A ferris wheel cabin safety detection and abnormality identification method based on machine vision

Abstract

The application discloses a kind of based on machine vision's Ferris wheel hanging car safety detection and abnormality identification method, including the following steps: take boarding platform and circumferential position image, determine hanging car position in combination with encoder and sensor;Image and structural features are extracted to component state data by door body, door lock, door gap, pressure rod, fastener and suspension arm area correction;Characteristic representation is generated by state data and rotation position index unfolding splicing;A set of position conditioned Koopman operators is trained based on the representation, block sparse is used and growth constraint is set in crack sub-block;Residual, spectral shift and logic violation result are obtained by using operator set for cross-ring prediction and cyclic consistency modeling;Composite anomaly score is generated based on the result and grade is determined;Grade is output and maintenance suggestion is generated.The application introduces a set of position conditioned Koopman operators, realizes cross-ring prediction and cyclic consistency modeling of Ferris wheel hanging car.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent detection, in particular to a Ferris wheel cabin safety detection and abnormality identification method based on machine vision. BACKGROUND

[0002] The existing Ferris wheel cabin safety detection mainly relies on manual inspection and single-point sensor monitoring. Manual inspection is usually carried out when the device is shut down, and the detection period is long, so it is difficult to find abnormalities in the running process in time. Although the single-point sensor can monitor the state of the door lock or pressure rod, the coverage is limited, and it cannot comprehensively detect multiple key components such as the door body, door gap, fastener and suspension arm, and it is prone to false positives and false negatives during long-term operation.

[0003] In recent years, machine vision and deep learning technologies have been gradually applied to the field of equipment safety detection, but most existing methods use conventional image recognition or classification models, which lack the ability to model the periodic rotation scene of the Ferris wheel, and cannot effectively handle cross-circle alignment, periodic consistency and slow evolution anomalies, resulting in insufficient accuracy in detecting gradual risks such as crack propagation, gradually widening door gap and loose fastener.

[0004] Therefore, how to provide a Ferris wheel cabin safety detection and abnormality identification method based on machine vision is a problem that needs to be solved by those skilled in the art. SUMMARY

[0005] One object of the present application is to provide a Ferris wheel cabin safety detection and abnormality identification method based on machine vision. The present application models the state of the cabin components of the Ferris wheel during periodic rotation by introducing a set of position-conditioned Koopman operators, combines cross-circle same-position prediction and whole-circle cycle consistency modeling, and can capture gradual abnormal changes. In the operator set, a block sparse structure is used to logically couple the door lock and door gap, the pressure rod and the human body posture, and the fastener and the suspension arm, thereby reducing the interference of unrelated components. In the crack detection corresponding operator sub-block, growth constraints are set, and a composite anomaly score is generated by combining the prediction residual, the spectral shift result and the logical constraint violation result, thereby realizing abnormality level determination, so as to provide higher precision and stronger targeted detection and early warning capability for cabin operation safety.

[0006] According to the Ferris wheel cabin safety detection and abnormality identification method based on machine vision of the present application, the following steps are included:

[0007] Obtain images installed on the Ferris wheel boarding platform and the circumferential position, and determine the position of the cabin in the rotation process in combination with the rotation angle signal;

[0008] The areas of the gondola door, door lock, door gap, pressure bar, fasteners and suspension arm are corrected, and image features and structural features are extracted to form component status data;

[0009] The component state data is subjected to feature transformation to generate a feature representation carrying rotational position information;

[0010] Based on the feature representation, a set of position-conditional Koopman operators is established, and a block-sparse structure is adopted to reflect the logical coupling relationship between different components. Growth constraints are set in the operator sub-blocks corresponding to crack detection.

[0011] The Koopman operator set is used to perform cross-cycle same-position prediction and full-cycle cyclic consistency modeling to obtain prediction residuals, spectral shift results and logical constraint violation results.

[0012] Based on the predicted residuals, spectral shift results, and logical constraint violation results, a composite anomaly score is generated. The anomaly level of the gondola component is determined according to the composite anomaly score, and the anomaly level and maintenance recommendations are output.

[0013] Furthermore, the rotational position is obtained by an angle encoder installed on the Ferris wheel drive unit and calibrated in conjunction with a zero-position sensor at the boarding platform. The rotational position is synchronized with the image acquisition timestamp.

[0014] Furthermore, the steps for generating component status data include:

[0015] Perspective transformation is performed based on the camera's intrinsic and extrinsic parameters to correct the areas of the gondola door, door lock, door gap, pressure bar, fasteners, and suspension arm.

[0016] Within the corrected area, image features are extracted through edge detection, semantic segmentation, and contour recognition, including the door contour, door seam edge, fastener contour, and cantilever area.

[0017] Structural features are extracted through key point detection and geometric calculation, including door gap width, door lock lever angle, and the relative distance between the lever position and key points on the human body.

[0018] The image features and structural features are combined to form component status data.

[0019] Furthermore, the feature transformation includes:

[0020] When performing feature transformation on the component state data, the value of the rotation position is concatenated with the component state data;

[0021] The rotational position is periodically expanded using sine and cosine functions;

[0022] The unfolding results are combined with the component state data to generate a high-dimensional feature representation carrying rotational position information.

[0023] Furthermore, the construction of the conditionalized set of Koopman operators includes:

[0024] Based on the rotation position signal output by the angle encoder, the rotation process of one revolution of the Ferris wheel is divided into several position intervals, each interval corresponding to a specific cabin orientation;

[0025] The high-dimensional feature representation carrying rotational position information is input into the training module, and the corresponding Koopman operator unit is trained in each interval to characterize the change pattern of component state data within the interval.

[0026] Divide the intervals according to the rotation position and train the corresponding Koopman operator units in each interval. Combine all the operator units to form an operator set.

[0027] During training, block sparsity constraints are applied to the operator set so that the door lock and door gap, pressure bar and human posture, fastener and suspension arm form mutually coupled operator sub-blocks, while the remaining components remain sparse.

[0028] Growth constraints are set in the operator sub-block corresponding to crack detection to characterize the trend of crack or corrosion gradually expanding over time.

[0029] Furthermore, the operations performed by the set of Koopman operators include:

[0030] After the gondola completes one revolution, the component status data at the same position in this revolution is input into the Koopman operator unit of the corresponding position interval to obtain the prediction data for the same position in the next revolution.

[0031] For each position, the actual data of the next rotation is compared with the predicted data obtained in the previous step, and the prediction residual sequence is calculated and recorded.

[0032] At each position, spectral feature descriptions are extracted based on the corresponding Koopman operator unit and input data. The spectral features of the current loop and the previous loop at the same position are obtained and compared. The spectral shift results are calculated and recorded.

[0033] Establish a list of logical relationships between door locks and door gaps, pressure bars and human posture, and fasteners and suspension arms;

[0034] At each location, the component status data is checked item by item, the triggered violation items are recorded, and logical constraint violation results are generated.

[0035] The prediction residuals, spectral shift results, and logical constraint violation results are summarized by location index and gondola number.

[0036] Furthermore, the generation of the anomaly level result includes:

[0037] Input the obtained predicted residual sequence, spectral shift results, and logical constraint violation results into the anomaly score generation module;

[0038] In the anomaly score generation module, the prediction residual, spectral shift result and logical constraint violation result are linearly combined according to preset weights to obtain the composite anomaly score.

[0039] After the composite anomaly score is generated, it is compared item by item with the stored threshold table;

[0040] Based on the comparison results, the composite anomaly scores are divided into different levels, corresponding to three intervals: safe, warning, and dangerous, to form anomaly level results;

[0041] Output the anomaly level result and bind the result with the gondola number and location index.

[0042] Furthermore, the generation of the maintenance recommendations includes:

[0043] Store multiple maintenance rules, each rule corresponding to a risk level and a specific component category;

[0044] After generating the risk classification results, the maintenance rules that match the risk level and component category are invoked.

[0045] Based on the invoked maintenance rules, maintenance suggestions are generated, including the maintenance components, maintenance types, and priorities, and then stored in association with the risk classification results.

[0046] The risk classification results and maintenance suggestions are simultaneously output to the host computer interface, control system, or work order management module as reference information for maintenance personnel.

[0047] The beneficial effects of this invention are:

[0048] This invention introduces a set of position-conditional Koopman operators to model the state of gondolas at different positions during the periodic rotation of a Ferris wheel. This enables cross-cycle same-position prediction and consistent modeling of the entire cycle, and can capture the gradual changes that gondolas undergo during long-term operation, thereby improving the sensitivity and accuracy of anomaly detection.

[0049] This invention employs a block-sparse structure in the operator set, dividing logically related components such as door locks and door gaps, pressure bars and human postures, fasteners and suspension arms into blocks for modeling. This maintains logical coupling between components, avoids redundant connections between unrelated components, makes the detection results more consistent with actual operating rules, and reduces the false alarm rate.

[0050] This invention sets growth constraints in the operator sub-block corresponding to crack detection, enabling the operator to characterize the trend of crack, corrosion and spalling gradually expanding over time. It also combines the prediction residual, spectral shift results and logic constraint violation results to generate a composite anomaly score, establishing an anomaly level determination mechanism, thereby providing a reliable basis for risk classification and maintenance decisions. Attached Figure Description

[0051] 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:

[0052] Fig. 1 This is a flowchart of a Ferris wheel gondola safety detection and anomaly identification method based on machine vision proposed in this invention;

[0053] Fig. 2 This invention provides a link graph for generating anomaly scores in a Ferris wheel gondola safety detection and anomaly identification method based on machine vision.

[0054] Fig. 3 This is a schematic diagram of the block-sparse structure of a Ferris wheel gondola safety detection and anomaly identification method based on machine vision proposed in this invention. Detailed Implementation

[0055] 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.

[0056] refer to Figs. 1-3 A method for safety detection and anomaly identification of Ferris wheel gondolas based on machine vision includes the following steps:

[0057] Acquire images installed on the Ferris wheel boarding platform and at the circumference, and combine them with rotation angle signals to determine the position of the gondola during rotation;

[0058] The areas of the gondola door, door lock, door gap, pressure bar, fasteners and suspension arm are corrected, and image features and structural features are extracted to form component status data;

[0059] The component state data is subjected to feature transformation to generate a feature representation carrying rotational position information;

[0060] Based on the feature representation, a set of position-conditional Koopman operators is established, and a block-sparse structure is adopted to reflect the logical coupling relationship between different components. Growth constraints are set in the operator sub-blocks corresponding to crack detection.

[0061] The Koopman operator set is used to perform cross-cycle same-position prediction and full-cycle cyclic consistency modeling to obtain prediction residuals, spectral shift results and logical constraint violation results.

[0062] Based on the predicted residuals, spectral shift results, and logical constraint violation results, a composite anomaly score is generated. The anomaly level of the gondola component is determined according to the composite anomaly score, and the anomaly level and maintenance recommendations are output.

[0063] In this embodiment, images of the Ferris wheel gondolas are acquired and their positions during rotation are determined through the collaborative work of sensors and cameras. Specifically, a photoelectric incremental angle encoder is installed on the Ferris wheel drive unit to output pulse signals that change with rotation; an infrared through-beam zero-position sensor and a fixed reference mark are set at the boarding platform to perform zero-position calibration when the gondola passes the reference position. The signals from the encoder and the zero-position sensor are connected to a trigger control unit. The trigger control unit is responsible for counting and frequency division of the encoder signals to generate a fixed number of position indices within one rotation, and for resetting the position count when the zero-position signal is triggered, thereby ensuring the accuracy of the position indices and consistency across rotations.

[0064] After generating the position index, the trigger control unit outputs the frequency-divided signal as an external trigger pulse to the industrial camera, simultaneously triggering the strobe light source. Upon receiving the external trigger, the industrial camera begins exposure and writes the frame number, timestamp, position index, and circle number into its metadata, achieving a one-to-one correspondence between image acquisition and rotation position information. To avoid position errors caused by signal jitter or missing pulses, the trigger control unit also performs jitter removal and interpolation processing on the encoder input signal. When an anomaly is detected, a validity flag is marked, and phase correction is performed after re-triggered at zero position. The output contains a record of image data and the corresponding position index, providing input for subsequent component area correction and feature extraction.

[0065] The intrinsic and extrinsic parameters of the industrial camera were calibrated using a checkerboard calibration board pre-set at the boarding platform, yielding the camera's distortion coefficients and perspective matrix. In the image preprocessing module, OpenCV's perspective transformation function was called to perform geometric correction on the areas of the gondola door, door lock, door gap, pressure bar, fasteners, and suspension arm, ensuring that each component maintains a uniform scale and positional relationship across different rotational revolutions. Within the corrected areas:

[0066] For the door gap area, the Canny operator is used for edge detection to extract the edge lines of the door gap, and the width of the door gap is calculated by combining the pixel spacing.

[0067] For the door lock and lever area, the key point detection network based on HRNet is used to locate the end point of the door lock lever and the end point of the lever, and output the coordinates of the key points;

[0068] For the fastener and cantilever area, a semantic segmentation model based on DeepLabv3+ is used to extract the contour region, and morphological processing is used to obtain the fastener shape parameters and cantilever boundary.

[0069] For the door area, the outline recognition method is used to extract the outer border of the door, and its planar position and tilt angle are calculated based on the perspective-corrected matrix.

[0070] Edges, contours, and segmentation results extracted from the images are categorized as image features, while door gap width, door lock angle, relative position of the pressure bar, fastener geometry, cantilever boundary, and door position parameters are categorized as structural features. During data organization, these results are encoded into vector forms; for example, edge point coordinates are compressed into endpoint and length descriptions, and masked regions are converted into contour polygon parameters. All image and structural features are concatenated according to a fixed field order to form a unified multidimensional vector, which represents the component status data for the corresponding position and circle number.

[0071] In this embodiment, the feature transformation includes:

[0072] The rotational position index is aligned with the component status data and concatenated within the same record. To ensure that the rotational position reflects periodic characteristics, the rotational position index is periodically expanded using sine and cosine functions, and its sine and cosine values ​​are calculated respectively. The results are used as two new positional feature dimensions. The calculation process is performed in real-time in the edge computing unit using the NumPy numerical computing library.

[0073] The concatenated results are organized into a feature vector containing multiple fields:

[0074] Numerical fields from component status data, such as door gap width, door lock lever angle, key point coordinate difference of pressure bar, fastener geometric parameters, suspension arm profile parameters, door position and angle;

[0075] Image feature encoding fields from component status data, such as parameterized descriptions of edge segments and segmented contours;

[0076] Sine and cosine feature fields derived from rotational position expansion.

[0077] The above fields are stored sequentially to form a high-dimensional feature representation vector. The same feature transformation steps are performed on each frame of the image to obtain a high-dimensional feature set containing the part state and rotation position, providing input for the subsequent establishment of a set of position-conditional Koopman operators.

[0078] In this embodiment, the construction of the conditional Koopman operator set includes:

[0079] Based on the rotation position index, a full rotation process is divided into several intervals, for example, intervals formed at fixed angles. The high-dimensional feature representations of the corresponding intervals are used as input, and a training module built using the PyTorch framework is used to independently train a Koopman operator unit within each interval, obtaining a linear operator that describes the state evolution of that interval. After training all intervals, the obtained operator units are combined to form a set of position-conditionalized Koopman operators.

[0080] During training, a block sparsity constraint is applied to the operator matrix. Specifically:

[0081] The operator matrix is ​​divided into multiple sub-blocks, each sub-block corresponding to a group of components that have physical or logical connections;

[0082] Coupler sub-blocks are retained on three sets of components: door lock and door gap, pressure bar and human posture, and fastener and suspension arm.

[0083] For other components that are not directly related, L1 regularization constraints are used to force the corresponding sub-blocks to be sparse to zero.

[0084] For the sub-blocks used for crack detection, growth constraints are set during the training phase. Specifically:

[0085] During the optimization process, the spectral normalization method is invoked to adjust the eigenvalues ​​of the operator matrix;

[0086] The principal eigenvalues ​​of the cracked sub-blocks are restricted to a range greater than one.

[0087] This allows the sub-block to naturally amplify input anomalies during the learning process, thereby capturing the expansion trend of cracks or corrosion.

[0088] After training and constraint application are completed for each interval, the operator units of all intervals are combined in index order to form a complete set of Koopman operators. The trained operator set is stored in matrix form in the model management module of the edge computing device and mapped to the corresponding rotation position intervals for subsequent prediction.

[0089] In this embodiment, the operations performed by the Koopman operator set include:

[0090] After the gondola completes one rotation, the high-dimensional feature representation of a certain position on that rotation is input into the corresponding Koopman operator unit. Matrix multiplication is then used to obtain the predicted data for the same position on the next rotation. This same operation is performed on all positions throughout the entire rotation to obtain the complete predicted rotation data.

[0091] The cross-cycle prediction results are compared position-by-position with the actual component status data collected in the next cycle. The NumPy numerical computing library is used to calculate the Euclidean distance and generate a residual sequence. This residual sequence is stored by position index and serves as one of the inputs for subsequent anomaly score calculation.

[0092] At each position interval, the corresponding Koopman operator unit performs eigenvalue decomposition to extract its principal eigenvalues ​​and corresponding eigenvectors. The spectral features extracted from the same position in two adjacent loops are compared to obtain the spectral offset, and the offset results are stored in vector form.

[0093] The system calls upon a pre-defined logical rule base to verify each logical relationship between the door lock and door gap, the pressure bar and human posture, and the fasteners and suspension arm. The logical verification is performed by the rule matching engine, which records violations as logical constraint violation results and binds them to the corresponding location index and gondola number.

[0094] In this embodiment, the generation of the anomaly level result includes:

[0095] The three types of results are uniformly organized, and the numerical value of the prediction residual, the change in spectral shift, and the violation of logical constraints are combined into a complete record with the position index as the primary key. Each record contains the corresponding gondola number and circle number.

[0096] The compiled records are input into the calculation module, which sequentially reads the three fields: prediction residual, spectral shift, and logic violation. These fields are then weighted according to predefined weights to obtain a composite anomaly score. The weights are trained using historical data before device deployment and remain unchanged during operation.

[0097] A threshold table is established based on historical operational data. Specifically, under normal operating conditions, composite anomaly scores are collected over a long period, their distribution under various location indices is statistically analyzed, and tiered thresholds are determined based on the distribution characteristics. The threshold table is stored internally and can be periodically updated during operation based on newly collected data.

[0098] The composite anomaly score calculated in real time is compared with the corresponding threshold in the threshold table, and three levels are output: safety, warning, and danger. These levels are then bound to the gondola number and location index and stored in the result record, providing input for subsequent risk classification and maintenance recommendations.

[0099] In this embodiment, the generation of the maintenance suggestion includes:

[0100] Each anomaly level result is merged with its corresponding gondola number, rotation position index, and circle number into a single record. This record includes: composite anomaly score, anomaly level, timestamp, and camera number. This record is then added to the results set for the current shift, forming a searchable index list.

[0101] Based on the result set, a mapping from location index to score is established, and a linear color mapping table (with a fixed color sequence corresponding to low to high scores) is used to color the scores at each location. The perspective correction results of the door are used as the base map, and color blocks and location labels are superimposed in a vector overlay manner within the same coordinate system to generate a complete abnormal score map; local images are then output for each gondola, retaining the original resolution and location information.

[0102] Read the logical constraint violation results, sort them by component category and violation code, forming individual entries. Each entry includes: component category (door lock, door gap, pressure bar, fastener, suspension arm, door body), violation code, occurrence location index, circle number, corresponding score and grade, and referenced keyframe number. Write the entries into the violation list table for the current shift.

[0103] For each entry, locate keyframe images, invoke the overlay rendering process, and overlay the following onto the frame: door seam edge line, key points of the door lock and pressure bar, outline of fasteners and suspension arm, and outer frame of the door. Label the image with text indicating the door seam width, door lock lever angle, relative coordinates of the pressure bar and human body key points, fastener geometric parameters, and the door's planar position and tilt angle. Archive the rendered keyframes in pairs with the original frames, naming them according to the following rules: gondola number, position index, and circle number.

[0104] Aggregate the risk rating results for each shift by gondola number, calculate the highest risk rating and corresponding location index for each gondola, and generate a risk rating table. The risk rating table fields include: gondola number, highest risk rating, involved components, list of location indices, score range, and list of keyframe numbers.

[0105] Establish a rule mapping table, using triples (component category, violation code, exception level) as keys, mapping to suggestion text and handling priority. Look up each violation in the table to generate a corresponding suggestion; when the same component in the same compartment appears repeatedly in multiple location indices, merge them into a single suggestion and summarize the location indices. The suggestion text includes: handling action (re-tighten, adjust, replace, review), required tool or spare part name, and suggested completion deadline.

[0106] The anomaly score chart, violation list, evidence keyframes, risk classification table, and maintenance recommendations are compiled into a single-shift report according to a fixed template. The first page of the report lists the gondola number and the highest level comparison table, and is accompanied by local anomaly score charts and evidence frames for each gondola. Exported files are prepared in both printable and parsable data formats (e.g., one formatted document, one structured table, and one image catalog). The catalog names are uniformly named, including the date and shift identifier, and archived in the shift results database for traceability.

[0107] Example 1

[0108] To verify the feasibility of this invention in practice, it was applied to a Ferris wheel with a diameter of 120m and a rated rotation speed of 0.25rpm. A 5-megapixel industrial camera (16mm low-distortion lens, external trigger exposure 1 / 1000s) was placed on the boarding platform and at the circumference. A 2048PPR photoelectric incremental encoder (A / B / Z three-channel) was installed in the drive unit, and an infrared through-beam zero-position sensor was set on the boarding platform. The trigger control unit discretized one circumference into 72 position indices (each 5° increment), and the cameras were triggered synchronously with the strobe. Eight consecutive days of data collection were used as "normal samples" to establish a threshold table; subsequently, online detection and judgment were performed on operating days. Feature extraction and region correction were performed according to the instructions; feature transformation involved periodically expanding the position indices using sin / cos and concatenating them with the component state vector. Koopman training used "position intervals" as units, with each interval yielding a 32×32 linear operator; the operator matrix was divided into three groups (door locks...). Door gaps, pressure bars Human posture, fasteners L1 sparsification is applied to the cantilever arm, and spectral normalization is used to maintain the principal eigenvalue > 1 for the crack detection sub-block. The residual r, the change in principal value Δλ of the operator spectrum, and the logical rule violation flag v (0 / 1) are obtained from the online phase cross-loop same-position prediction. The composite score is defined as S = 0.5·r′ + 0.3·Δλ′ + 0.2·v, where r′ and Δλ′ are the Min–Max normalization results based on historical normal data; the thresholds T_warn and T_danger for each position are obtained from 8 days of normal samples (in this embodiment, the typical thresholds near the gate area are approximately T_warn ≈ 0.62 and T_danger ≈ 0.78).

[0109] Table 1 Online Testing

[0110]

[0111] As shown in Table 1, when the gondola passes index 11, the difference between the Koopman operator prediction and the actual observation increases, the spectral features shift, and the logical rule "the door gap should not widen when the door lock is closed" is triggered. The composite anomaly score of the three types of signals reaches 0.68, exceeding the warning threshold (0.62). The system outputs "warning" and generates maintenance suggestions "recheck door lock closure and adjust the latch".

[0112] In the gate area indices 11–13, this invention, through joint modeling of “location-conditional Koopman + block sparsity + growth constraint”, can provide early warning and timely shutdown decision for progressive gate gap widening under controllable noise conditions; non-gate areas maintain low scores and low false alarm rate, verifying the feasibility and effectiveness of this invention in actual operation scenarios.

[0113] 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 safety inspection and anomaly identification of Ferris wheel gondolas based on machine vision, characterized in that, Includes the following steps: Images captured by industrial cameras positioned on the Ferris wheel boarding platform and around the circumference are used to determine the position of the gondola during rotation, combined with rotation angle signals. The areas of the gondola door, door lock, door gap, pressure bar, fasteners and suspension arm are corrected, and image features and structural features are extracted to form component status data; The component state data is subjected to feature transformation to generate a feature representation carrying rotational position information; Based on the feature representation, a set of position-conditional Koopman operators is established, and a block-sparse structure is adopted to reflect the logical coupling relationship between different components. Growth constraints are set in the operator sub-blocks corresponding to crack detection. The Koopman operator set is used to perform cross-cycle same-position prediction and full-cycle cyclic consistency modeling to obtain prediction residuals, spectral shift results and logical constraint violation results. Based on the predicted residuals, spectral shift results, and logical constraint violation results, a composite anomaly score is generated, and the anomaly level of the gondola component is determined according to the composite anomaly score, and the anomaly level and maintenance recommendations are output. The construction of the conditionalized set of Koopman operators includes: Based on the rotation position signal output by the angle encoder, the rotation process of one revolution of the Ferris wheel is divided into several position intervals, each interval corresponding to a specific cabin position; The high-dimensional feature representation carrying rotational position information is input into the training module, and the corresponding Koopman operator unit is trained in each interval to characterize the change pattern of component state data within the interval. Divide the intervals according to the rotation position and train the corresponding Koopman operator units in each interval. Combine all the operator units to form an operator set. During training, block sparsity constraints are applied to the operator set so that the door lock and door gap, pressure bar and human posture, fastener and suspension arm form mutually coupled operator sub-blocks, while the remaining components remain sparse. Growth constraints are set in the operator sub-block corresponding to crack detection to characterize the trend of crack or corrosion gradually expanding over time.

2. The method for safety detection and anomaly identification of Ferris wheel gondolas based on machine vision according to claim 1, characterized in that, The rotation position is obtained by an angle encoder installed on the Ferris wheel drive unit and calibrated in conjunction with a zero-position sensor at the boarding platform. The rotation position is synchronized with the image acquisition timestamp.

3. The method for safety detection and anomaly identification of Ferris wheel gondolas based on machine vision according to claim 1, characterized in that, The steps for generating component status data include: Based on the camera's intrinsic and extrinsic parameters, perspective transformation is performed to correct the areas of the gondola door, door lock, door gap, pressure bar, fasteners, and suspension arm. Within the corrected area, image features are extracted through edge detection, semantic segmentation, and contour recognition, including door contour, door seam edge, fastener contour, and cantilever contour. Structural features are extracted through key point detection and geometric calculation, including door gap width, door lock lever angle, and the relative distance between the lever position and key points on the human body. The image features and structural features are combined to form component status data.

4. The method for safety detection and anomaly identification of Ferris wheel gondolas based on machine vision according to claim 1, characterized in that, The feature transformation includes: When performing feature transformation on the component state data, the value of the rotation position is concatenated with the component state data; The rotational position is periodically expanded using sine and cosine functions; The unfolding results are combined with the component state data to generate a high-dimensional feature representation carrying rotational position information.

5. The method for safety detection and anomaly identification of Ferris wheel gondolas based on machine vision according to claim 1, characterized in that, The operations performed by the set of Koopman operators include: After the gondola completes one revolution, the component status data at the same position in this revolution is input into the Koopman operator unit of the corresponding position interval to obtain the prediction data for the same position in the next revolution. For each position, the actual data of the next rotation is compared with the predicted data obtained in the previous step, and the prediction residual sequence is calculated and recorded. At each position, spectral feature descriptions are extracted based on the corresponding Koopman operator unit and input data. The spectral features of the current loop and the previous loop at the same position are obtained and compared. The spectral shift results are calculated and recorded. Establish a list of logical relationships between door locks and door gaps, pressure bars and human posture, and fasteners and suspension arms; At each location, the component status data is checked item by item, the triggered violation items are recorded, and logical constraint violation results are generated. The prediction residuals, spectral shift results, and logical constraint violation results are summarized by location index and gondola number.

6. The method for safety detection and anomaly identification of Ferris wheel gondolas based on machine vision according to claim 1, characterized in that, The generation of the anomaly level result includes: Input the obtained predicted residual sequence, spectral shift results, and logical constraint violation results into the anomaly score generation module; In the anomaly score generation module, the prediction residual, spectral shift result and logical constraint violation result are linearly combined according to preset weights to obtain the composite anomaly score. After the composite anomaly score is generated, it is compared item by item with the stored threshold table; Based on the comparison results, the composite anomaly scores are divided into different levels, corresponding to three intervals: safe, warning, and dangerous, to form anomaly level results; Output the anomaly level result and bind the result with the gondola number and location index.

7. The method for safety detection and anomaly identification of Ferris wheel gondolas based on machine vision according to claim 1, characterized in that, The generation of the maintenance recommendations includes: Store multiple maintenance rules, each rule corresponding to a risk level and a specific component category; After generating the risk classification results, the maintenance rules that match the risk level and component category are invoked. Based on the invoked maintenance rules, maintenance suggestions are generated, including the maintenance components, maintenance types, and priorities, and then stored in association with the risk classification results. The risk classification results and maintenance suggestions are simultaneously output to the host computer interface, control system, or work order management module as reference information for maintenance personnel.

Images