An elevator door failure detection method based on image enhancement

Abstract

The application discloses an elevator door fault detection method based on image enhancement, and relates to the technical field of image enhancement, and comprises the following steps: collecting running state image data in the running process of an elevator door, and generating an elevator door standard image sequence; performing image enhancement processing, and generating an elevator door enhanced image sequence; constructing a door body mixed visual state matrix; performing independent component separation by using a FastICA algorithm, and generating a door body independent fault source matrix; constructing an improved Hindmarsh Rose neural network, and generating a door body fault evolution state sequence; determining an elevator door fault evolution path, and generating an elevator door fault detection result and an elevator door fault early warning result. By introducing the FastICA algorithm and the improved Hindmarsh Rose neural network, the elevator door fault source decoupling analysis, the fault propagation path modeling, the fault type identification and the fault risk early warning are realized.

Description

Technical Field

[0001] This invention relates to the field of image enhancement technology, and in particular to an elevator door fault detection method based on image enhancement. Background Technology

[0002] With the continuous increase in the number of elevators in urban buildings, the elevator door system, as the most frequently moving actuator during elevator operation, directly affects elevator safety and passenger experience. Over long-term operation, elevator doors are susceptible to mechanical wear, dust accumulation on guide rails, component aging, installation deviations, and changes in the external environment, gradually leading to malfunctions such as abnormal door gaps, door misalignment, poor closure, and guide rail jamming. To ensure safe elevator operation, continuous monitoring of the elevator door's operating status is usually necessary to promptly detect potential faults. With the development of image acquisition equipment and intelligent monitoring technology, utilizing visual information for automatic detection of elevator door operation is gradually becoming an important research direction in the field of intelligent elevator operation and maintenance.

[0003] In existing technologies, most elevator door fault detection methods acquire images of the elevator door in operation using camera equipment. They then utilize edge detection, target recognition, and feature extraction to obtain operational features such as door gap width, door leaf position, and guide rail status. These features are then combined with threshold judgment, statistical analysis, or machine learning models to achieve fault identification. Some technologies improve fault detection accuracy by constructing time-series feature sequences to analyze the changes in the elevator door's operating state. However, existing methods typically focus on single fault feature analysis or single-moment state determination, failing to adequately consider the coupled influence of multiple fault factors and making it difficult to effectively characterize the state evolution patterns during fault propagation. Furthermore, their ability to depict fault development paths and fault risk change trends is limited, leading to inaccurate fault type identification and insufficient early warning capabilities under complex operating conditions.

[0004] Therefore, how to provide an image enhancement-based elevator door fault detection method is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an image-enhanced elevator door fault detection method. This invention utilizes image enhancement technology, independent component separation technology, and an improved Hindmarsh Rose neural network to perform visual state analysis on the door gap region, door leaf edge region, and sill guide rail region during elevator door operation. The implementation process of elevator door fault detection, fault propagation modeling, and fault early warning is described in detail. By extracting features such as door gap width variation, door closure offset, door leaf edge displacement, sill guide rail texture variation, door closure speed, and door jitter, a hybrid visual state matrix of the door is constructed. Independent component separation technology is used to generate an independent fault source matrix of the door, achieving decoupled representation of multiple fault sources. Furthermore, an improved Hindmarsh Rose neural network, incorporating a piecewise-gated hypergraph coupling structure, a hierarchical time-scale memory mechanism, and a heterogeneous homogeneous track structure, is used to model the fault state evolution process, realizing fault propagation path analysis, fault type identification, and fault risk early warning. This invention's method has the advantages of strong fault feature representation capability, accurate fault propagation relationship characterization, high fault identification accuracy, and strong fault early warning capability.

[0006] An image enhancement-based elevator door fault detection method according to an embodiment of the present invention includes: Collect operational status image data of the elevator door during operation, perform preprocessing on the operational status image data, and generate a standard image sequence of the elevator door; Image enhancement processing is performed on the standard image sequence of the elevator door to enhance the structural features of the door gap area, door lock area, door leaf edge area and sill guide rail area, and generate an enhanced image sequence of the elevator door. Based on the enhanced image sequence of elevator doors, the features of door gap width change, door closure offset, door leaf edge displacement, sill guide rail texture change, door closure speed and door shaking are extracted to construct a hybrid visual state matrix of the door. The hybrid visual state matrix of the door is centered and whitened, and the FastICA algorithm is used to separate the independent components of the processed hybrid visual state matrix of the door to generate the independent fault source matrix of the door. An improved HindmarshRose neural network is constructed based on the independent fault source matrix of the door body. The state evolution of the multi-path fault propagation relationship between the abnormal sources of door gaps, abnormal sources of closure, offset sources of door leaves, stuck sources of ground sills and closure disturbance sources is carried out, and the heterogeneous and homogeneous fault state trajectory is constructed to generate the door body fault evolution state sequence. The elevator door fault evolution path is determined based on the door fault evolution state sequence, and the elevator door fault type is identified. By combining the door independent fault source matrix, the door fault evolution state sequence, and the elevator door fault evolution path, elevator door fault detection results and elevator door fault early warning results are generated.

[0007] Optionally, the operating status image data specifically includes door opening process image data, door closing process image data, door gap area image data, door edge area image data, sill guide rail area image data, and acquisition timestamp data.

[0008] Optionally, the preprocessing of the running state image data specifically includes image distortion correction, inter-frame registration, image denoising, brightness normalization, and time alignment.

[0009] Optionally, generating the enhanced elevator door image sequence includes: Based on the collected timestamp data, the images of the door gap area, door leaf edge area, and sill guide rail area in the standard image sequence of the elevator door are read, and the mean gray value of the door gap, the mean gradient value of the door gap edge, the mean gray value of the door leaf edge, the mean gradient value of the door leaf edge, the mean gray value of the sill guide rail, and the mean texture response value of the sill guide rail are calculated respectively. The door gap enhancement coefficient is calculated based on the brightness loss corresponding to the average gray value of the door gap and the edge loss corresponding to the average gradient of the door gap edge. The door leaf edge enhancement coefficient is calculated based on the brightness loss corresponding to the average gray value of the door leaf edge and the edge loss corresponding to the average gradient of the door leaf edge. The sill rail enhancement coefficient is calculated based on the brightness loss corresponding to the average gray value of the sill rail and the texture loss corresponding to the average texture response of the sill rail. Enhancement processing is performed on the door gap region image, door leaf edge region image, and sill guide rail region image according to the door gap enhancement coefficient, door leaf edge enhancement coefficient, and sill guide rail enhancement coefficient, respectively, to generate door gap enhanced region image, door leaf edge enhanced region image, and sill guide rail enhanced region image, and then written into the elevator door standard image sequence to generate elevator door enhanced image sequence.

[0010] Optionally, constructing the gate hybrid visual state matrix includes: Read the elevator door enhanced image sequence according to the collected timestamp data, extract the left and right edge positions of the door gap from the door gap enhanced area image, extract the door leaf edge position from the door leaf edge enhanced area image, and extract the sill guide rail texture response value from the sill guide rail enhanced area image; The door gap width is calculated based on the left and right edge positions of the door gap under the same collection timestamp. The door gap width change feature is extracted based on the difference in door gap width corresponding to adjacent collection timestamps. The door closure offset feature is extracted based on the position offset between the center position of the door gap and the closing center position of the elevator door. The door leaf edge displacement feature is extracted based on the difference in door leaf edge positions corresponding to adjacent collection timestamps. The textural change features of the sill guide rail are extracted based on the difference in the texture response values ​​of the sill guide rail corresponding to adjacent acquisition timestamps. The door closing speed features are extracted based on the door gap width change and time interval corresponding to adjacent acquisition timestamps. The door shaking features are extracted based on the door edge position fluctuation corresponding to consecutive acquisition timestamps. Based on the collected timestamp data, the characteristics of door gap width change, door closure offset, door edge displacement, sill guide rail texture change, door closure speed, and door shaking are written into the same state row, and all state rows are arranged in the order of collection time to construct a hybrid visual state matrix of the door.

[0011] Optionally, generating the independent fault source matrix of the gate includes: The state features in the hybrid visual state matrix of the door are divided into channels. The door gap width change feature, door leaf edge displacement feature, door closing speed feature and door shaking feature are divided into high frequency motion channels, and the door closing offset feature and sill guide rail texture change feature are divided into low frequency structure channels. The state features in the high-frequency motion channel and the low-frequency structure channel are centered respectively to eliminate the overall offset between different state features in the high-frequency motion channel and the low-frequency structure channel, thus obtaining the high-frequency centered state submatrix and the low-frequency centered state submatrix. A streaming Sketch-Cov statistical structure is constructed based on the high-frequency centered state submatrix and the low-frequency centered state submatrix, and the second-order statistical information of the high-frequency motion channel and the low-frequency structure channel is compressed and stored. High-frequency block diagonal whitening matrix and low-frequency block diagonal whitening matrix are constructed respectively. The high-frequency centered state submatrix is ​​whitened based on the high-frequency block diagonal whitening matrix, and the low-frequency centered state submatrix is ​​whitened based on the low-frequency block diagonal whitening matrix. The whitened high-frequency motion channel and the whitened low-frequency structure channel are then aligned and spliced ​​according to the acquisition timestamp data to obtain the whitened state matrix. Based on the FastICA algorithm, independent component separation is performed on the whitening state matrix to construct a topological correlation graph between the door gap width variation feature, door closure offset feature, door leaf edge displacement feature, sill guide rail texture variation feature, door closure speed feature, and door shaking feature. Sparse constraints are applied to the independent component separation process to generate a feature-related independent component set. Based on the feature-related independent component set, independence objective, sparsity objective and fault discrimination objective are established. The independent separation direction is updated through multi-objective collaborative optimization. An adversarial discrimination mechanism is introduced to reduce the correlation between different independent components, and an independent component matrix is ​​obtained. Calculate the correspondence between each independent component in the independent component matrix and the characteristics of door gap width variation, door closure offset, door leaf edge displacement, sill guide rail texture variation, door closure speed, and door shaking. Label each independent component as a door gap anomaly source, closure anomaly source, door leaf offset source, sill jamming source, and closure disturbance source, and construct an independent fault source matrix for the door.

[0012] Optionally, generating the gate failure evolution state sequence includes: An improved Hindmarsh Rose neural network is constructed, comprising a fault state mapping unit, a multipath coupling evolution unit, a slowly varying memory modulation unit, and a heteroclinic orbit unit. The fault state mapping unit reads the independent fault source matrix of the door body, maps the abnormal sources of door gaps, abnormal sources of closure, sources of door leaf offset, sources of ground sill jamming, and sources of closure disturbance to the corresponding fault state nodes, and forms a set of fault state nodes. The multi-path coupling evolution unit reads the set of fault state nodes, constructs the gate structure hyperedge, the closure control hyperedge and the gate control communication relationship, generates the piecewise-gated hypergraph, and constructs the mechanical clearance hysteresis channel, the closure control hysteresis channel and the damping hysteresis channel to form a tensor hysteresis coupling structure, executes multi-path fault propagation evolution, and generates a multi-path fault propagation state sequence. The slow-variable memory modulation unit reads the multi-path fault propagation state sequence, adjusts the state evolution order according to the door closing speed characteristics and door jitter characteristics, and adjusts the propagation hysteresis according to the door gap width change characteristics, door closing offset characteristics and sill guide rail texture change characteristics, thus constructing a variable-order-variable-hysteresis evolution structure. In the variable-order-variable-hysteresis evolution structure, event-triggered state nodes and continuous evolution state nodes are set, and the state transition results are fused with the continuous evolution results to generate an event-continuous fused fault propagation state sequence. The heterodox and coherent orbit unit reads the event continuous fusion fault propagation state sequence, takes each fault state node as a game participant node, takes the excitation state of each fault state node as the game strategy, takes the coupling strength between fault state nodes as the game payoff, performs multi-objective game modulation on the event continuous fusion fault propagation state sequence, generates the modulated fault propagation state sequence, constructs the heterodox and coherent fault state orbit, and outputs the gate fault evolution state sequence. The improved Hindmarsh Rose neural network was trained, and the fault state prediction error and fault trajectory reconstruction error were used as joint optimization objectives. The parameters in the fault state mapping unit, multi-path coupling evolution unit, slow-varying memory modulation unit and hetero-homogeneous trajectory unit were continuously optimized. Training was stopped when the change of the joint optimization objective was less than 0.001 for 10 consecutive iterations.

[0013] Optionally, generating elevator door fault detection results and elevator door fault early warning results includes: Read the door failure evolution state sequence according to the collected timestamp data, extract the state transition relationship between normal closing state, door leaf offset state, door gap abnormal state, closing abnormal state and closing failure state, and generate the elevator door failure evolution path. Elevator door fault types are identified based on the termination state of the elevator door fault evolution path, state transition characteristics, and fault contribution relationships in the independent fault source matrix of the door body. Read the fault source information and status information corresponding to the elevator door fault type, and write the elevator door fault type, fault source information, status information, fault location and collection timestamp data into the elevator door fault detection result; Read the fault states that have not reached the closed failure state in the elevator door fault evolution path, and generate elevator door fault early warning results by combining the changing trend and state transition direction of the corresponding fault source information.

[0014] The beneficial effects of this invention are: This invention proposes an image enhancement-based elevator door fault detection method. First, targeted image enhancement processing is performed on the door gap area, door leaf edge area, and sill guide rail area during elevator door operation. By improving the quality of edge and structural information in key areas, the accuracy of extracting door gap width variation features, door closing offset features, door leaf edge displacement features, sill guide rail texture variation features, door closing speed features, and door shaking features is improved. This reduces the impact of image quality fluctuations, illumination changes, and local occlusion on fault detection results, providing a stable and reliable data foundation for fault analysis.

[0015] This invention further utilizes independent component separation technology to decouple various visual state features, separating the independent fault sources corresponding to different fault factors such as door gap abnormalities, closure abnormalities, door leaf offset, sill jamming, and closure disturbance, thereby achieving accurate characterization of the contribution relationship of fault sources under complex fault conditions and avoiding misjudgment problems caused by the superposition of multiple fault features.

[0016] By constructing an improved Hindmarsh Rose neural network that incorporates a piecewise-gated hypergraph coupling structure, a hierarchical time-scale memory mechanism, and a heterodox homology trajectory structure, the state transition relationships and evolutionary laws during fault propagation are modeled, enabling dynamic tracking and analysis of elevator door fault development paths. Based on the fault evolution path, fault contribution relationships, and fault state change trends, fault detection and early warning results are generated. This not only enables fault type identification but also allows for the early detection of risk trends before the fault develops into a closed failure state, improving the accuracy of elevator door fault detection, the reliability of fault location, and the timeliness of fault early warning. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an elevator door fault detection method based on image enhancement proposed in this invention; Figure 2 This is a data flow diagram illustrating the FastICA algorithm for an elevator door fault detection method based on image enhancement proposed in this invention. Figure 3 This is a schematic diagram of the structure of an improved HindmarshRose neural network for an elevator door fault detection method based on image enhancement proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 , Figure 2 and Figure 3 An image enhancement-based elevator door fault detection method includes: Collect operational status image data of the elevator door during operation, perform preprocessing on the operational status image data, and generate a standard image sequence of the elevator door; Image enhancement processing is performed on the standard image sequence of the elevator door to enhance the structural features of the door gap area, door lock area, door leaf edge area and sill guide rail area, and generate an enhanced image sequence of the elevator door. Based on the enhanced image sequence of elevator doors, the features of door gap width change, door closure offset, door leaf edge displacement, sill guide rail texture change, door closure speed and door shaking are extracted to construct a hybrid visual state matrix of the door. The hybrid visual state matrix of the door is centered and whitened, and the FastICA algorithm is used to separate the independent components of the processed hybrid visual state matrix of the door to generate the independent fault source matrix of the door. An improved HindmarshRose neural network is constructed based on the independent fault source matrix of the door body. The state evolution of the multi-path fault propagation relationship between the abnormal sources of door gaps, abnormal sources of closure, offset sources of door leaves, stuck sources of ground sills and closure disturbance sources is carried out, and the heterogeneous and homogeneous fault state trajectory is constructed to generate the door body fault evolution state sequence. The elevator door fault evolution path is determined based on the door fault evolution state sequence, and the elevator door fault type is identified. By combining the door independent fault source matrix, the door fault evolution state sequence, and the elevator door fault evolution path, elevator door fault detection results and elevator door fault early warning results are generated.

[0020] In this embodiment, the operating status image data specifically includes door opening process image data, door closing process image data, door gap area image data, door edge area image data, sill guide rail area image data, and acquisition timestamp data.

[0021] In this embodiment, the preprocessing of the running state image data specifically includes image distortion correction, inter-frame registration, image denoising, brightness normalization, and time alignment.

[0022] In this embodiment, generating the elevator door enhanced image sequence includes: Based on the collected timestamp data, images of the door gap region, door leaf edge region, and sill guide rail region from the standard image sequence of the elevator door are read. The mean grayscale value of the door gap, the mean gradient value of the door gap edge, the mean grayscale value of the door leaf edge, the mean gradient value of the door leaf edge, the mean grayscale value of the sill guide rail, and the mean texture response value of the sill guide rail are calculated respectively. The mean grayscale value of the door gap, the mean gradient value of the door gap edge, the mean grayscale value of the door leaf edge, the mean gradient value of the door leaf edge, the mean grayscale value of the sill rail, and the mean texture response value of the sill rail are calculated separately. Read the images of the door gap area, door edge area, and sill guide rail area corresponding to the current moment according to the collected timestamp data, and divide the gray value of each pixel in each area image by 255 to obtain the normalized gray value. The mean grayscale value of the door gap is obtained by summing all normalized grayscale values ​​within the door gap region image and dividing by the total number of pixels. The normalized grayscale difference between horizontally adjacent pixels and vertically adjacent pixels in the door gap region image is then calculated. The mean gradient value of the door gap edge is obtained by summing the absolute values ​​of all normalized grayscale differences and dividing by the total number of grayscale differences. Similarly, the mean grayscale value of the door edge is obtained by summing all normalized grayscale values ​​within the door leaf edge region image and dividing by the total number of pixels. The mean grayscale value of the door leaf edge is then calculated by summing the absolute values ​​of horizontally adjacent pixels and vertically adjacent pixels in the door leaf edge region image. The normalized grayscale difference between adjacent pixels in the direction is summed and divided by the total number of grayscale differences to obtain the average gradient value of the door edge. The normalized grayscale values ​​in the sill rail area image are summed and divided by the total number of pixels to obtain the average grayscale value of the sill rail. The normalized grayscale difference between adjacent pixels in the horizontal direction and adjacent pixels in the vertical direction in the sill rail area image is calculated. The normalized grayscale difference between adjacent pixels in the horizontal direction and adjacent pixels in the vertical direction is summed and divided by the total number of grayscale differences to obtain the average texture response value of the sill rail. The door gap enhancement coefficient is calculated based on the brightness loss corresponding to the average grayscale value of the door gap and the edge loss corresponding to the average gradient value of the door gap edge. The door panel edge enhancement coefficient is calculated based on the brightness loss corresponding to the average grayscale value of the door panel edge and the edge loss corresponding to the average gradient value of the door panel edge. The sill rail enhancement coefficient is calculated based on the brightness loss corresponding to the average grayscale value of the sill rail and the texture loss corresponding to the average texture response value of the sill rail. Where: The brightness loss corresponding to the average grayscale value of the door gap refers to the difference between the standard brightness value of the image of the door gap area and the average grayscale value of the door gap, where the standard brightness value is taken as 0.8; The edge missing amount corresponding to the mean gradient of the door gap edge refers to the difference between the standard edge gradient value and the mean gradient of the door gap edge in the image of the door gap region, where the standard edge gradient value is taken as 0.6; The calculation of the door gap reinforcement coefficient is as follows: Read the brightness loss corresponding to the average gray value of the door gap and the edge loss corresponding to the average gradient of the door gap edge. Multiply the brightness loss by 0.4 and the edge loss by 0.6. Sum the products to obtain the door gap enhancement coefficient. The brightness loss corresponding to the average gray value of the door edge refers to the difference between the standard brightness value of the image of the door edge region and the average gray value of the door edge, where the standard brightness value is 0.8; The edge missing value corresponding to the mean gradient of the door leaf edge refers to the difference between the standard edge gradient value and the mean gradient of the door leaf edge region image, where the standard edge gradient value is taken as 0.6. The calculation of the door leaf edge reinforcement coefficient is as follows: Read the brightness loss corresponding to the average gray value of the door edge and the edge loss corresponding to the average gradient of the door edge. Multiply the brightness loss by 0.4 and the edge loss by 0.6. Sum the products to obtain the door edge enhancement coefficient. The texture response of the ground sill rail refers to the quantitative representation value of the degree of gray-level change between adjacent pixels in the image of the ground sill rail region; The brightness loss corresponding to the grayscale mean of the ground sill guide rail refers to the difference between the standard brightness value of the image of the ground sill guide rail area and the grayscale mean of the ground sill guide rail, where the standard brightness value is taken as 0.8; The texture missing amount corresponding to the mean texture response of the sill rail refers to the difference between the standard texture response value of the sill rail region image and the mean texture response of the sill rail, where the standard texture response value is 0.5; The reinforcement coefficient of the ground sill guide rail is as follows: Read the brightness missing amount corresponding to the average gray value of the ground sill rail and the texture missing amount corresponding to the average texture response of the ground sill rail. Multiply the brightness missing amount by 0.5 and the texture missing amount by 0.5. Sum the product results to obtain the ground sill rail enhancement coefficient. Enhancement processing is performed on the door gap region image, door leaf edge region image, and sill guide rail region image according to the door gap enhancement coefficient, door leaf edge enhancement coefficient, and sill guide rail enhancement coefficient, respectively, generating door gap enhanced region image, door leaf edge enhanced region image, and sill guide rail enhanced region image, and then written into the elevator door standard image sequence to generate an elevator door enhanced image sequence, wherein: Enhancement processing is performed on the images of the door gap region, door leaf edge region, and sill guide rail region according to the enhancement coefficients of the door gap, door leaf edge, and sill guide rail, respectively. Specifically: Read the normalized grayscale value corresponding to each pixel in the door gap area image, multiply the normalized grayscale value by the door gap enhancement coefficient to obtain the enhanced grayscale value of the corresponding pixel, and assign 1 to pixels with enhanced grayscale values ​​greater than 1 to generate the door gap enhanced area image. Read the normalized grayscale value corresponding to each pixel in the door edge area image, multiply the normalized grayscale value by the door edge enhancement coefficient to obtain the enhanced grayscale value of the corresponding pixel, and assign 1 to pixels with enhanced grayscale values ​​greater than 1 to generate the door edge enhanced area image. Read the normalized grayscale value corresponding to each pixel in the sill rail area image, multiply the normalized grayscale value by the sill rail enhancement coefficient to obtain the enhanced grayscale value of the corresponding pixel, and assign 1 to pixels with enhanced grayscale values ​​greater than 1 to generate the sill rail enhanced area image. Generate an enhanced image sequence of elevator doors, specifically as follows: The original image frames in the standard image sequence of the elevator door are read according to the acquisition timestamp data. The door gap area in the original image frame is replaced with the door gap enhancement area image, the door leaf edge area in the original image frame is replaced with the door leaf edge enhancement area image, and the sill guide rail area in the original image frame is replaced with the sill guide rail enhancement area image. The enhanced image frames under the corresponding acquisition timestamp are obtained. All enhanced image frames are arranged in the order of acquisition timestamps to generate the elevator door enhanced image sequence.

[0023] In this embodiment, constructing the hybrid visual state matrix of the door includes: The elevator door enhanced image sequence was read according to the acquisition timestamp data. The left and right edge positions of the door gap were extracted from the door gap enhanced area image, the door leaf edge position was extracted from the door leaf edge enhanced area image, and the sill rail texture response value was extracted from the sill rail enhanced area image. The left and right edge positions of the door gap refer to the pixel coordinate positions corresponding to the left boundary and the right boundary of the door gap in the enhanced area image of the door gap. Door edge position refers to the pixel coordinate position corresponding to the outer boundary of the door in the enhanced image of the door edge; The door gap width is calculated based on the left and right edge positions of the door gap under the same acquisition timestamp. Door gap width variation features are extracted based on the difference in door gap width between adjacent acquisition timestamps. Door closure offset features are extracted based on the positional offset between the center position of the door gap and the closing center position of the elevator door. Door leaf edge displacement features are extracted based on the difference in door leaf edge positions between adjacent acquisition timestamps. The calculation of the door gap width is as follows: Read the horizontal pixel coordinates of the left edge and the right edge of the door gap at the same collection timestamp, calculate the difference between the horizontal pixel coordinates of the right edge and the left edge of the door gap, and obtain the width of the door gap at the corresponding collection timestamp. The extraction of door gap width variation features is as follows: Read the door gap width corresponding to two adjacent collection timestamps in the order of collection timestamps, calculate the difference between the door gap width corresponding to the current collection timestamp and the door gap width corresponding to the previous collection timestamp, and obtain the door gap width change characteristics. The extraction of the door closure offset features is as follows: Read the horizontal pixel coordinates corresponding to the left edge of the door gap and the right edge of the door gap. Add the horizontal pixel coordinates corresponding to the left edge of the door gap and the right edge of the door gap and divide by 2 to obtain the center position of the door gap. Read the horizontal pixel coordinates corresponding to the center position of the elevator door closure. Calculate the absolute value of the difference between the horizontal pixel coordinates corresponding to the center position of the door gap and the horizontal pixel coordinates corresponding to the center position of the elevator door closure to obtain the door closure offset feature. The center position of the elevator door closure refers to the center pixel position when the left and right doors of the elevator door are completely closed. The extraction of door edge displacement features is as follows: Read the door edge position corresponding to two adjacent acquisition timestamps in the order of acquisition timestamps, read the horizontal pixel coordinate value of the door edge position corresponding to the current acquisition timestamp and the horizontal pixel coordinate value of the door edge position corresponding to the previous acquisition timestamp, calculate the absolute value of the difference between the two, and obtain the door edge displacement feature. The textural variation features of the sill guide rail are extracted based on the difference in texture response values ​​corresponding to adjacent acquisition timestamps. The door closing speed features are extracted based on the change in door gap width and time interval corresponding to adjacent acquisition timestamps. Door shaking features are extracted based on the fluctuation of door leaf edge positions corresponding to consecutive acquisition timestamps. The extraction of texture variation features of the sill guide rail is as follows: Read the texture response values ​​of the sill rail corresponding to two adjacent collection timestamps in the order of collection timestamps, calculate the difference between the texture response value of the sill rail corresponding to the current collection timestamp and the texture response value of the sill rail corresponding to the previous collection timestamp, and obtain the texture change characteristics of the sill rail. The extraction of the door closing speed features is as follows: Read the door gap width corresponding to two adjacent collection timestamps in the order of collection timestamps, calculate the difference between the door gap width corresponding to the current collection timestamp and the door gap width corresponding to the previous collection timestamp to obtain the door gap width change, read the time interval between the current collection timestamp and the previous collection timestamp, and divide the door gap width change by the time interval to obtain the door closing speed characteristic. The extraction of door shaking features is as follows: The door edge position corresponding to five collection timestamps is read continuously in the order of collection timestamps. The horizontal pixel coordinate value of the door edge position corresponding to each collection timestamp is read. The difference between the maximum and minimum values ​​of the five horizontal pixel coordinate values ​​is calculated to obtain the fluctuation amount of the door edge position. The fluctuation amount of the door edge position is used as the door shaking feature. Based on the collected timestamp data, the characteristics of door gap width change, door closure offset, door edge displacement, sill guide rail texture change, door closure speed, and door shaking are written into the same state row, and all state rows are arranged in the order of collection time to construct a hybrid visual state matrix of the door.

[0024] In this embodiment, generating the independent fault source matrix of the gate includes: The state features in the hybrid visual state matrix of the door are divided into channels. The door gap width change feature, door leaf edge displacement feature, door closing speed feature and door shaking feature are divided into high frequency motion channels, and the door closing offset feature and sill guide rail texture change feature are divided into low frequency structure channels. Centering processing is performed on the state features in the high-frequency motion channel and the low-frequency structure channel respectively to eliminate the overall offset between different state features in the high-frequency motion channel and the low-frequency structure channel, resulting in a high-frequency centered state submatrix and a low-frequency centered state submatrix, where: The state features in the high-frequency motion channel and the low-frequency structure channel are respectively centered, specifically as follows: The door gap width variation features, door leaf edge displacement features, door closing speed features, and door shaking features in the high-frequency motion channel are read in the order of the acquisition timestamps. The mean values ​​of the door gap width variation features, door leaf edge displacement features, door closing speed features, and door shaking features are calculated respectively. The feature values ​​corresponding to each acquisition timestamp are subtracted from the corresponding feature mean values ​​to obtain the centered door gap width variation features, centered door leaf edge displacement features, centered door closing speed features, and centered door shaking features. The door closing offset features and sill guide rail texture variation features in the low-frequency structure channel are centered in the same way. The high-frequency centered state submatrix and the low-frequency centered state submatrix are obtained as follows: The centralized door gap width variation features, centralized door leaf edge displacement features, centralized door closing speed features, and centralized door shaking features are arranged in the order of collection timestamps to form a high-frequency centralized state sub-matrix. The centralized door closing offset features and centralized sill guide rail texture variation features are arranged in the order of collection timestamps to form a low-frequency centralized state sub-matrix. A streaming Sketch-Cov statistical structure is constructed based on the high-frequency and low-frequency centered state submatrices, and the second-order statistical information of the high-frequency motion channel and the low-frequency structure channel is compressed and stored. High-frequency block diagonal whitening matrices and low-frequency block diagonal whitening matrices are constructed respectively, where: A streaming Sketch-Cov statistical structure is constructed based on high-frequency and low-frequency centered state submatrices, specifically as follows: The door gap width variation feature, door edge displacement feature, door closing speed feature, door shaking feature, door closing offset feature, and sill guide rail texture variation feature are read in the order of the acquisition timestamp. The door gap width variation feature, door edge displacement feature, door closing speed feature, door shaking feature, door closing offset feature, and sill guide rail texture variation feature corresponding to the same acquisition timestamp are arranged in a fixed order to form a feature group. The feature group corresponding to the current acquisition timestamp is multiplied by the transpose of the feature group corresponding to the current acquisition timestamp to obtain the second-order statistical matrix at the current time. The second-order statistical matrix at the current time is added to the cumulative second-order statistical matrix corresponding to the previous acquisition timestamp, and the cumulative second-order statistical matrix is ​​updated to obtain the streaming Sketch-Cov statistical structure. The second-order statistical information of the high-frequency motion channel and the low-frequency structure channel is compressed and stored, specifically as follows: Read the cumulative second-order statistical matrix from the streaming Sketch-Cov statistical structure, retain the main diagonal elements and the elements adjacent to the main diagonal elements, delete the remaining elements, and obtain the compressed second-order statistical information. Construct the high-frequency block diagonal whitening matrix and the low-frequency block diagonal whitening matrix respectively, as follows: Read the statistical results of the corresponding door gap width change characteristics, door leaf edge displacement characteristics, door closing speed characteristics and door shaking characteristics from the compressed second-order statistical information, form a high-frequency block diagonal covariance matrix, calculate the reciprocal of the square root of each diagonal element of the high-frequency block diagonal covariance matrix, and write it into the corresponding diagonal position to obtain the high-frequency block diagonal whitening matrix. Read the statistical results of the corresponding door closure offset features and ground sill guide rail texture change features in the compressed second-order statistical information, form a low-frequency block diagonal covariance matrix, calculate the reciprocal of the square root of each diagonal element of the low-frequency block diagonal covariance matrix, and write it into the corresponding diagonal position to obtain the low-frequency block diagonal whitening matrix. Whitening transformation is performed on the high-frequency centered state submatrix based on the high-frequency block diagonal whitening matrix, and whitening transformation is performed on the low-frequency centered state submatrix based on the low-frequency block diagonal whitening matrix. The whitened high-frequency motion channel and the whitened low-frequency structure channel are then aligned and stitched together according to the acquisition timestamp data to obtain the whitened state matrix, where: The whitening transformation of the high-frequency centered state submatrix is ​​performed based on the high-frequency block diagonal whitening matrix, specifically as follows: Read the door gap width variation features, door leaf edge displacement features, door closing speed features, and door shaking features from the high-frequency centered state sub-matrix. Multiply the feature values ​​corresponding to each acquisition timestamp by the element values ​​at the corresponding diagonal positions of the high-frequency block diagonal whitening matrix to obtain the whitened door gap width variation features, whitened door leaf edge displacement features, whitened door closing speed features, and whitened door shaking features. The whitening transformation of the low-frequency centered state submatrix is ​​performed based on the low-frequency block diagonal whitening matrix, specifically as follows: Read the door closure offset feature and the sill guide rail texture change feature in the low-frequency centered state sub-matrix, and multiply the feature value corresponding to each acquisition timestamp by the element value at the corresponding diagonal position of the low-frequency block diagonal whitening matrix to obtain the whitened door closure offset feature and the whitened sill guide rail texture change feature. The whitening state matrix is ​​obtained as follows: Read the whitened door gap width change feature, whitened door leaf edge displacement feature, whitened door closing speed feature, whitened door shaking feature, whitened door closing offset feature, and whitened sill guide rail texture change feature in the order of collection timestamp. Arrange the 6 whitening features corresponding to the same collection timestamp in the order of door gap width change feature, door leaf edge displacement feature, door closing speed feature, door shaking feature, door closing offset feature, and sill guide rail texture change feature to form a whitening feature group. Arrange all whitening feature groups in the order of collection timestamp to obtain the whitening state matrix. Based on the FastICA algorithm, independent component separation is performed on the whitening state matrix to construct a topological correlation graph among the features of door gap width variation, door closure offset, door edge displacement, sill guide rail texture variation, door closure speed, and door jitter. Sparse constraints are applied to the independent component separation process to generate a set of feature-correlated independent components, where: Independent component separation of the whitened state matrix is ​​performed based on the FastICA algorithm, specifically as follows: Six independent separation directions are established based on the six whitening features in the whitening state matrix. The separation coefficients corresponding to each independent separation direction are initialized to 0.5. Each independent separation direction is multiplied by the whitening feature group to obtain the corresponding independent component value. The fourth power of the difference between the independent component value and the mean of the independent component is calculated to obtain the fourth central moment. The square of the difference between the independent component value and the mean of the independent component is calculated and then the average of the two values ​​is squared to obtain the square of the second central moment. The difference between the fourth central moment and the square of the second central moment is calculated to obtain the non-Gaussianity evaluation value. Read the difference between the current independent component value and the mean of the independent components. Sum the absolute values ​​of all differences and divide by the total number of independent components to obtain the distribution deviation value. Multiply the current separation coefficient by the distribution deviation value to obtain the updated separation coefficient. Multiply the updated separation coefficient by the whitening feature group to obtain the new independent component value. Recalculate the fourth central moment, the square of the second central moment, and the non-Gaussianity evaluation value according to the new independent component value. Save the non-Gaussianity evaluation value corresponding to this iteration. Starting from the second iteration, calculate the difference between the current non-Gaussianity evaluation value and the non-Gaussianity evaluation value corresponding to the previous iteration. The iteration stops when the difference corresponding to 10 consecutive iterations is less than 0.001; the independent component values ​​corresponding to all independent separation directions are arranged in the order of acquisition timestamps to form an independent component matrix, and the independent component results are obtained. A topological relationship graph is constructed between the characteristics of door gap width variation, door closure offset, door leaf edge displacement, sill guide rail texture variation, door closure speed, and door vibration. Specifically: Read the cumulative correlation statistics matrix in the streaming Sketch-Cov statistical structure, establish connection edges for the corresponding features with correlation statistics values ​​greater than 0.3 in the cumulative correlation statistics matrix, and use the door gap width change feature, door closing offset feature, door leaf edge displacement feature, sill guide rail texture change feature, door closing speed feature, and door shaking feature as nodes. Map the connection edges with the corresponding nodes to obtain the topological correlation graph. Sparse constraints refer to retaining edges with an association statistic greater than 0.3 in the topological graph and deleting edges with an association statistic less than or equal to 0.3. Generate a set of feature-associative independent components, specifically: Read the connection relationships in the independent component results and topological association graph, associate the independent component values ​​with the connection relationship, and arrange all the associated independent component values ​​in the order of collection timestamp to obtain the feature-associated independent component set; Based on the feature-correlated independent component set, independence, sparsity, and fault discriminative objectives are established. The independent separation direction is updated through multi-objective collaborative optimization, and an adversarial discrimination mechanism is introduced to reduce the correlation between different independent components, resulting in an independent component matrix, where: Based on the feature-correlated independent component set, independence objectives, sparsity objectives, and fault discriminative objectives are established, specifically as follows: Read the independent component values ​​in the feature-associated independent component set, calculate the average of the absolute values ​​of the differences between any two independent component values ​​as the independence target, and calculate the ratio between the number of independent component values ​​corresponding to the connecting edges in the feature-associated independent component set and the total number of independent component values ​​as the sparsity target. Calculate the average of the absolute values ​​of the differences between each independent component value and the door gap width variation feature, door closure offset feature, door leaf edge displacement feature, sill guide rail texture variation feature, door closure speed feature, and door shaking feature as the fault discrimination target. The multi-objective collaborative optimization method refers to calculating the comprehensive optimization value according to the weight of independence objective (0.4), sparsity objective (0.2), and fault discrimination objective (0.4), and then performing iterative updates based on the comprehensive optimization value. The independent separation direction is updated through a multi-objective collaborative optimization approach, specifically as follows: Read the independence target, sparsity target, and fault discrimination target corresponding to the current independent separation direction, calculate the comprehensive optimization value according to the corresponding weight, multiply the difference between the current comprehensive optimization value and the previous comprehensive optimization value by 0.01 as the direction correction value, and superimpose the direction correction value onto the corresponding independent separation direction to obtain the updated independent separation direction. Recalculate the independent component value using the updated independent separation direction, and recalculate the independence target, sparsity target, and fault discrimination target. The adversarial discriminant mechanism refers to a mechanism that constructs a discriminant process and a correction process based on the degree of difference and similarity between independent components. The discriminant process identifies highly correlated independent components, and the correction process reduces the correlation between highly correlated independent components. An adversarial discriminant mechanism is introduced to reduce the correlation between different independent components, specifically: Read any two independent components from the independent component matrix, calculate the average of the absolute values ​​of the differences between the corresponding independent component values ​​at all collection timestamps, and obtain the component difference value. Independent components with a component difference value less than 0.1 are identified as highly correlated independent components. The independent component values ​​corresponding to the highly correlated independent components are increased and decreased by 0.01 times the component difference value, respectively, to obtain the corrected independent component values. The component difference values ​​between each independent component are recalculated. When the component difference values ​​corresponding to all independent components are greater than or equal to 0.1, the independent component values ​​with reduced correlation are obtained. The independent component matrix is ​​obtained as follows: Read the independent component values ​​after the correlation reduction according to the collection timestamp order, and arrange all independent component values ​​according to the correspondence of independent separation direction to form an independent component matrix; Calculate the correspondence between each independent component in the independent component matrix and the characteristics of door gap width variation, door closure offset, door leaf edge displacement, sill guide rail texture variation, door closure speed, and door vibration. Label each independent component as a door gap anomaly source, closure anomaly source, door leaf offset source, sill jamming source, and closure disturbance source, respectively, to construct an independent fault source matrix for the door, where: The correspondence between each independent component in the independent component matrix and the characteristics of door gap width variation, door closure offset, door leaf edge displacement, sill guide rail texture variation, door closure speed, and door vibration is calculated as follows: Calculate the average absolute value of the difference between each independent component and the characteristics of door gap width variation, door closing offset, door edge displacement, sill guide rail texture variation, door closing speed, and door shaking. The feature with the smallest average absolute value of the difference is determined as the associated feature of the corresponding independent component. The independent fault source matrix of the gate is constructed as follows: Independent components corresponding to the door gap width variation characteristics are marked as door gap anomaly sources, independent components corresponding to the door body closure offset characteristics are marked as closure anomaly sources, independent components corresponding to the door leaf edge displacement characteristics are marked as door leaf offset sources, independent components corresponding to the sill guide rail texture variation characteristics are marked as sill jamming sources, and independent components corresponding to the door body closing speed characteristics and door body shaking characteristics are marked as closure disturbance sources. All fault sources are arranged in the order of collection timestamps to form a door body independent fault source matrix.

[0025] In this embodiment, generating the gate failure evolution state sequence includes: An improved Hindmarsh Rose neural network is constructed, comprising a fault state mapping unit, a multipath coupling evolution unit, a slowly varying memory modulation unit, and a heteroclinic orbit unit, wherein: The improved Hindmarsh Rose neural network is constructed as follows: Based on the traditional Hindmarsh Rose neural network state variable mapping structure, a fault source state mapping process is introduced to obtain the fault state mapping unit. Based on the traditional Hindmarsh Rose neural network neuron connection structure, a multi-fault source association connection process is introduced to obtain the multi-path coupling evolution unit. Based on the traditional Hindmarsh Rose neural network slow-varying state variable update structure, a historical fault state modulation process is introduced to obtain the slow-varying memory modulation unit. Based on the traditional Hindmarsh Rose neural network state evolution output structure, a heterocentric homophonic orbit construction process is introduced to obtain the heterocentric homophonic orbit unit, thus forming an improved Hindmarsh Rose neural network. The fault state mapping unit reads the independent fault source matrix of the door body, maps the abnormal sources of door gaps, abnormal sources of closure, door leaf offset sources, sill jamming sources, and closure disturbance sources to corresponding fault state nodes, and forms a set of fault state nodes, wherein: The fault state mapping unit includes: Fault source register: Stores the independent component values ​​corresponding to door gap abnormality source, closure abnormality source, door leaf offset source, sill jamming source and closure disturbance source; Node mapping register: Stores the fault source number, the fault state node number, and the independent component value corresponding to the fault state node; Node index pool: Stores the numbering correspondence between door gap fault status nodes, closure fault status nodes, offset fault status nodes, jamming fault status nodes, and disturbance fault status nodes; The set of fault state nodes is as follows: Read the independent component values ​​corresponding to the door gap anomaly source, the closure anomaly source, the door leaf offset source, the sill jamming source, and the closure disturbance source. Map the door gap anomaly source to a door gap fault state node, the closure anomaly source to a closure fault state node, the door leaf offset source to an offset fault state node, the sill jamming source to a jamming fault state node, and the closure disturbance source to a disturbance fault state node. Arrange them in the order of door gap fault state node, closure fault state node, offset fault state node, jamming fault state node, and disturbance fault state node to form a fault state node set. The multi-path coupling evolution unit reads the set of fault state nodes, constructs the gate structure hyperedge, the closure control hyperedge, and the gated communication relationship, generates a piecewise-gated hypergraph, and constructs mechanical clearance hysteresis channels, closure control hysteresis channels, and damped hysteresis channels to form a tensorized hysteresis coupling structure. It then executes multi-path fault propagation evolution to generate a multi-path fault propagation state sequence, where: The multi-path coupling evolution unit includes: Super-edge register: stores the node number combination corresponding to the super-edge of the gate structure and the node number combination corresponding to the super-edge of the closure control; Communication relationship register: Stores the gating communication relationships between faulty nodes; Hysteresis channel accumulator: stores the mechanical backlash hysteresis value sequence, the closure control hysteresis value sequence, and the damped hysteresis value sequence; Coupled tensor pool: Stores the mechanical backlash hysteresis, closure control hysteresis, and damped hysteresis values ​​in the tensor hysteresis coupling structure; The gate structure hyperedge, closure control hyperedge, and gate communication relationships are constructed as follows: Record the fault state node number of the door gap as 1, the fault state node number of the offset as 2, the fault state node number of the jamming as 3, the fault state node number of the closure as 4, and the fault state node number of the disturbance as 5. Record the node number combination 1, 2, 3 to obtain the door structure super-edge. Record the node number combination 4, 5, 1 to obtain the closure control super-edge. Read the independent component values ​​corresponding to each fault state node, calculate the average value of the absolute value of the difference between the independent component values ​​corresponding to any two fault state nodes, record the nodes with an average value less than 0.1 as gated communication relationships, and associate and map the gate structure hyperedge, closure control hyperedge and gated communication relationships to generate a segmented-gated hypergraph. Constructing a mechanical clearance hysteresis channel, a closed-loop control hysteresis channel, and a damped hysteresis channel, specifically as follows: Read the independent component values ​​corresponding to the door gap fault state node and the offset fault state node. Use the door gap fault state node as the starting node and the offset fault state node as the ending node to establish a mechanical clearance hysteresis channel. Read the independent component values ​​corresponding to the closure fault state node and the disturbance fault state node. Use the closure fault state node as the starting node and the disturbance fault state node as the ending node to establish a closure control hysteresis channel. Read the independent component values ​​corresponding to the jamming fault state node and the disturbance fault state node. Use the jamming fault state node as the starting node and the disturbance fault state node as the ending node to establish a damping hysteresis channel. Calculate the difference between the independent component values ​​corresponding to the starting node and the ending node of each channel according to the acquisition timestamp order, and write the difference as the hysteresis value of the corresponding channel into the corresponding channel. The tensor hysteresis coupling structure is composed of: The hysteresis values ​​in the mechanical clearance hysteresis channel, the hysteresis value in the closed control hysteresis channel, and the hysteresis value in the damping hysteresis channel are read in the order of the acquisition timestamps. The hysteresis values ​​in the mechanical clearance hysteresis channel, the hysteresis value in the closed control hysteresis channel, and the hysteresis value in the damping hysteresis channel are aligned and arranged according to the unified acquisition timestamps to form a three-dimensional hysteresis data set. The three-dimensional hysteresis data set is arranged in the order of the acquisition timestamps to form a tensor hysteresis coupling structure. Perform multi-path fault propagation evolution, specifically as follows: Read the independent component values ​​corresponding to each fault state node, as well as the hysteresis values ​​in the mechanical clearance hysteresis channel, the closure control hysteresis channel, and the damping hysteresis channel, according to the acquisition timestamp order. Add the independent component values ​​corresponding to the door gap fault state node to the hysteresis values ​​in the mechanical clearance hysteresis channel to obtain the offset fault state node update value. Add the independent component values ​​corresponding to the closure fault state node to the hysteresis values ​​in the closure control hysteresis channel to obtain the disturbance fault state node update value. Add the independent component values ​​corresponding to the jamming fault state node to the hysteresis values ​​in the damping hysteresis channel to obtain the jamming fault state node update value. Arrange all fault state node update values ​​according to the acquisition timestamp order to generate a multi-path fault propagation state sequence. The slow-varying memory modulation unit reads the multi-path fault propagation state sequence, adjusts the state evolution order based on the door closing speed characteristics and door jitter characteristics, and adjusts the propagation hysteresis based on the door gap width variation characteristics, door closing offset characteristics, and sill guide rail texture variation characteristics, thus constructing a variable-order-variable-hysteresis evolution structure, wherein: The slow-varying memory modulation unit includes: Evolution order register: stores the state evolution order; Propagation hysteresis register: stores the amount of propagation hysteresis; Transition Status Register: Stores the event trigger status node, status transition result, and corresponding acquisition time; Continuous state register: stores continuous evolution state nodes, continuous evolution results, and corresponding propagation hysteresis; Fusion Accumulation Pool: Stores a continuous fusion of fault propagation state sequences of events; The state evolution order is adjusted based on the characteristics of the gate closing speed and the gate shaking characteristics, specifically as follows: Read the gate closing speed characteristics and gate jitter characteristics under the corresponding acquisition timestamp of the multi-path fault propagation state sequence, calculate the average value of the gate closing speed characteristics and the gate jitter characteristics. When the average value is less than 0.3, set the state evolution order to 1. When the average value is greater than or equal to 0.3 and less than 0.6, set the state evolution order to 2. When the average value is greater than or equal to 0.6, set the state evolution order to 3. The propagation hysteresis is adjusted based on the characteristics of door gap width variation, door closure offset, and sill guide rail texture variation. Specifically: Read the characteristics of door gap width variation, door closure offset, and sill guide rail texture variation, calculate the average value of the three characteristics. When the average value is less than 0.3, the propagation hysteresis is set to 1 acquisition timestamp period. When the average value is greater than or equal to 0.3 and less than 0.6, the propagation hysteresis is set to 2 acquisition timestamp periods. When the average value is greater than or equal to 0.6, the propagation hysteresis is set to 3 acquisition timestamp periods. Constructing a variable-order-variable-hysteresis evolution structure, specifically: The state evolution order is associated with the independent component values ​​of the fault state nodes in the multi-path fault propagation state sequence, and the propagation hysteresis is associated with the collection timestamp order in the multi-path fault propagation state sequence, forming a variable-order-variable-hysteresis evolution structure. In a variable-order-variable-hysteresis evolution structure, event-triggered state nodes and continuous evolution state nodes are set, and the state transition results are fused with the continuous evolution results to generate an event-continuously fused fault propagation state sequence, wherein: The settings for event-triggered state nodes and continuous evolution state nodes are as follows: Read each fault state node in the variable-order-variable-hysteresis evolution structure. When the absolute value of the difference between the independent component values ​​of the fault state nodes corresponding to two adjacent acquisition timestamps is greater than 0.2, set the corresponding fault state node as an event-triggered state node. When the absolute value of the difference between the independent component values ​​of the fault state nodes corresponding to two adjacent acquisition timestamps is less than or equal to 0.2, set the corresponding fault state node as a continuous evolution state node. The state transition result refers to the result obtained by multiplying the independent component value of the fault state node corresponding to the event-triggered state node by the state evolution order; The continuous evolution result refers to the average of the independent component values ​​of the fault state node corresponding to the acquisition timestamps of the forward propagation hysteresis of the continuous evolution state node. Generate a continuous fusion of event-driven fault propagation state sequences, specifically: Read the independent component value and state evolution order of the fault state node corresponding to the event triggering state node. Multiply the independent component value of the fault state node with the state evolution order to obtain the state transition result. Read the independent component value and propagation hysteresis of the fault state node corresponding to the continuously evolving state node. Calculate the average value of the independent component value of the fault state node corresponding to the propagation hysteresis of the previous acquisition timestamps to obtain the continuous evolution result. Arrange the state transition result and the continuous evolution result according to the acquisition timestamp order to generate the event continuous fusion fault propagation state sequence. The heterocentric coherent orbit unit reads the continuously fused fault propagation state sequence of events, treats each fault state node as a game participant node, uses the excitation state of each fault state node as the game strategy, and uses the coupling strength between fault state nodes as the game payoff. It then performs multi-objective game modulation on the continuously fused fault propagation state sequence of events to generate a modulated fault propagation state sequence, constructs a heterocentric coherent fault state orbit, and outputs the gate fault evolution state sequence, where: The heteroclinic orbital unit includes: Node status register: stores the independent component values ​​corresponding to each faulty node; Coupling strength table: Stores the coupling strength between nodes in a fault state; Profit Register: Stores the game payout corresponding to each faulty node; Policy Register: Stores policies for increasing the independent component value of a fault state node and policies for decreasing the independent component value of a fault state node; Track pool: Stores the modulated fault propagation state sequence and the track with different-homogeneous fault states; The game-playing nodes refer to the door gap fault state node, closure fault state node, offset fault state node, jamming fault state node, and disturbance fault state node. The activation state of each fault state node refers to the independent component value of the corresponding fault state node at the current acquisition timestamp. Game strategy refers to the adjustment method of increasing or decreasing the independent component value of the corresponding fault state node. The coupling strength between faulty state nodes is the result obtained by dividing 1 by the sum of the absolute value of the difference between the independent component values ​​corresponding to the two faulty state nodes and 0.001; The game payoff refers to the sum of all coupling strengths between the current faulty node and its connected faulty nodes. Perform multi-objective game modulation on the continuously fused fault propagation state sequence of the event, specifically: Read the independent component values ​​of the fault state nodes corresponding to each participating node in the game, calculate the absolute value of the difference between the current fault state node's independent component value and the independent component values ​​of the connected fault state nodes, divide 1 by the sum of the absolute value of the difference and 0.001 to obtain the corresponding coupling strength, accumulate all the coupling strengths corresponding to the current fault state node to obtain the game payoff corresponding to the current fault state node, when the game payoff corresponding to increasing the independent component value of the fault state node is greater than the game payoff corresponding to decreasing the independent component value of the fault state node, increase the corresponding independent component value of the fault state node by 0.01, when the game payoff corresponding to increasing the independent component value of the fault state node is less than or equal to the game payoff corresponding to decreasing the independent component value of the fault state node, decrease the corresponding independent component value of the fault state node by 0.01, and obtain the modulated fault propagation state sequence; Constructing a heterogeneous homophonic fault state orbit is specifically as follows: Read the independent component values ​​of the fault state nodes in the modulated fault propagation state sequence according to the acquisition timestamp order, establish a connection between the independent component values ​​of the fault state nodes corresponding to adjacent acquisition timestamps, arrange all the connection relationships in time order, and form a fault state track with different destinations and the same modulation. The output gate fault evolution state sequence is as follows: Read the independent component values ​​of the fault state nodes in the heterogeneous and homogeneous fault state track according to the acquisition timestamp order, arrange all the independent component values ​​of the fault state nodes in time order, and output the gate body fault evolution state sequence. An improved Hindmarsh Rose neural network was trained, using fault state prediction error and fault trajectory reconstruction error as joint optimization objectives. The parameters in the fault state mapping unit, multi-path coupling evolution unit, slowly varying memory modulation unit, and heterocentric coherent trajectory unit were continuously optimized. Training stopped when the change in the joint optimization objective was less than 0.001 for 10 consecutive iterations. The improved Hindmarsh Rose neural network was trained as follows: The historical gate independent fault source matrix and the corresponding gate fault evolution state sequence are read as training samples. The gate independent fault source matrix is ​​input into the fault state mapping unit, the multi-path coupling evolution unit, the slow-varying memory modulation unit, and the heterodox and co-homogeneous orbit unit to obtain the predicted fault state node independent component values ​​and the predicted heterodox and co-homogeneous fault state orbit. The square of the difference between the predicted independent component values ​​of fault state nodes and the actual independent component values ​​of fault state nodes is calculated, and the average of the squared differences corresponding to all collection timestamps is calculated to obtain the fault state prediction error. The square of the difference between the independent component values ​​of each track node in the predicted heterodox and homodox fault state track and the corresponding independent component values ​​of the track node in the actual heterodox and homodox fault state track is calculated, and the average of the squared differences corresponding to all track nodes is calculated to obtain the fault track reconstruction error. The fault state prediction error and the fault track reconstruction error are weighted and summed according to weights of 0.6 and 0.4 to obtain the joint optimization objective. The influence of the joint optimization objective on each parameter in the fault state mapping unit, multi-path coupling evolution unit, slow-varying memory modulation unit, and hetero-homogeneous orbit unit is calculated. Each parameter is adjusted according to a learning rate of 0.001. After one parameter adjustment, the fault state prediction error, fault orbit reconstruction error, and joint optimization objective are recalculated. Training is stopped when the change in the joint optimization objective corresponding to 10 consecutive iterations is less than 0.001, and the converged improved HindmarshRose neural network is obtained.

[0026] In this embodiment, generating elevator door fault detection results and elevator door fault early warning results includes: The door failure evolution state sequence is read according to the collected timestamp data. The state transition relationships between the normal closing state, door leaf offset state, door gap abnormal state, closure abnormal state, and closure failure state are extracted, and the elevator door failure evolution path is generated. Extract the state transition relationships between the normal closing state, door leaf offset state, door gap abnormal state, closure abnormal state, and closure failure state, specifically as follows: Read the independent component values ​​of the fault state nodes in the fault evolution state sequence of the gate body according to the collection timestamp order, determine the state corresponding to the fault source with the largest independent component value as the current state, record the change relationship between adjacent states according to the collection timestamp order, and obtain the state transition relationship. The elevator door fault evolution path is generated as follows: Read the state transition relationships in the order of the collection timestamps, connect all the state transition relationships end to end, and generate the elevator door failure evolution path; Elevator door fault types are identified based on the termination state of the elevator door fault evolution path, state transition characteristics, and fault contribution relationships in the independent fault source matrix of the door body. The termination state of the elevator door failure evolution path refers to the state corresponding to the last collected timestamp in the elevator door failure evolution path. The fault contribution relationship in the independent fault source matrix of the door body refers to the magnitude relationship of the independent component values ​​of the door gap abnormal source, closure abnormal source, door leaf offset source, sill jamming source and closure disturbance source under the same acquisition timestamp. The identification of elevator door malfunction types is as follows: Read the termination state and fault contribution relationship of the elevator door fault evolution path, determine the fault source with the largest independent component value corresponding to the termination state as the fault type, and obtain the elevator door fault type. Read the fault source information and status information corresponding to the elevator door fault type, and write the elevator door fault type, fault source information, status information, fault location, and data collection timestamp into the elevator door fault detection results, including: The fault source information and status information corresponding to the elevator door fault type refer to the fault source name, the independent component value corresponding to the fault source, the fault status name, and the corresponding collection timestamp. Read the fault states that have not reached the closed failure state in the elevator door fault evolution path, and combine the changing trends and state transition directions of the corresponding fault source information to generate elevator door fault early warning results, where: The fault state that has not reached the closure failure state in the elevator door fault evolution path refers to the fault state other than the closure failure state in the elevator door fault evolution path. The changing trend and state transition direction of the corresponding fault source information refer to the increase or decrease of the independent component values ​​of the fault source corresponding to adjacent acquisition timestamps and the transition of the current state to the next state. Generate elevator door malfunction warning results, specifically: Read the independent component value of the fault source corresponding to the fault state that has not reached the closed failure state; calculate the change in the independent component value of the fault source corresponding to adjacent acquisition timestamps; when the independent component value of the fault source corresponding to 3 consecutive acquisition timestamps increases and the state transition direction points to the closed failure state, generate a fault warning sign; arrange the fault state name, fault source name, independent component value of the fault source, state transition direction, fault warning sign and acquisition timestamp data to generate elevator door fault warning result.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to a simulated environment for monitoring the operation of an elevator door in a building. This environment included one split-type elevator door, one set of door area image acquisition devices, and one edge computing terminal. The image acquisition devices collected elevator door opening and closing images, door gap area images, door leaf edge area images, sill guide rail area images, and acquisition timestamp data at 25 frames per second. A total of 7200 sets of opening and closing image sequences were collected in the simulation, including 4800 training samples, 1200 verification samples, and 1200 test samples. The training samples included 1800 normal opening and closing samples, 900 abnormal door gap samples, 760 door leaf offset samples, 720 sill jamming samples, and 620 closure disturbance samples, used to train the improved Hindmarsh Rose neural network and verify the fault source separation effect.

[0028] During a door closing test, the average grayscale value of the door gap area in frame 28 was 0.57, and the average gradient value of the door gap edge was 0.38; the average grayscale value of the door leaf edge was 0.63, and the average gradient value of the door leaf edge was 0.44; the average grayscale value of the sill rail was 0.51, and the average texture response value of the sill rail was 0.31. The standard brightness value was set to 0.8, the standard edge gradient value to 0.6, and the standard texture response value to 0.5. The door gap brightness loss was 0.23, and the door gap edge loss was 0.22. Based on a brightness weight of 0.4 and an edge weight of 0.6, the door gap enhancement was calculated to be 0.224, with an enhancement coefficient of 1.224. The door leaf edge enhancement was 0.164, with an enhancement coefficient of 1.164. The sill rail enhancement was 0.240, with an enhancement coefficient of 1.240. After enhancement, the average grayscale value of the door gap increased from 0.57 to 0.70, the average gradient value of the door gap edge increased from 0.38 to 0.47, and the average texture response value of the sill rail increased from 0.31 to 0.38. The door gap edge and rail texture became more continuous in the image. The enhanced image frames were written into the elevator door enhanced image sequence.

[0029] From frames 32 to 36, the system reads the left and right edge positions of the door gap from the enhanced area image. The door gap widths in frames 32 to 36 are 20 pixels, 19 pixels, 18 pixels, 18 pixels, and 17 pixels, respectively, with width variation features of 1, 1, 0, and 1. In frame 36, the horizontal pixel coordinate of the door gap center is 627.5, the elevator door closing center position is 625.0, and the door closing offset feature is 2.5 pixels. The horizontal pixel coordinates of the door leaf edge in frames 32 to 36 are 611, 613, 616, 618, and 619, respectively, with a door jitter feature of 8 pixels. The texture response values ​​of the sill rail in frames 32 to 36 are 0.48, 0.53, 0.58, 0.62, and 0.64, respectively, with texture variation features of sill rail in frames 0.05, 0.05, 0.04, and 0.02, respectively. Since the time interval between adjacent frames is 0.04 seconds, the change in the door gap width corresponding to the 36th frame is 1 pixel, and the door closing speed characteristic is 25 pixels per second, which is significantly lower than the average of 73 pixels per second during the normal door closing phase.

[0030] The system writes the door gap width variation features, door edge displacement features, door closing speed features, and door jitter features into the high-frequency motion channel, and the door closing offset features and sill guide rail texture variation features into the low-frequency structure channel. After centering and block diagonal whitening, the whitened feature group of frame 36 is as follows: door gap width variation feature 0.22, door edge displacement feature 0.41, door closing speed feature -0.76, door jitter feature 0.68, door closing offset feature -0.13, and sill guide rail texture variation feature -0.09. The FastICA algorithm performs independent component separation on this whitened feature group, obtaining independent component values ​​for door gap anomaly source 0.42, closure anomaly source 0.38, door offset source 0.51, sill jamming source 0.79, and closure disturbance source 0.57. The correlation statistics between the ground sill jamming source and the texture change characteristics of the ground sill guide rail are 0.71, which is greater than 0.3, and are retained as connecting edges in the topological correlation graph. The component difference value between the ground sill jamming source and the closed disturbance source is adjusted from 0.08 to 0.14 by the adversarial discrimination mechanism, so that the ground sill jamming source and the closed disturbance source form distinguishable fault source information in the independent fault source matrix of the gate body.

[0031] After improving the Hindmarsh Rose neural network to read the independent fault source matrix of the door body, the sill jamming source is mapped to a jamming fault state node, the closure disturbance source is mapped to a disturbance fault state node, and the door offset source is mapped to an offset fault state node. In frame 36, the independent component value of the jamming fault state node is 0.79, the independent component value of the disturbance fault state node is 0.57, and the independent component value of the offset fault state node is 0.51. The hysteresis value of the mechanical clearance hysteresis channel is 0.42 for the door gap anomaly source minus 0.51 for the door offset source, resulting in -0.09. The hysteresis value of the closure control hysteresis channel is 0.38 for the closure anomaly source minus 0.57 for the closure disturbance source, resulting in -0.19. The hysteresis value of the damping hysteresis channel is 0.79 for the sill jamming source minus 0.57 for the closure disturbance source, resulting in 0.22. The system adds the independent component value of the jamming fault state node (0.79) to the damped hysteresis channel value (0.22) to obtain the updated value of the jamming fault state node (1.01), and generates a multi-path fault propagation state sequence.

[0032] The slow-varying memory modulation unit reads the normalized value of the door closing speed feature (0.26) and the normalized value of the door jitter feature (0.63) from frame 36. The average of these two values ​​is 0.445, therefore the state evolution order is set to 2. It also reads the normalized value of the door gap width change feature (0.20), the normalized value of the door closing offset feature (0.25), and the texture change feature of the sill rail (0.02), with an average of 0.157. Therefore, the propagation hysteresis is set to one acquisition timestamp period. The absolute difference between the independent component value of the jamming fault state node in frame 36 and that in frame 35 is 0.24, which is greater than 0.2. Therefore, it is set as an event-triggered state node, and the state transition result is 1.01 multiplied by 2, resulting in 2.02. The absolute difference between adjacent frames of the closing fault state node is 0.07, so it is set as a continuous evolution state node. The continuous evolution result is the independent component value corresponding to the previous acquisition timestamp (0.35). The system arranges the state transition results and continuous evolution results according to the acquisition timestamp, and generates a continuous fusion fault propagation state sequence.

[0033] The heterodox and co-homogeneous orbital unit performs game-theoretic modulation on the continuous fusion of the fault propagation state sequence. In frame 36, the independent component value of the jamming fault state node is 2.02, and the independent component value of the disturbance fault state node is 0.19. The absolute difference between the two is 1.83, and the coupling strength is 1 divided by 1.831, resulting in 0.546. The absolute difference between the jamming fault state node and the offset fault state node is 1.69, and the coupling strength is 0.591. The current game payoff for the jamming fault state node is 1.137. If the independent component value of the jamming fault state node is increased by 0.01, the game payoff becomes 1.131; if it is decreased by 0.01, the game payoff becomes 1.144. Therefore, the system modulates the independent component value of this node to 2.01. After modulation, the fault propagation state sequence forms a heterodox and co-homogeneous fault state orbital between the normal closure state, door offset state, door gap abnormal state, closure abnormal state, and closure failure approaching state. A ground sill jamming warning result is output 5.6 seconds before the closure failure state occurs.

[0034] In 1200 test samples, the traditional method correctly identified 1035 samples, with an overall accuracy of 86.25%, while the method of this invention correctly identified 1159 samples, with an overall accuracy of 96.58%. In 150 samples of ground sill blockage, the traditional method correctly identified 119 samples, with an accuracy of 79.33%, while the method of this invention correctly identified 142 samples, with an accuracy of 94.67%. In 200 samples of closed disturbance, the traditional method correctly identified 164 samples, with an accuracy of 82.00%, while the method of this invention correctly identified 191 samples, with an accuracy of 95.50%. In 180 samples of complex faults, the traditional method correctly identified 134 samples, with an accuracy of 74.44%, while the method of this invention correctly identified 169 samples, with an accuracy of 93.89%. The traditional method has an average warning lead time of 2.1 seconds, while the method of this invention has an average warning lead time of 6.4 seconds. The traditional method has an average false alarm rate of 7.8%, while the method of this invention has an average false alarm rate of 2.1%. The traditional method has an average missed alarm rate of 8.6%, while the method of this invention has an average missed alarm rate of 2.7%. The simulation results show that this invention can separate independent fault sources under conditions of door gap abnormalities, door leaf misalignment, sill jamming, and closure disturbances that are coupled together, and output fault warning results in advance based on the fault evolution state sequence.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting elevator door faults based on image enhancement, characterized in that, include: Collect operational status image data of the elevator door during operation, perform preprocessing on the operational status image data, and generate a standard image sequence of the elevator door; Image enhancement processing is performed on the standard image sequence of the elevator door to enhance the structural features of the door gap area, door lock area, door leaf edge area and sill guide rail area, and generate an enhanced image sequence of the elevator door. Based on the enhanced image sequence of elevator doors, the features of door gap width change, door closure offset, door leaf edge displacement, sill guide rail texture change, door closure speed and door shaking are extracted to construct a hybrid visual state matrix of the door. The hybrid visual state matrix of the door is centered and whitened, and the FastICA algorithm is used to separate the independent components of the processed hybrid visual state matrix of the door to generate the independent fault source matrix of the door. An improved HindmarshRose neural network is constructed based on the independent fault source matrix of the door body. The state evolution of the multi-path fault propagation relationship between the abnormal sources of door gaps, abnormal sources of closure, offset sources of door leaves, stuck sources of ground sills and closure disturbance sources is carried out, and the heterogeneous and homogeneous fault state trajectory is constructed to generate the door body fault evolution state sequence. The elevator door fault evolution path is determined based on the door fault evolution state sequence, and the elevator door fault type is identified. By combining the door independent fault source matrix, the door fault evolution state sequence, and the elevator door fault evolution path, elevator door fault detection results and elevator door fault early warning results are generated.

2. The elevator door fault detection method based on image enhancement according to claim 1, characterized in that, The operational status image data specifically includes image data of the door opening process, image data of the door closing process, image data of the door gap area, image data of the door leaf edge area, image data of the sill and guide rail area, and data of the acquisition timestamp.

3. The elevator door fault detection method based on image enhancement according to claim 1, characterized in that, The preprocessing of the running state image data specifically includes image distortion correction, inter-frame registration, image denoising, brightness normalization, and time alignment.

4. The elevator door fault detection method based on image enhancement according to claim 1, characterized in that, The generated elevator door enhanced image sequence includes: Based on the collected timestamp data, the images of the door gap area, door leaf edge area, and sill guide rail area in the standard image sequence of the elevator door are read, and the mean gray value of the door gap, the mean gradient value of the door gap edge, the mean gray value of the door leaf edge, the mean gradient value of the door leaf edge, the mean gray value of the sill guide rail, and the mean texture response value of the sill guide rail are calculated respectively. The door gap enhancement coefficient is calculated based on the brightness loss corresponding to the average gray value of the door gap and the edge loss corresponding to the average gradient of the door gap edge. The door leaf edge enhancement coefficient is calculated based on the brightness loss corresponding to the average gray value of the door leaf edge and the edge loss corresponding to the average gradient of the door leaf edge. The sill rail enhancement coefficient is calculated based on the brightness loss corresponding to the average gray value of the sill rail and the texture loss corresponding to the average texture response of the sill rail. Enhancement processing is performed on the door gap region image, door leaf edge region image, and sill guide rail region image according to the door gap enhancement coefficient, door leaf edge enhancement coefficient, and sill guide rail enhancement coefficient, respectively, to generate door gap enhanced region image, door leaf edge enhanced region image, and sill guide rail enhanced region image, and then written into the elevator door standard image sequence to generate elevator door enhanced image sequence.

5. The elevator door fault detection method based on image enhancement according to claim 1, characterized in that, The construction of the gate hybrid visual state matrix includes: Read the elevator door enhanced image sequence according to the collected timestamp data, extract the left and right edge positions of the door gap from the door gap enhanced area image, extract the door leaf edge position from the door leaf edge enhanced area image, and extract the sill guide rail texture response value from the sill guide rail enhanced area image; The door gap width is calculated based on the left and right edge positions of the door gap under the same collection timestamp. The door gap width change feature is extracted based on the difference in door gap width corresponding to adjacent collection timestamps. The door closure offset feature is extracted based on the position offset between the center position of the door gap and the closing center position of the elevator door. The door leaf edge displacement feature is extracted based on the difference in door leaf edge positions corresponding to adjacent collection timestamps. The textural change features of the sill guide rail are extracted based on the difference in the texture response values ​​of the sill guide rail corresponding to adjacent acquisition timestamps. The door closing speed features are extracted based on the door gap width change and time interval corresponding to adjacent acquisition timestamps. The door shaking features are extracted based on the door edge position fluctuation corresponding to consecutive acquisition timestamps. Based on the collected timestamp data, the characteristics of door gap width change, door closure offset, door edge displacement, sill guide rail texture change, door closure speed, and door shaking are written into the same state row, and all state rows are arranged in the order of collection time to construct a hybrid visual state matrix of the door.

6. The elevator door fault detection method based on image enhancement according to claim 1, characterized in that, The generated gate-independent fault source matrix includes: The state features in the hybrid visual state matrix of the door are divided into channels. The door gap width change feature, door leaf edge displacement feature, door closing speed feature and door shaking feature are divided into high frequency motion channels, and the door closing offset feature and sill guide rail texture change feature are divided into low frequency structure channels. The state features in the high-frequency motion channel and the low-frequency structure channel are centered respectively to eliminate the overall offset between different state features in the high-frequency motion channel and the low-frequency structure channel, thus obtaining the high-frequency centered state submatrix and the low-frequency centered state submatrix. A streaming Sketch-Cov statistical structure is constructed based on the high-frequency centered state submatrix and the low-frequency centered state submatrix, and the second-order statistical information of the high-frequency motion channel and the low-frequency structure channel is compressed and stored. High-frequency block diagonal whitening matrix and low-frequency block diagonal whitening matrix are constructed respectively. The high-frequency centered state submatrix is ​​whitened based on the high-frequency block diagonal whitening matrix, and the low-frequency centered state submatrix is ​​whitened based on the low-frequency block diagonal whitening matrix. The whitened high-frequency motion channel and the whitened low-frequency structure channel are then aligned and spliced ​​according to the acquisition timestamp data to obtain the whitened state matrix. Based on the FastICA algorithm, independent component separation is performed on the whitening state matrix to construct a topological correlation graph between the door gap width variation feature, door closure offset feature, door leaf edge displacement feature, sill guide rail texture variation feature, door closure speed feature, and door shaking feature. Sparse constraints are applied to the independent component separation process to generate a feature-related independent component set. Based on the feature-related independent component set, independence objective, sparsity objective and fault discrimination objective are established. The independent separation direction is updated through multi-objective collaborative optimization. An adversarial discrimination mechanism is introduced to reduce the correlation between different independent components, and an independent component matrix is ​​obtained. Calculate the correspondence between each independent component in the independent component matrix and the characteristics of door gap width variation, door closure offset, door leaf edge displacement, sill guide rail texture variation, door closure speed, and door shaking. Label each independent component as a door gap anomaly source, closure anomaly source, door leaf offset source, sill jamming source, and closure disturbance source, and construct an independent fault source matrix for the door.

7. The elevator door fault detection method based on image enhancement according to claim 1, characterized in that, The generated gate failure evolution state sequence includes: An improved Hindmarsh Rose neural network is constructed, comprising a fault state mapping unit, a multipath coupling evolution unit, a slowly varying memory modulation unit, and a heteroclinic orbit unit. The fault state mapping unit reads the independent fault source matrix of the door body, maps the abnormal sources of door gaps, abnormal sources of closure, sources of door leaf offset, sources of ground sill jamming, and sources of closure disturbance to the corresponding fault state nodes, and forms a set of fault state nodes. The multi-path coupling evolution unit reads the set of fault state nodes, constructs the gate structure hyperedge, the closure control hyperedge and the gate control communication relationship, generates the piecewise-gated hypergraph, and constructs the mechanical clearance hysteresis channel, the closure control hysteresis channel and the damping hysteresis channel to form a tensor hysteresis coupling structure, executes multi-path fault propagation evolution, and generates a multi-path fault propagation state sequence. The slow-variable memory modulation unit reads the multi-path fault propagation state sequence, adjusts the state evolution order according to the door closing speed characteristics and door jitter characteristics, and adjusts the propagation hysteresis according to the door gap width change characteristics, door closing offset characteristics and sill guide rail texture change characteristics, thus constructing a variable-order-variable-hysteresis evolution structure. In the variable-order-variable-hysteresis evolution structure, event-triggered state nodes and continuous evolution state nodes are set, and the state transition results are fused with the continuous evolution results to generate an event-continuous fused fault propagation state sequence. The heterodox and coherent orbit unit reads the event continuous fusion fault propagation state sequence, takes each fault state node as a game participant node, takes the excitation state of each fault state node as the game strategy, takes the coupling strength between fault state nodes as the game payoff, performs multi-objective game modulation on the event continuous fusion fault propagation state sequence, generates the modulated fault propagation state sequence, constructs the heterodox and coherent fault state orbit, and outputs the gate fault evolution state sequence. The improved Hindmarsh Rose neural network was trained, and the fault state prediction error and fault trajectory reconstruction error were used as joint optimization objectives. The parameters in the fault state mapping unit, multi-path coupling evolution unit, slow-varying memory modulation unit and hetero-homogeneous trajectory unit were continuously optimized. Training was stopped when the change of the joint optimization objective was less than 0.001 for 10 consecutive iterations.

8. The elevator door fault detection method based on image enhancement according to claim 1, characterized in that, The generation of elevator door fault detection results and elevator door fault early warning results includes: Read the door failure evolution state sequence according to the collected timestamp data, extract the state transition relationship between normal closing state, door leaf offset state, door gap abnormal state, closing abnormal state and closing failure state, and generate the elevator door failure evolution path. Elevator door fault types are identified based on the termination state of the elevator door fault evolution path, state transition characteristics, and fault contribution relationships in the independent fault source matrix of the door body. Read the fault source information and status information corresponding to the elevator door fault type, and write the elevator door fault type, fault source information, status information, fault location and collection timestamp data into the elevator door fault detection result; Read the fault states that have not reached the closed failure state in the elevator door fault evolution path, and generate elevator door fault early warning results by combining the changing trend and state transition direction of the corresponding fault source information.

Images