism is defined: positive rewards are given for successful repairs and negative rewards are given for failed repairs. By continuously simulating the execution of different repair strategies (in a virtual environment or on historical data), the agent learns the optimal repair strategy selection method. Simultaneously, a genetic algorithm is used to perform crossover and mutation operations on the repair strategy's operational steps and execution sequence to generate new candidate repair strategies. The new strategies are then compared with the original strategies in a simulated environment. Based on the test results, the superior strategy is selected and updated in the corresponding entry in the knowledge base, ensuring that a more effective repair solution can be adopted the next time a similar fault occurs.

[0021] See also Figure 2 , a schematic diagram of a second embodiment of a method for automatic diagnosis and repair of electronic splicing screen faults provided by an embodiment of the present invention, the method comprising: S21. For image data from multi-source data, a GAN network is used to perform denoising, and an image registration algorithm based on feature point matching is used for image registration. An improved convolutional neural network algorithm is used to extract features from the pre-processed image data to obtain image feature vectors. In this embodiment, the image data is input into the generator of GAN for denoising to generate a denoised image; the denoised image is layered by a Gaussian pyramid, and the FAST corner detector is used to detect feature points in each layer. The corner points are screened by the Harris response function to obtain the top N feature points with the highest response values. For each feature point, its main direction is calculated and a rotation-invariant BRIEF descriptor is generated; a bidirectional kd tree is used for fast feature point matching, the Euclidean distance of the matching point pairs is calculated, and a ratio test is used to eliminate false matches; bilinear interpolation is used for sub-pixel resampling, and a local adaptive window is introduced in the interpolation process to complete image registration and obtain preprocessed image data; a deep separable convolutional neural network architecture is constructed, which includes 5 residual modules and 3 attention modules, each of which consists of channel attention and spatial attention; the preprocessed image data is input into the deep separable convolutional neural network, the channel attention adopts the squeeze excitation mechanism of SENet, the spatial attention generates a 2D attention map through convolution, and the image feature vector is output from the global average pooling layer.

[0022] S22. For electrical parameter data in multi-source data, the isolation forest algorithm is used to detect electrical parameter anomalies. The dynamic time warping algorithm is combined to perform pattern matching on the changing trends of the electrical parameters. The ReliefF algorithm is used to calculate feature weights. The electrical parameter features are screened from the preprocessed electrical parameter data, and the electrical feature vector is obtained through principal component analysis and dimensionality reduction. In this embodiment, the electrical parameter data is divided into sliding windows, and the BiLSTM network is used to predict and fill in the missing data of the electrical parameter data; a sample subset is extracted from the electrical parameter data after missing value processing, and features and segmentation points are recursively randomly selected in the feature space to generate an isolation tree, and the path length of each sample point in all isolation trees is calculated to determine the abnormal point; for the detected abnormal point, a dynamic time warping algorithm is used to perform pattern matching of the change trend to find the optimal alignment path to obtain the preprocessed electrical parameter data; the ReliefF algorithm is applied to calculate the weights of the features in the preprocessed electrical parameter data, and features with weights higher than a preset threshold are screened out; the screened electrical parameter features are subjected to principal component analysis dimensionality reduction to obtain an electrical feature vector.

[0023] S23. For the communication data in the multi-source data, perform semantic analysis on the communication data and convert it into a word vector sequence. Then input it into the LSTM network to capture the temporal dependency in the communication data, and combine it with the wavelet transform algorithm to extract the communication feature vector.

[0024] In this embodiment, the communication data is parsed layer by layer to extract the protocol header fields and payload content, and data cleaning is performed to identify and filter invalid data packets and noise data. The Word2Vec algorithm is used to convert the denoised data into a word vector sequence, which is input into the LSTM network. The long-distance temporal dependencies are captured through dynamic adjustment of the forget gate, input gate, and output gate. The temporal features output by the LSTM are decomposed by wavelet transform, and the temporal features extracted by the LSTM and the frequency domain features obtained by the wavelet transform are spliced to form a communication feature vector.

[0025] In this embodiment, a denoising architecture that combines a generative adversarial network (GAN) with a variational autoencoder (VAE) is adopted. First, the original image is input into the encoder network of the VAE, and a latent space representation is generated through reparameterization techniques to capture the structural features of the image. At the same time, the noisy image is input into the generator of the GAN to generate a denoised candidate image. A multi-scale discriminator network is designed to perform feature matching and discrimination on the VAE reconstructed image and the GAN generated image, and the generator parameters are optimized through adversarial training. The final denoised image is output while retaining the structural constraints of the VAE and the detail generation capability of the GAN, effectively removing mixed noises such as salt and pepper noise and Gaussian noise. An improved ORB feature point detection algorithm is adopted. First, The denoised image is layered using a Gaussian pyramid. The FAST corner detector is used to detect feature points in each layer, and the corner points are finely screened using the Harris response function, retaining the top N feature points with the highest response values. For each feature point, its main direction is calculated and a rotation-invariant BRIEF descriptor is generated. To enhance the feature extraction capability in low-texture areas, an adaptive threshold adjustment mechanism based on image entropy is introduced, and the detection threshold is lowered in areas with low texture complexity to ensure uniform distribution of feature points. A bidirectional kd tree is used to achieve fast feature point matching, and the Euclidean distance of matching point pairs is calculated, and a ratio test is used to eliminate false matches. For the retained matching point pairs, the RANSAC algorithm is used to estimate the homography matrix. In the RANSAC During the iterative process, a weight calculation based on the structural similarity of the feature point neighborhood is introduced to give higher weights to matching points in texture-rich areas; after each iteration, the spatial distribution entropy of the inner point set is calculated to judge the model fitting quality, and the iteration is terminated early when the entropy value is lower than the set threshold to improve the computational efficiency; the image is perspective transformed according to the estimated homography matrix; bilinear interpolation is used for sub-pixel resampling, and a local adaptive window is introduced in the interpolation process. The window size is dynamically adjusted according to the local image gradient change. A smaller window is used in the edge area to retain details, and a larger window is used in the flat area to improve noise resistance; after the registration is completed, the structural similarity index (SSIM) of the overlapping area of the image is calculated to evaluate the registration quality. If the value is lower than the threshold, the algorithm returns to the feature point detection step and readjusts the parameters. A deep separable convolutional neural network architecture is constructed, which includes 5 residual modules and 3 attention modules. The input image is first subjected to spatial pyramid pooling to extract multi-scale features, and then the features are transformed in sequence through the residual modules. Each attention module consists of channel attention and spatial attention. The channel attention adopts the squeeze excitation mechanism of SENet, and the spatial attention generates a 2D attention map through convolution. During the feature extraction process, the stochastic depth technique is used for regularization, and some residual connections are randomly discarded to prevent overfitting. Finally, the image feature vector is output from the global average pooling layer of the network.The t-SNE algorithm is used to perform nonlinear dimensionality reduction on high-dimensional feature vectors. Locality preservation constraints are introduced during the dimensionality reduction process to ensure that features of similar fault modes maintain proximity in the low-dimensional space. The reduced feature vectors are then vector-quantized and the K-means algorithm is used to divide the feature space into multiple codebooks, with each feature vector mapped to the nearest codebook index. To improve retrieval efficiency, a multi-level hash table structure is designed to store codebook information, enabling fast fault pattern matching and retrieval.

[0026] In this embodiment, the electrical parameter data is first divided into sliding windows, each window containing continuous time series sampling points; the data in the window is Z-score standardized to eliminate the dimensional influence between different parameters; a missing value filling strategy is designed, and a bidirectional long short-term memory network (BiLSTM) is used to predict and fill the missing data, and the front-to-back dependency of the time series data is used to improve the filling accuracy; at the same time, abnormal sampling point detection is implemented, outliers exceeding the 3σ range are marked, and smoothing is performed through local weighted regression; an integrated isolation forest model is constructed for anomaly detection; a sample subset is randomly extracted from the electrical parameter data, and features and segmentation points are randomly selected in the feature space recursively to generate an isolation tree; the probability of each sample point in the The path lengths in all isolated trees are such that the shorter the path, the more likely it is to be an anomaly. Samples are divided into normal and abnormal categories by calculating the anomaly score threshold. To improve sensitivity to early faults, a weighting mechanism is introduced during model training, giving higher weights to data close to the time of fault occurrence. For the detected abnormal period, a dynamic time warping (DTW) algorithm is used to match the pattern of change trends. First, a historical fault pattern library is constructed, and the electrical parameter change curves of known fault types are stored in segments. For the abnormal sequence to be detected, the DTW distance between it and each template in the pattern library is calculated, and the optimal alignment path is found through a dynamic programming algorithm. To reduce computational complexity, a pruning strategy is used to terminate the calculation of impossible paths in advance. Finally, the output is The top K fault modes with the highest matching degree and their similarity scores are obtained; the ReliefF algorithm is applied to calculate the weight of each electrical parameter feature; for each sample, the nearest neighbor (NearHit) is found in samples of the same type, and the K nearest neighbors (NearMiss) are found in samples of different types; the feature weight is updated according to the degree of difference between the feature value NearHit and NearMiss, and a stable feature weight vector is obtained after T consecutive iterations; in order to adapt to the high-dimensional characteristics of electrical parameter data, a feature grouping strategy is introduced, and the related electrical parameters are regarded as a feature group for weight calculation to improve the accuracy of feature screening; based on the feature weight calculated by the ReliefF algorithm, a threshold θ is set to retain features with weights higher than the threshold ; Sort the importance of the filtered features, use the recursive feature elimination (RFE) strategy, gradually eliminate the features with the lowest weight and re-evaluate the model performance until the preset number of features is reached; to avoid feature redundancy, calculate the mutual information matrix between features, and retain the features with higher weights for feature pairs with mutual information higher than the threshold; perform principal component analysis (PCA) dimensionality reduction on the filtered electrical parameter features; first calculate the feature covariance matrix, and solve the eigenvalues and eigenvectors through singular value decomposition (SVD); select the eigenvectors corresponding to the first N largest eigenvalues to form a projection matrix, and project the original features into a low-dimensional space; to retain the time series features, perform time domain expansion on the features before PCA transformation, and splice the features of adjacent time steps into a high-dimensional vector;The resulting low-dimensional feature vector is the electrical parameter feature vector, which is used as input for subsequent fault diagnosis models. To capture the dynamic characteristics of electrical parameters, the feature vector after PCA dimensionality reduction is subjected to temporal enhancement. A gated recurrent unit (GRU) network is constructed, taking feature vectors of consecutive time steps as input and capturing temporal dependencies through hidden layer state transfer. An attention mechanism is added to the GRU output layer to automatically assign importance weights to features at different time steps. Finally, temporal features are extracted from the attention-weighted output and concatenated with the original feature vector to form an enhanced electrical parameter feature vector, improving the diagnostic capabilities for gradual faults.

[0027] In this embodiment, a protocol parsing engine based on a finite state machine (FSM) is constructed, and state transition rules are defined according to the communication protocol specification; the captured communication data packets are parsed layer by layer to extract the protocol header fields and payload content; a regular expression pattern matcher is designed to identify and filter invalid data packets and noise data; for the TCP / IP protocol, the complete session flow is reconstructed and the order of the fragmented data packets is restored; protocol anomaly detection is implemented, and protocol formats that do not conform to the specification are marked to provide prior knowledge for subsequent analysis; the cleaned communication payload data is segmented into byte or character sequences to construct an N-gram model; different N values (usually 1-5) are set to generate multi-gram grammatical units, and the number of each N-gram in the data set is counted. The frequency of occurrence in the word vector is calculated; the information entropy and mutual information of N-gram are calculated to evaluate its semantic importance; the chi-square test is applied to screen discriminative N-gram features and build a dictionary library; to process variable-length sequences, a sliding window mechanism is used to generate fixed-length N-gram sequences to ensure the consistency of input dimensions; the Word2Vec algorithm is used to convert N-gram features into low-dimensional dense word vectors; a CBOW (continuous bag of words model) or Skip-gram model is constructed, and the appropriate context window size and embedding dimension are set; negative sampling technology is introduced during training to improve computational efficiency, and high-frequency words are subsampled to balance the word frequency distribution; to capture the temporal characteristics of communication data, a time window constraint is added to the word vector training , so that the word vectors of adjacent time steps have semantic relevance; the generated word vector matrix is used as the input of the subsequent LSTM network; a multi-layer bidirectional LSTM network structure is constructed, including input layer, hidden layer and output layer; the word vector sequence is input into the LSTM network, and the long-distance temporal dependency is captured through the dynamic adjustment of the forget gate, input gate and output gate; to prevent the gradient disappearance problem, the gate unit update mechanism and gradient clipping strategy are adopted; Dropout regularization is added to the hidden layer, and some neuron connections are randomly discarded to improve the generalization ability; an attention mechanism layer is designed to automatically focus on the features of the key time steps and generate a weighted temporal feature representation; the temporal features of the LSTM output are decomposed by wavelet transform; and the appropriate wavelet basis function is selected. The time series signal is decomposed into approximate coefficients and detail coefficients of different frequency bands by using the number of decomposition layers (such as Daubechies wavelet). Threshold processing is performed on the coefficients of each sub-band to remove noise interference and retain fault feature information. The energy distribution characteristics of each frequency band are calculated, and the frequency band energy entropy is extracted as part of the feature vector. The detail signal is further decomposed by applying wavelet packet transform to obtain more refined frequency features. The time series features extracted by LSTM and the frequency domain features obtained by wavelet transform are spliced to form a high-dimensional feature vector. Linear discriminant analysis (LDA) is used for supervised dimensionality reduction to maximize the inter-class separation and minimize the intra-class variance. To preserve the nonlinear feature relationship, kernel principal component analysis (KPCA) is used for nonlinear dimensionality reduction.A feature importance assessment model is constructed, and the Gini index of each feature is calculated based on the random forest algorithm to screen out the most discriminative feature subset. The resulting low-dimensional feature vector is used as the feature representation of the communication data. An adaptive time window mechanism is designed to dynamically adjust the time window length based on the communication traffic characteristics. Shorter time windows are used during peak traffic periods to capture rapidly changing features, while longer time windows are used during low traffic periods to obtain more comprehensive statistical characteristics. A sampling strategy combining sliding time windows and jumping time windows is implemented to balance real-time performance and computational efficiency. To handle sudden abnormal events, a time window overlap mechanism is designed to ensure that key events are not missed. Through time window adjustment, the feature extraction process can adapt to changes in different communication scenarios. See also Figure 3 , a schematic diagram of a third embodiment of a method for automatic diagnosis and repair of electronic splicing screen faults provided by an embodiment of the present invention, the method comprising: S31, for the image feature vector, the electrical feature vector, and the communication feature vector, using a normalization method to map all feature values to the same interval; S32. Use a Transformer-based feature fusion network to deeply interact and fuse the processed image feature vector, electrical feature vector, and communication feature vector under the action of self-attention and cross-attention, and output the fused feature vector; S33. Input the fused feature vector into the fault diagnosis model composed of CNN network, RNN network and GNN network, and integrate it through the fully connected layer to output the fault diagnosis result of the electronic splicing screen.

[0028] In this embodiment, the CNN network extracts and abstracts images through convolutional layers and pooling layers; the RNN network uses a cyclic structure to process the temporal features in the data and capture the changes in the data over time; the GNN network constructs a graph structure based on the connection relationship between electronic splicing screen devices to analyze the topological relationship and fault propagation path between devices.

[0029] In this embodiment, the fault diagnosis result includes at least the fault type, fault location and fault cause.

[0030] In this embodiment, the data distribution range of the image data feature vector, electrical parameter feature vector and communication data feature vector is first counted respectively; in view of the dimensional differences of different feature vectors, a normalization method is used to map all eigenvalues to the same interval, such as [0,1] or [-1,1]; if the feature vector dimensions are inconsistent, they are adjusted to the same dimension by zero padding or using dimensionality reduction methods such as principal component analysis (PCA) to ensure the feasibility of subsequent fusion operations; a feature fusion network based on Transformer is constructed; the processed image, electrical parameter and communication data feature vectors are sequentially connected to form a long sequence and input into the Transformer; the multi-head attention mechanism of Transformer will Automatically calculate the attention weights between different feature vectors, focusing on the importance of each feature in fault diagnosis; after being processed by multiple Transformer layers, different types of feature vectors are allowed to achieve deep interaction and fusion under the action of self-attention and cross-attention, and finally output the fused feature vector; the results obtained by fusion of feature layers and the results obtained by preliminary diagnosis using simple classification models (such as support vector machines SVM and logistic regression) of each source data feature vector separately are used as "evidence" in Dempster-Shafer evidence theory; an initial degree of confidence is assigned to each piece of evidence, which reflects the degree to which the evidence supports the conclusion of a certain fault category; then, using the combination rules in evidence theory, calculate The joint trust and likelihood of each fault category under multiple evidences; by comparing the joint trust and likelihood corresponding to different fault categories, the final fused fault diagnosis decision is obtained; a hybrid neural network model composed of convolutional neural network (CNN), recurrent neural network (RNN) and graph neural network (GNN) is built; the comprehensive feature vector after feature fusion is input into the corresponding sub-network according to different characteristics such as image, time sequence, topological relationship, etc.; CNN extracts and abstracts image-related features through convolutional layer and pooling layer; RNN uses its cyclic structure to process the time sequence features in electrical parameters and communication data, and captures the change law of data over time; GNN constructs graph structure according to the connection relationship between electronic splicing screen devices The system then constructs a network structure to analyze the topological relationships between devices and the fault propagation paths. The processed results of each sub-network are integrated through a fully connected layer, and the fault diagnosis results are finally output. A small amount of sample data of new fault types is prepared. The parameters of the hybrid neural network fault diagnosis model are initialized. For different tasks (i.e., samples of different new fault types), one or more gradient descent updates are performed on each sample to obtain the model parameters for each task. The average loss of these updated parameters across all tasks is calculated, and the initial model parameters are updated through backpropagation using this average loss as the target. In this way, the model can quickly adapt to newly emerging fault types, and when faced with a small number of new samples, only a small number of gradient updates are required to achieve good diagnostic results.The comprehensive feature vector obtained through feature fusion is fully input into the optimized hybrid neural network fault diagnosis model. The model calculates and processes the input feature vector layer by layer according to the trained parameters and network structure. After feature extraction and fusion of each sub-network and decision-making at the fully connected layer, it finally outputs the fault diagnosis results of the electronic splicing screen, including information such as fault type and possible fault cause.

[0031] In this embodiment, fault type: clearly points out the fault category of the electronic splicing screen, such as display fault (flower screen, black screen, color cast, etc.), electrical fault (power module fault, driver board fault, etc.), communication fault (signal transmission interruption, protocol error, etc.), mechanical fault (abnormal splicing gap, loose bracket, etc.); fault location: accurately locate the specific location of the fault. For a screen composed of multiple splicing units, determine which splicing screen, which circuit board, or which communication line has a problem; if it is a software fault, point out whether there is a problem with the operating system, driver, or specific configuration file; fault cause: analyze the root cause of the fault, which may be aging of electronic components, external environment interference (such as unstable voltage, power failure, etc.), etc. magnetic interference), software code vulnerabilities, human error, loose connection, or the result of multiple factors; Fault severity: graded assessment of the severity of the fault, such as minor fault (only affects local display effect, does not affect the overall function), moderate fault (partial function is limited, affecting normal use), severe fault (the equipment is basically unable to operate and needs to be handled immediately); Fault probability: combined with historical data and current operating status, gives a probability value of the possibility of fault occurrence, helps predict fault risks, and provides a reference for maintenance decisions; Associated impact range: describes the impact range of the fault on other parts of the electronic splicing screen and the entire system, such as whether it will cause abnormalities in adjacent splicing units, whether it will affect the interaction with external devices, etc.

[0032] The above shows and describes 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 above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for automatic diagnosis and repair of electronic splicing screen faults, characterized in that: The method comprises the following steps: S1. During the operation of the electronic splicing screen, multiple sensors and monitoring devices deployed in the splicing screen system are used to collect multi-source data in real time; S2. Preprocessing and feature extraction are performed on the collected multi-source data to obtain image feature vectors, electrical feature vectors, and communication feature vectors; S3, fusing the extracted image feature vector, electrical feature vector, and communication feature vector to form a comprehensive feature vector, inputting the comprehensive feature vector into a fault diagnosis model for processing, and outputting a fault diagnosis result; S4. Determine the repair strategy based on the fault diagnosis results and automatically perform the repair operation. After the repair is completed, collect data again for verification. If the fault is still not eliminated, optimize the repair strategy.

2. The method for automatic diagnosis and repair of electronic splicing screen faults according to claim 1, characterized in that: Step S1 includes the following steps: S11, using a high-definition camera to shoot the display screen of the spliced screen in real time to obtain display image data; S12. Collect the working current, voltage and temperature data of each module of the splicing screen through the current sensor, voltage sensor and temperature sensor to obtain electrical parameter data; S13, monitoring the communication signals between the splicing screen and the signal source and the control host, recording the communication protocol data, data transmission rate and number of error frames, and obtaining the communication data.

3. The method for automatic diagnosis and repair of electronic splicing screen faults according to claim 1, characterized in that: Step S2 includes the following steps: S21. For image data from multi-source data, a GAN network is used to perform denoising, and an image registration algorithm based on feature point matching is used for image registration. An improved convolutional neural network algorithm is used to extract features from the pre-processed image data to obtain image feature vectors. S22. For electrical parameter data in multi-source data, the isolation forest algorithm is used to detect electrical parameter anomalies. The dynamic time warping algorithm is combined to perform pattern matching on the changing trends of the electrical parameters. The ReliefF algorithm is used to calculate feature weights. The electrical parameter features are screened from the preprocessed electrical parameter data, and the electrical feature vector is obtained through principal component analysis and dimensionality reduction. S23. For the communication data in the multi-source data, perform semantic analysis on the communication data and convert it into a word vector sequence. Then input it into the LSTM network to capture the temporal dependency in the communication data, and combine it with the wavelet transform algorithm to extract the communication feature vector.

4. The method for automatically diagnosing and repairing electronic splicing screen faults according to claim 3, wherein: Step S21 includes the following steps: Input the image data into the GAN generator for denoising to generate a denoised image; The denoised image is layered using a Gaussian pyramid. The FAST corner detector is used to detect feature points in each layer. The corner points are filtered using the Harris response function to obtain the top N feature points with the highest response values. For each feature point, its main direction is calculated and a rotation-invariant BRIEF descriptor is generated. Use bidirectional kd-tree for fast feature point matching, calculate the Euclidean distance of matching point pairs and use ratio test to eliminate false matches; Bilinear interpolation is used for sub-pixel resampling. A local adaptive window is introduced during the interpolation process to complete image registration and obtain preprocessed image data. Construct a deep separable convolutional neural network architecture, which includes 5 residual modules and 3 attention modules. Each attention module consists of channel attention and spatial attention. The preprocessed image data is input into a depthwise separable convolutional neural network. The channel attention adopts the squeeze excitation mechanism of SENet, the spatial attention generates a 2D attention map through convolution, and the image feature vector is output from the global average pooling layer.

5. The method for automatic diagnosis and repair of electronic splicing screen faults according to claim 3, characterized in that: Step S22 includes the following steps: The electrical parameter data is divided into sliding windows, and the BiLSTM network is used to predict and fill in the missing data of the electrical parameter data; Extract a sample subset from the electrical parameter data after missing value processing, recursively randomly select features and split points in the feature space to generate an isolation tree, and calculate the path length of each sample point in all isolation trees to determine the abnormal point; For the detected abnormal points, the dynamic time warping algorithm is used to perform pattern matching of the change trend, find the optimal alignment path, and obtain the pre-processed electrical parameter data; Apply the ReliefF algorithm to calculate the weights of features in the preprocessed electrical parameter data and filter out features whose weights are higher than a preset threshold; The filtered electrical parameter features are subjected to principal component analysis and dimensionality reduction to obtain the electrical feature vector.

6. The method for automatically diagnosing and repairing electronic splicing screen faults according to claim 3, wherein: Step S23 includes the following steps: Parse communication data layer by layer, extract protocol header fields and payload content, and perform data cleaning to identify and filter invalid data packets and noise data; The Word2Vec algorithm is used to convert the denoised data into a word vector sequence, which is then input into the LSTM network. The long-distance temporal dependencies are captured through the dynamic adjustment of the forget gate, input gate, and output gate. The time series features output by LSTM are decomposed by wavelet transform, and the time series features extracted by LSTM and the frequency domain features obtained by wavelet transform are spliced to form a communication feature vector.

7. The method for automatically diagnosing and repairing electronic splicing screen faults according to claim 1, wherein: Step S3 includes the following steps: S31, for the image feature vector, the electrical feature vector, and the communication feature vector, using a normalization method to map all feature values to the same interval; S32. Use a Transformer-based feature fusion network to deeply interact and fuse the processed image feature vector, electrical feature vector, and communication feature vector under the action of self-attention and cross-attention, and output the fused feature vector; S33. Input the fused feature vector into the fault diagnosis model composed of CNN network, RNN network and GNN network, and integrate it through the fully connected layer to output the fault diagnosis result of the electronic splicing screen.

8. The method for automatically diagnosing and repairing electronic splicing screen faults according to claim 7, wherein: Step S33 includes the following steps: The CNN network extracts and abstracts images through convolutional layers and pooling layers; The RNN network uses a cyclic structure to process the temporal characteristics of the data and capture the changing patterns of the data over time; The GNN network constructs a graph structure based on the connection relationship between electronic splicing screen devices and analyzes the topological relationship and fault propagation path between devices.

9. The method for automatically diagnosing and repairing electronic splicing screen faults according to claim 7, wherein: The fault diagnosis result includes at least the fault type, fault location and fault cause.

10. A system for implementing the method for automatic fault diagnosis and repair of an electronic splicing screen according to claim 1, characterized in that: The system includes: Real-time acquisition module, used to collect multi-source data in real time during the operation of the electronic splicing screen through various sensors and monitoring equipment deployed in the splicing screen system; Feature extraction module, used to preprocess and extract features from the collected multi-source data to obtain image feature vectors, electrical feature vectors, and communication feature vectors; The fault diagnosis module is used to fuse the extracted image feature vector, electrical feature vector and communication feature vector to form a comprehensive feature vector, input the comprehensive feature vector into the fault diagnosis model for processing, and output the fault diagnosis result; The automatic repair module is used to determine the repair strategy based on the fault diagnosis results and automatically perform the repair operation. After the repair is completed, data is collected again for verification. If the fault is still not eliminated, the repair strategy is optimized.

Images