An elevator door lock type adaptive identification method and system for precision measurement

Abstract

The application discloses a precision measurement-oriented elevator door lock model adaptive identification method and system, which comprises the following steps: obtaining an original image of an elevator door lock and preprocessing to form an adaptive image with a customized rectangular proportion; a lightweight customized feature extraction network is constructed and trained, including a pre-processing module and a 3-level cascaded customized Inception module, the customized Inception module adopts a four-branch collaborative architecture combining factorized long strip convolution, double deformable convolution and SENet; the image to be identified is input into the feature extraction network, and after the output feature map is sequentially subjected to global average pooling, a full connection layer and a Softmax function, the model classification result of the elevator door lock is output; based on the model classification result, a standard parameter is called from a preset database to generate a closed-loop control instruction, a mechanical arm is driven to perform pose compensation and dynamic adjustment, and automatic precision quantitative measurement is performed. The application can improve the accuracy of elevator door lock model identification, and realize intelligent industrial detection closed loop from visual identification to precision measurement execution.

Description

Technical Field

[0001] This invention belongs to the field of automated detection and industrial visual recognition technology, specifically relating to an adaptive recognition method and system for elevator door lock models for precision measurement. Background Technology

[0002] Elevators, as an indispensable means of transportation in modern high-rise buildings, are directly related to the safety of users' lives and property. Elevator door locks not only effectively prevent elevator doors from opening accidentally under abnormal external forces, but also prevent accidents caused by operational malfunctions. However, the enclosed environment and complex structure of elevator shafts have always posed significant challenges to the automated visual recognition and status monitoring of door lock components.

[0003] In recent years, deep learning technology has been widely introduced into the field of industrial visual inspection. In particular, architectures represented by multi-branch deep networks have demonstrated high feature extraction efficiency in handling routine image classification tasks. However, when directly applied to the extremely specialized and precise industrial scenario of elevator shafts, existing technologies have revealed the following serious limitations:

[0004] First, the physical morphology suffers from "structural mismatch" and poor adaptability to spatial deformation. From a geometric morphological perspective, elevator door locks, hooks, and their minute meshing gaps typically exhibit a highly pronounced flattened, elongated distribution. Existing deep learning models often force a standard square input resolution, which directly disrupts the true physical length-to-width ratio of slender mechanical components. Furthermore, in actual inspections, industrial cameras are constrained by robotic arms or installation poses, resulting in slight deviations in the shooting angle and perspective distortion. The fixed, rigid square convolutional kernels in traditional networks cannot effectively adapt to this irregular spatial deformation, easily leading to missed detections or misjudgments.

[0005] Secondly, deep network architectures are highly susceptible to "semantic submersion" of microscopic features. Differences in highly similar elevator door lock models often depend solely on micrometer-level local slots or edge markings. Existing deep networks, in pursuit of generalization capabilities, blindly stack depth, causing these crucial microscopic mechanical features to be completely obliterated during multi-layer pooling downsampling, resulting in an extremely high misclassification rate when dealing with highly similar industrial components.

[0006] Finally, the complex lighting and redundant background in the elevator shaft lead to severe contamination of the feature channels. The lighting conditions inside the elevator shaft are extremely poor and uneven, and the oil stains on the surface and the strong local reflections of the stainless steel metal create extremely complex high-frequency background noise. Traditional deep networks often treat all feature channels equally when performing feature fusion, failing to actively identify and suppress these interfering noises at the lower level, resulting in the semantic features of the core metal components being contaminated by high-light noise.

[0007] In summary, existing deep learning image recognition technologies struggle to overcome bottlenecks such as input distortion, feature overload, and noise pollution when dealing with elevator door locks, and lack hardware-level coordination at the execution level. Therefore, developing an automated detection system that can adapt to elongated physical shapes, resist spatial perspective distortion, actively truncate to protect microscopic features, and perform precise physical quantification measurements based on visual recognition results has become a critical technical challenge that urgently needs to be addressed in the field of intelligent elevator operation and maintenance. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide an adaptive identification method and system for elevator door lock models for precision measurement. This invention can improve the accuracy of elevator door lock model identification and realize an intelligent industrial inspection closed loop from visual recognition to precision measurement execution.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: an adaptive identification method for elevator door lock models for precision measurement, comprising the following steps:

[0010] S1. Data Acquisition and Preprocessing: Acquire the original image containing the elevator door lock; Based on the actual physical geometric features of the elevator door lock, which is flat and elongated, dynamically adjust the resolution of the original image to a customized rectangular ratio that matches the actual physical aspect ratio of the target to obtain an adapted image, so as to avoid spatial compression distortion of the mechanical structure, and thus form a dataset.

[0011] S2. Construct a lightweight, customized feature extraction network, including a preprocessing module and a 3-level cascaded customized Inception module to prevent semantic submersion of microscopic mechanical features during multi-level downsampling. Each customized Inception module includes four parallel functional branches to achieve collaborative optimization against physical interference in the well shaft. The first branch performs channel compression using 1×1 convolutions, the second branch uses 1×1 convolutions followed by deformable convolutions to compensate for macroscopic perspective distortion, the third branch uses 1×1 convolutions followed by factorized strip convolutions to extract the strip-shaped features of the door lock, and the fourth branch uses 3×3 max pooling followed by deformable convolutions to spatially align microscopic gap features. The output of each branch is connected to an SENet channel attention module to suppress noise from metal mirror reflections. The feature extraction network is trained using a training set.

[0012] S3. Input the image to be recognized into the trained feature extraction network. The feature map output by the feature extraction network is processed by the global average pooling layer (using global feature mapping to overcome the size uncertainty caused by non-standard rectangular input), the fully connected layer and the Softmax function in sequence, and then outputs the elevator door lock model classification result.

[0013] S4. Based on the output model classification results, retrieve the corresponding factory physical standard parameters for the model from the preset database, including the standard distance from the lock hook to the baseline. Using this standard distance as the absolute physical reference, generate closed-loop control commands to drive the robotic arm equipped with a camera to perform pose compensation and dynamic adjustment. Perform automated precision quantitative measurement on the actual relative position of the current elevator door lock's lock hook and the baseline to determine the actual engagement depth.

[0014] Furthermore, in step S1, the resolution of the original image is dynamically adjusted to a customized rectangular ratio, specifically including:

[0015] S101. Extract the effective bounding box of the elevator door lock in the original image, calculate its aspect ratio, and scale the image to a non-standard rectangular size with an aspect ratio between 2:1 and 3:1.

[0016] S102. Localized restricted contrast enhancement (CLAHE) and morphological filtering are used to suppress reflections on the metal surface and messy lines in the well background.

[0017] Further, in step S2, the preprocessing module is used to perform preliminary feature extraction and spatial dimensionality reduction on the adapted image, and then perform multi-scale deformation correction and feature extraction through a cascaded customized Inception module; the preprocessing module includes a 7×7 convolution module, a batch normalization module, a LeakyReLU activation module and a 3×3 max pooling module in sequence.

[0018] Further, in step S2, the deformable convolutions in the second and fourth branches are fourth-generation deformable convolutions (DCNv4); the factorized strip convolutions in the third branch are 1×5 convolutions followed by 5×1 convolutions.

[0019] Furthermore, in step S2, the output of each branch is connected in series with the SENet channel attention module. Specifically, before the output feature maps of the four branches are spliced ​​and fused, the feature channels of each branch are weighted by the SENet channel attention module to suppress the highlight noise channel. Then, the feature maps of the four branches after weighting are spliced ​​together.

[0020] Further, in step S2, the training method of the feature extraction network includes:

[0021] The training set is divided into multiple mini-batches, each mini-batch containing a set number of samples;

[0022] Each mini-batch is input into a lightweight, customized feature extraction network built on the GoogLeNet neural network model to obtain the predicted probability of each category. The loss value of the mini-batch is then calculated using the cross-entropy loss function to obtain the difference between the predicted result and the true label. The expression for the cross-entropy loss function is as follows:

[0023]

[0024] Where L represents the cross-entropy loss function, N represents the total number of categories, and y i p represents the true label of the i-th category. i This represents the predicted probability of the i-th category;

[0025] Calculate the gradient of the loss function with respect to the model parameters, and use the Adam optimizer to dynamically adjust the model's weights and biases until the set number of training epochs is reached;

[0026] After training is complete, save the trained feature extraction network.

[0027] Furthermore, the weight update formula for the Adam optimizer is:

[0028]

[0029] in, This indicates the model weights for the current round. The updated model weights are represented by η, which represents the learning rate. This represents the gradient of the loss function under the current weights.

[0030] Furthermore, step S3 also includes: inputting the test set into the trained feature extraction network, outputting the model classification result, and evaluating the model classification result:

[0031] Calculate the accuracy of the model classification results:

[0032]

[0033] Where Accuracy represents the accuracy rate, TP represents the total number of samples correctly predicted as positive, TN represents the total number of samples correctly predicted as negative, FP represents the total number of samples incorrectly predicted as positive, and FN represents the total number of samples incorrectly predicted as negative.

[0034] When the accuracy reaches 80% or higher, the model classification result is deemed to meet the requirements; otherwise, the feature extraction network is retrained by increasing the training set samples, adjusting hyperparameters, or optimizing the network model.

[0035] The present invention also provides an elevator door lock model adaptive recognition system for performing the above method, comprising:

[0036] The data acquisition and preprocessing module is used to acquire the original image of the elevator door lock and dynamically adjust its resolution to a customized rectangular ratio that matches the physical aspect ratio of the door lock in order to construct a dataset.

[0037] The network construction and training module is used to build a lightweight, customized feature extraction network, including a preprocessing module and a 3-level cascaded customized Inception module. Each customized Inception module includes four parallel functional branches: the first branch performs channel compression using 1×1 convolutions; the second branch uses 1×1 convolutions followed by deformable convolutions to compensate for macroscopic perspective distortion; the third branch uses 1×1 convolutions followed by factorized strip convolutions to extract the strip-shaped features of door locks; and the fourth branch uses 3×3 max pooling followed by deformable convolutions to spatially align microscopic gap features. The output of each branch is connected to an SENet channel attention module. The feature extraction network is trained using a training set.

[0038] The classification decision module is used to call the trained feature extraction network and input the test set or the image to be identified to output the model classification result of the elevator door lock, which is used for model evaluation or subsequent module applications.

[0039] The linkage measurement execution module is used to automatically retrieve the corresponding factory physical standard parameters based on the identified model, and drive the robotic arm to perform a precise quantitative measurement of the actual meshing distance based on the standard parameters.

[0040] Compared with existing technologies, this invention has the following advantages: Addressing the issue of the flat, elongated physical characteristics of elevator door locks and the tendency for perspective deformation of the meshing gaps in narrow shaft conditions, this invention performs deep physical reconstruction of the Inception module during the feature extraction stage. It innovatively constructs a four-branch physical collaborative architecture, including factorized elongated convolution, dual deformable convolution (DCNv4) for macro- and micro-deformation correction, and full-branch positive channel attention (SENet). This accurately corrects and purifies irregular mechanical edges and high-brightness reflective interference. In the inference and decision-making stage, it abandons deep architectures and redundant auxiliary branches that cause "semantic submersion" of micro-features, relying on a lightweight 3-level backbone and global dynamic pooling to directly output high-precision model prediction. Furthermore, based on the identified specific model, the system automatically retrieves the corresponding factory physical standard parameters (especially the standard distance from the lock hook to the baseline) and generates closed-loop control commands accordingly. This commands guide the camera-equipped robotic arm to perform precise quantitative measurement of the current actual meshing distance based on the physical baseline, completely realizing a closed-loop intelligent industrial inspection system from visual perception to hardware execution. Attached Figure Description

[0041] Figure 1This is a flowchart of an adaptive elevator door lock model identification method for precision measurement provided in an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of a lightweight, customized feature extraction network and a feature extraction and classification decision process in an embodiment of the present invention;

[0043] Figure 3 This is a schematic diagram of the structure of the customized Inception module in an embodiment of the present invention. Detailed Implementation

[0044] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0045] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0046] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0047] like Figure 1 As shown in the figure, this embodiment provides an adaptive identification method for elevator door lock models for precision measurement, and its implementation steps are as follows.

[0048] S1. Data Acquisition and Preprocessing: Acquire the original image containing the elevator door lock; based on the actual physical geometric characteristics of the elevator door lock, which is flat and elongated, dynamically adjust the resolution of the original image to a customized rectangular ratio that matches the actual physical aspect ratio of the target, thus obtaining an adapted image to avoid spatial compression distortion of the mechanical structure, thereby forming a dataset.

[0049] In actual industrial settings, high frame rate industrial cameras such as the MV-CA060-10GC, equipped with corresponding lenses, are guided by robotic arms or inspection robots to designated workstations to capture images of elevator door locks. All captured images are then combined to obtain the required raw image.

[0050] Then, the original image is preprocessed.

[0051] S101. Perform adaptive scaling on the original image. Considering the extremely obvious flat and elongated physical geometric features of the elevator door lock and hook, this method abandons the traditional approach of forcibly scaling the network to a square. This method first extracts the effective bounding box of the elevator door lock in the original image, calculates its aspect ratio, and dynamically scales it proportionally to a non-standard customized rectangular size (in this embodiment, the aspect ratio is maintained between 2:1 and 3:1) to ensure that the elongated mechanical structure and the tiny meshing gaps do not undergo any spatial compression distortion before being input into the network.

[0052] S102. Basic Image Enhancement. To address the strong reflections that easily occur on the metal surfaces of the pulleys and door locks in the image, Local Restricted Contrast Enhancement (CLAHE) is used to suppress overexposure, and morphological filtering is applied to suppress metal surface reflections and messy lines in the well background, providing a structurally clear input for subsequent feature extraction by the network.

[0053] S2. Network Construction and Training: A lightweight, customized feature extraction network is constructed, including a preprocessing module and a 3-level cascaded customized Inception module to prevent semantic submersion of microscopic mechanical features during multi-level downsampling. Each customized Inception module includes four parallel functional branches to achieve collaborative optimization against physical interference in the well shaft. The first branch performs channel compression using 1×1 convolutions, the second branch uses 1×1 convolutions cascaded with deformable convolutions to compensate for macroscopic perspective distortion, the third branch uses 1×1 convolutions cascaded with factorized strip convolutions to extract the strip-shaped features of the door lock, and the fourth branch uses 3×3 max pooling cascaded with deformable convolutions to spatially align microscopic gap features. Each branch's output is cascaded with an SENet channel attention module to suppress noise from metallic mirror reflections. The structure of the feature extraction network is as follows: Figure 2 As shown. Then, the feature extraction network is trained using the training set.

[0054] The preprocessing module performs preliminary feature extraction and spatial dimensionality reduction on the adapted image, followed by multi-scale deformation correction and feature extraction via a cascaded customized Inception module. The preprocessing module sequentially includes a 7×7 convolution module, a batch normalization module, a LeakyReLU activation module, and a 3×3 max pooling module.

[0055] like Figure 3 As shown, in the customized Inception module, this method designs four parallel cooperative branches to address physical disturbances in the wellbore environment:

[0056] The first branch preserves the cross-channel mapping of the original features through 1×1 convolution;

[0057] The second branch uses 1×1 convolution to reduce the dimension, and then concatenates it with fourth-generation deformable convolution (DCNv4) to use its adaptive offset field to compensate for the global perspective deformation of the macroscopic outline of the elevator door lock caused by the viewpoint of the robotic arm.

[0058] The third branch addresses the "slender and flat" geometric features of the elevator door lock. Instead of using conventional large-scale square convolutions, it adopts a spatial factorization strategy to construct a 1×1 convolution followed by 1×5 and 5×1 asymmetric strip convolutions. This reduces computational load while accurately capturing the shape and orientation of the long strip edges of the door lock.

[0059] The fourth branch addresses the pain point of spatial misalignment caused by local micro-imprints and fine slots by constructing an adversarial structure of "3×3 max pooling → DCNv4 → 1×1 convolution". Max pooling first extracts sharp mechanical high-frequency edges, and the second DCNv4 operator is specifically designed to perform extremely fine secondary adaptive geometric alignment for these micro-gap features that cause spatial misalignment.

[0060] Finally, before concatenating the feature maps from the four branches, an innovative approach involves orthogonally concatenating SENet channel attention modules at the end of each branch. This design forces all spatial features at all scales to undergo channel-level specular noise filtering via SENet before fusion, thereby precisely suppressing the specular artifacts caused by the stainless steel door lock under bright reflections at the underlying level. The activation function within the module is LeakyReLU.

[0061] The training method for the feature extraction network is as follows:

[0062] 1) Divide the training set into multiple mini-batches, each mini-batch containing a set number of samples. In this embodiment, each mini-batch contains 8 samples.

[0063] 2) Each mini-batch is input into a lightweight, customized feature extraction network built on the GoogLeNet neural network model to obtain the predicted probability of each category. The loss value of the mini-batch is then calculated using the cross-entropy loss function to obtain the difference between the predicted result and the true label. The expression for the cross-entropy loss function is:

[0064]

[0065] Where L represents the cross-entropy loss function, N represents the total number of categories, and y i p represents the true label of the i-th category. i This represents the predicted probability of the i-th category.

[0066] 3) Parameter optimization. Calculate the gradient of the loss function with respect to the model parameters, and dynamically adjust the model's weights and biases using the Adam optimizer. The weight update formula for the Adam optimizer is:

[0067]

[0068] in, This indicates the model weights for the current round. The updated model weights are represented by η, which represents the learning rate. This represents the gradient of the loss function under the current weights.

[0069] The feature extraction network is trained iteratively until the set number of training epochs is reached. After training is complete, the trained feature extraction network is saved.

[0070] S3. Model classification decision based on global dynamic pooling (inference testing phase)

[0071] The test set is input into the trained feature extraction network, which outputs a feature map. Then, as follows... Figure 2 As shown, after processing through a global average pooling layer, a fully connected layer, and a Softmax function, the elevator door lock model classification results are output.

[0072] After image features are processed and highly refined through a three-level customized Inception module, they are directly input into the Global Average Pooling layer at the end of the backbone network. The Global Average Pooling layer can adapt to various non-standard rectangular feature map sizes, compressing them into a one-dimensional global feature vector. This vector is then passed through a fully connected layer and standardized using the Softmax activation function. The formula for calculating the Softmax activation function is:

[0073]

[0074] in, This represents the linear combination output of neurons corresponding to the i-th category in the output layer. This represents the linear combination output of neurons corresponding to the j-th category in the output layer.

[0075] Then, the model classification results are evaluated, specifically as follows:

[0076] Calculate the accuracy of the model classification results:

[0077]

[0078] Where Accuracy represents the accuracy rate, TP represents the total number of samples correctly predicted as positive, TN represents the total number of samples correctly predicted as negative, FP represents the total number of samples incorrectly predicted as positive, and FN represents the total number of samples incorrectly predicted as negative.

[0079] When the accuracy reaches 80% or higher, the model classification result is deemed to meet the requirements; otherwise, the feature extraction network is retrained by increasing the training set samples, adjusting hyperparameters, or optimizing the network model.

[0080] After successful evaluation, use the model saving function to save the trained and evaluated model parameters for subsequent loading and use. Use the following function to save the model:

[0081] model.save('model.h1') # Save the model

[0082] from keras .models import load_model

[0083] model = load_model('model.h1') # Load the model

[0084] Then, the image to be identified is input into the saved feature extraction network, which outputs the model classification results of the elevator door lock.

[0085] S4. Execute a closed loop of physical linkage and precision measurement based on visual recognition results.

[0086] In actual industrial deployments, after accurately outputting the specific elevator door lock model through model classification decision-making, the system automatically retrieves the corresponding factory physical standard parameters from a preset database, primarily the standard distance parameter from the lock hook to the baseline (a physical reference baseline is engraved on the lock hook of each elevator at the factory). Using this standard distance parameter as the absolute physical reference, the system sends closed-loop control commands to the execution end, driving the camera-equipped robotic arm to perform dynamic pose compensation. After aligning the optimal observation angle, the system performs high-precision feature localization of the lock hook and physical baseline area for this specific elevator door lock model and performs sub-pixel-level automated precision quantization measurement of the current actual engagement distance. This step directly transforms the model result from pure visual recognition into physical guidance for the downstream robotic arm to perform measurements based on the factory baseline, realizing an intelligent detection closed loop from visual perception to precision physical measurement.

[0087] This embodiment also provides an elevator door lock model adaptive recognition system for performing the above method, including: a data acquisition and preprocessing module, a network construction and training module, a classification decision module, and a linkage measurement execution module.

[0088] The data acquisition and preprocessing module is used to acquire the original image of the elevator door lock and dynamically adjust its resolution to a customized rectangular ratio that matches the physical aspect ratio of the door lock in order to construct a dataset.

[0089] The network construction and training module is used to construct a lightweight, customized feature extraction network, including a preprocessing module and a 3-level cascaded customized Inception module. Each customized Inception module includes four parallel functional branches: the first branch performs channel compression using 1×1 convolution; the second branch uses 1×1 convolution followed by deformable convolution to compensate for macroscopic perspective distortion; the third branch uses 1×1 convolution followed by factorized strip convolution to extract the strip-shaped features of a door lock; and the fourth branch uses 3×3 max pooling followed by deformable convolution to spatially align microscopic gap features. The output of each branch is connected to an SENet channel attention module. The feature extraction network is trained using a training set.

[0090] The classification decision module is used to call the trained feature extraction network and input the test set or the image to be identified to output the elevator door lock model classification result for model evaluation or subsequent module application.

[0091] The linkage measurement execution module is used to automatically retrieve the corresponding factory physical standard parameters according to the identified model, and drive the robotic arm to perform a precise quantitative measurement of the actual meshing distance based on the standard parameters.

[0092] This embodiment also provides a computer-readable storage medium storing computer program instructions that, when executed by a processor, implement the above-described method.

[0093] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0094] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0095] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0096] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0097] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A precision measurement-oriented elevator door lock model adaptive recognition method, characterized by, Includes the following steps: S1. Data Acquisition and Preprocessing: Acquire the original image containing the elevator door lock; based on the actual physical geometric features of the elevator door lock, which is flat and elongated, dynamically adjust the resolution of the original image to a customized rectangular ratio that matches the actual physical aspect ratio of the target, thus obtaining the adapted image, which in turn constitutes the dataset. S2. Construct a lightweight, customized feature extraction network, including a preprocessing module and a 3-level cascaded customized Inception module. Each customized Inception module includes four parallel functional branches: the first branch performs channel compression using 1×1 convolutions; the second branch uses 1×1 convolutions followed by deformable convolutions to compensate for macroscopic perspective distortion; the third branch uses 1×1 convolutions followed by factorized strip convolutions to extract the strip-shaped features of door locks; and the fourth branch uses 3×3 max pooling followed by deformable convolutions to spatially align microscopic gap features. Each branch's output is connected to an SENet channel attention module to suppress noise from metallic mirror reflections. The feature extraction network is then trained using a training set. S3. Input the image to be identified into the trained feature extraction network. The feature map output by the feature extraction network is processed by the global average pooling layer, the fully connected layer and the Softmax function in sequence to output the elevator door lock model classification result. S4. Based on the output model classification results, retrieve the corresponding factory physical standard parameters for the model from the preset database, including the standard distance from the lock hook to the baseline. Using this standard distance as the absolute physical reference, generate closed-loop control commands to drive the robotic arm equipped with a camera to perform pose compensation and dynamic adjustment. Perform automated precision quantitative measurement on the actual relative position of the current elevator door lock's lock hook and the baseline to determine the actual engagement depth.

2. The adaptive elevator door lock model recognition method for precision measurement according to claim 1, characterized in that, In step S1, the resolution of the original image is dynamically adjusted to a customized rectangular ratio, specifically including: S101. Extract the effective bounding box of the elevator door lock in the original image, calculate its aspect ratio, and scale the image to a non-standard rectangular size with an aspect ratio between 2:1 and 3:

1. S102. Localized restricted contrast enhancement and morphological filtering are used to suppress reflections on the metal surface and messy lines in the well background.

3. The adaptive elevator door lock model recognition method for precision measurement according to claim 1, characterized in that, In step S2, the preprocessing module is used to perform preliminary feature extraction and spatial dimensionality reduction on the adapted image, and then multi-scale deformation correction and feature extraction are performed through a cascaded customized Inception module; the preprocessing module includes a 7×7 convolution module, a batch normalization module, a LeakyReLU activation module and a 3×3 max pooling module in sequence.

4. The adaptive elevator door lock model recognition method for precision measurement according to claim 1, characterized in that, In step S2, the deformable convolutions in the second and fourth branches are fourth-generation deformable convolutions DCNv4; the factorized strip convolutions in the third branch are 1×5 convolutions followed by 5×1 convolutions.

5. The adaptive elevator door lock model recognition method for precision measurement according to claim 1, characterized in that, In step S2, the output of each branch is connected in series with the SENet channel attention module. Specifically, before the output feature maps of the four branches are stitched together, the feature channels of each branch are weighted by the SENet channel attention module to suppress the highlight noise channel. Then, the feature maps of the four branches after weighting are stitched together.

6. The adaptive elevator door lock model recognition method for precision measurement according to claim 1, characterized in that, In step S2, the training method for the feature extraction network includes: The training set is divided into multiple mini-batches, each mini-batch containing a set number of samples; Each mini-batch is input into a lightweight, customized feature extraction network built on the GoogLeNet neural network model to obtain the predicted probability of each category. The loss value of the mini-batch is then calculated using the cross-entropy loss function to obtain the difference between the predicted result and the true label. The expression for the cross-entropy loss function is as follows: wherein L denotes a cross-entropy loss function, N denotes a total number of classes, y i denotes a true label of the i-th class, p i denotes a predicted probability of the i-th class; Calculate the gradient of the loss function with respect to the model parameters, and use the Adam optimizer to dynamically adjust the model's weights and biases until the set number of training epochs is reached; After training is complete, save the trained feature extraction network.

7. The adaptive elevator door lock model recognition method for precision measurement according to claim 6, characterized in that, The weight update formula for the Adam optimizer is: in, This indicates the model weights for the current round. The updated model weights are represented by η, where η represents the learning rate. This represents the gradient of the loss function under the current weights.

8. The adaptive elevator door lock model recognition method for precision measurement according to claim 1, characterized in that, Step S3 further includes: inputting the test set into the trained feature extraction network, outputting the model classification result, and evaluating the model classification result. Calculate the accuracy of the model classification results: Where Accuracy represents the accuracy rate, TP represents the total number of samples correctly predicted as positive, TN represents the total number of samples correctly predicted as negative, FP represents the total number of samples incorrectly predicted as positive, and FN represents the total number of samples incorrectly predicted as negative. When the accuracy reaches 80% or higher, the model classification result is deemed to meet the requirements; otherwise, the feature extraction network is retrained by increasing the training set samples, adjusting hyperparameters, or optimizing the network model.

9. An elevator door lock model adaptive recognition system for performing the method as described in any one of claims 1-8, characterized in that, include: The data acquisition and preprocessing module is used to acquire the original image of the elevator door lock and dynamically adjust its resolution to a customized rectangular ratio that matches the physical aspect ratio of the door lock in order to construct a dataset. The network construction and training module is used to build a lightweight, customized feature extraction network, including a preprocessing module and a 3-level cascaded customized Inception module. Each customized Inception module includes four parallel functional branches: the first branch performs channel compression using 1×1 convolutions; the second branch uses 1×1 convolutions followed by deformable convolutions to compensate for macroscopic perspective distortion; the third branch uses 1×1 convolutions followed by factorized strip convolutions to extract the strip-shaped features of door locks; and the fourth branch uses 3×3 max pooling followed by deformable convolutions to spatially align microscopic gap features. The output of each branch is connected to an SENet channel attention module. The feature extraction network is trained using a training set. The classification decision module is used to call the trained feature extraction network and input the test set or the image to be identified to output the model classification result of the elevator door lock, which is used for model evaluation or subsequent module applications. The linkage measurement execution module is used to automatically retrieve the corresponding factory physical standard parameters according to the identified model, and drive the robotic arm to perform a precise quantitative measurement of the actual meshing distance based on the standard parameters.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by a processor, the method described in any one of claims 1-8 is implemented.

Images