A cleaning detection method and system for an elevator maintenance scene
By constructing a logic-guided model and optimizing it with a specific loss function, the problem of low recognition accuracy of image classification models in elevator maintenance scenarios was solved, achieving efficient cleaning detection and reducing the need for manual review.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHONGSHAN SIDA TECH CO LTD
- Filing Date
- 2026-01-28
- Publication Date
- 2026-06-05
AI Technical Summary
Existing general image classification models have low recognition accuracy in elevator maintenance scenarios, cannot meet the differentiated needs of cleaning judgment standards, and rely on manual review, resulting in wasted resources.
A target dataset labeled with region and state labels is collected. A logically guided model is constructed using independent batch normalization (BN) layers in a convolutional neural network. The model is then optimized using a specific loss function. Combined with gamma correction and contrast limiting, targeted identification and interpretation are achieved.
It improves the accuracy of cleaning detection in elevator maintenance scenarios, can replace manual review, and reduces the use of human resources.
Smart Images
Figure CN122156718A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image recognition technology, and in particular to a cleaning detection method and system for elevator maintenance scenarios. Background Technology
[0002] The elevator maintenance industry operates on a 15-day cycle system, requiring each elevator to be cleaned and maintained every 15 days and photographed for record-keeping. Currently, there are approximately 8 million elevators in China, with an average of 300,000 to 400,000 maintenance photos being sent back daily. The current practice relies on manual visual inspection of each photo, and the mechanism of cross-verification by two people requires a large amount of human resources.
[0003] To alleviate manual labor pressure, some manufacturers have introduced general image classification models such as ResNet50 and EfficientNet to perform "clean / unclean" binary classification on maintenance images. However, maintenance images are collected from uncommon areas such as the elevator car top and elevator shaft pit. These areas vary significantly in shape; for example, the car top is a flat steel plate, while the pit has a recessed structure. There may also be interference factors such as dim lighting, oil stains reflecting light, and condensation mist. Furthermore, the cleanliness criteria for each area differ significantly. For instance, the car top must be free of any oil stains, while the pit must be free of any metal fragments but allows for a small amount of water. Existing general image classification models lack corresponding training materials and are not designed to differentiate between different cleanliness criteria, resulting in low recognition accuracy during deployment and failing to meet practical needs. Summary of the Invention
[0004] The first aspect of this embodiment discloses a cleaning inspection method for elevator maintenance scenarios, specifically including:
[0005] Collect a target dataset, which includes several visible light images labeled with region labels and status labels; For each visible light image, an independent BN layer corresponding to its region label is introduced into a convolutional neural network to construct a logically guided model. The logical guidance model is optimized using a loss function; The logic-guided model is used to perform a flow-based determination on the image to be recognized.
[0006] As an optional implementation, the area label includes the car roof, sill, and pit; The status labels include clean and unclean.
[0007] As an optional implementation, the step of introducing a corresponding independent Batch Normalization (BN) layer in a convolutional neural network for each visible light image based on its region label to construct a logically guided model includes: - MobileNetV3-Small1.0 with 4.2M parameters was used as the backbone network of the convolutional neural network. The ImageNet pre-training process freezes the global l-weights θ_shared; (to accelerate training and prevent overfitting). Each independent BN layer inference stage is routed via a 2-bit index, without increasing ROM.
[0008] As an optional implementation, a classification head GlobalAvgPool→Dropout(0.2)→FC(1024→2) is constructed for the convolutional neural network based on global average pooling.
[0009] As an optional implementation, the base loss of the loss function is L_cls = 0.7. CrossEntropy+ 0.3 FocalLoss(γ=2) is used to alleviate the situation of excessive clean labels in the target dataset.
[0010] As an optional implementation, the method further includes: Gamma correction is performed on the image to be identified in order to improve the contrast of dark areas in the image.
[0011] As an optional implementation, contrast limiting and the number of block grid divisions are applied to the image to be identified in order to suppress the reflective effect; The contrast ratio is limited to CLAHE clipLimit=2.0, and the number of block grid divisions is limited to tileGridSize=8×8.
[0012] As an optional implementation, the step of using the logical guidance model to perform a traffic splitting determination on the image to be recognized includes: The logical guidance model performs GPS field parsing and region indexing on the image to be identified; The image to be recognized is fed into a separate BN layer by forward computation; The classification head outputs the cleanliness probability corresponding to the image to be identified. If the cleaning probability is higher than the preset cleaning threshold, then the area indicated by the image to be identified is determined to meet the cleaning specifications.
[0013] As an optional implementation, the method further includes: The logic-guided model identifies and determines whether the image to be identified matches any preset region label. If it does not meet the requirements, the image to be identified is determined to be inconsistent with the input specifications.
[0014] The second aspect of this embodiment discloses a cleaning and inspection system for elevator maintenance scenarios, specifically including: The data acquisition and annotation unit is used to acquire a target dataset, which includes several visible light images labeled with region labels and status labels. The model building unit is used to introduce the corresponding independent BN layer in the convolutional neural network for each visible light image based on its region label, and to build a logically guided model. The model optimization unit is used to optimize the logic-guided model using a loss function; The detection and determination unit is used to perform a flow division determination on the image to be recognized using the logic-guided model.
[0015] Compared with the prior art, this embodiment has the following beneficial effects: Targeted datasets from different regions are collected and input into a convolutional network for training. Independent batch normalization (BN) layers are then deployed according to the region labels to achieve targeted recognition and interpretation, significantly improving recognition accuracy. This approach can replace manual review in real-world applications. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in this embodiment, the accompanying drawings used in the embodiment will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 This is a schematic diagram of the workflow of a cleaning and inspection method for elevator maintenance scenarios disclosed in this embodiment; Figure 2 This is a schematic diagram of the system structure of a cleaning and inspection system for elevator maintenance scenarios disclosed in this embodiment. Detailed Implementation
[0018] The technical solutions in this embodiment will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] Example 1 Please see Figure 1 This embodiment discloses a cleaning and inspection method for elevator maintenance scenarios, including: 101. Collect the target dataset, which includes several visible light images labeled with region labels and status labels.
[0020] In this embodiment, the area labels include the car roof, sill, and pit; Status labels include clean and unclean.
[0021] Understandably, area labels can be added or simplified according to actual needs.
[0022] Furthermore, for state labels that may have a third or more types, accurate calibration can be achieved through multi-branch judgment.
[0023] 102. For each visible light image, introduce the corresponding independent BN layer in the convolutional neural network based on its region label to construct a logically guided model.
[0024] In this embodiment, different regions are trained and identified independently in their corresponding BN layers.
[0025] As an optional implementation, for each visible light image, a corresponding independent Batch Normalization (BN) layer is introduced into the convolutional neural network based on its region label to construct a logically guided model, including: MobileNetV3-Small1.0 with 4.2M parameters was used as the backbone network of the convolutional neural network; The ImageNet pre-training process freezes the global l-weights θ_shared; (to accelerate training and prevent overfitting). Each independent BN layer inference stage is routed via a 2-bit index, without increasing ROM.
[0026] As an alternative implementation, a classification head GlobalAvgPool→Dropout(0.2)→FC(1024→2) is constructed for the convolutional neural network based on global average pooling.
[0027] Here, by setting up a separate BN layer, we can ensure that images in specific areas can be identified in a targeted manner, thereby significantly improving the recognition accuracy.
[0028] 103. Use loss function to optimize the logic-guided model.
[0029] In this embodiment, the base loss of the loss function is L_cls = 0.7. CrossEntropy + 0.3 FocalLoss(γ=2) is used to alleviate the situation of excessive clean labels in the target dataset.
[0030] In this case, if there is an excess of specific training data, the loss function can be used to adjust it and other training items to avoid the model from becoming localized.
[0031] 104. A logic-guided model is used to perform flow division and determination on the image to be recognized.
[0032] In this embodiment, gamma correction is performed on the image to be identified in order to improve the contrast of dark areas in the image.
[0033] In this embodiment, contrast limiting and the number of block grid divisions are limited for the image to be recognized in order to suppress the reflection effect; The contrast ratio is limited to CLAHE clipLimit=2.0, and the number of block grid divisions is limited to tileGridSize=8×8.
[0034] As an optional implementation, a logic-guided model is used to perform traffic splitting determination on the image to be recognized, including: The logical guidance model performs GPS field parsing and region indexing on the image to be identified; The independent BN layer corresponding to the split input of the image to be recognized is calculated by forward computation; The classification head outputs the cleanliness probability of the image to be identified. If the cleaning probability is higher than the preset cleaning threshold, the area indicated by the image to be identified is determined to meet the cleaning specifications.
[0035] As an optional implementation, the logic-guided model identifies whether the image to be identified matches any preset region label; If it does not meet the requirements, the image to be recognized is determined to be inconsistent with the input specifications.
[0036] Here, the trained logic-guided model can identify whether the image to be identified was captured accurately, avoiding staged images, and can also accurately divert the image to its corresponding recognition area for targeted interpretation, thereby accurately determining whether the cleaning and maintenance work has been completed properly.
[0037] Compared with the prior art, this embodiment has the following beneficial effects: Targeted datasets from different regions are collected and input into a convolutional network for training. Independent batch normalization (BN) layers are then deployed according to the region labels to achieve targeted recognition and interpretation, significantly improving recognition accuracy. This approach can replace manual review in real-world applications.
[0038] Example 2 Please see Figure 2 This embodiment discloses a cleaning and inspection system for elevator maintenance scenarios, comprising: The data acquisition and annotation unit is used to acquire the target dataset, which includes several visible light images labeled with region labels and status labels. The model building unit is used to introduce the corresponding independent BN layer in the convolutional neural network for each visible light image based on its region label, and build a logically guided model. The model optimization unit is used to guide the model by employing loss function optimization logic. The detection and judgment unit is used to perform a flow division judgment on the image to be recognized using a logic-guided model.
Claims
1. A cleaning inspection method for elevator maintenance scenarios, characterized in that, include: Collect a target dataset, which includes several visible light images labeled with region labels and status labels; For each visible light image, an independent BN layer corresponding to its region label is introduced into a convolutional neural network to construct a logically guided model. The logical guidance model is optimized using a loss function; The logic-guided model is used to perform a flow-based determination on the image to be recognized.
2. The cleaning and inspection method for elevator maintenance scenarios according to claim 1, characterized in that, include: The area labels include the car roof, sill, and pit; The status labels include clean and unclean.
3. The cleaning and inspection method for elevator maintenance scenarios according to claim 1, characterized in that, The step of introducing a corresponding independent Batch Normalization (BN) layer into a convolutional neural network for each visible light image based on its region label to construct a logically guided model includes: MobileNetV3-Small1.0 with 4.2M parameters was used as the backbone network of the convolutional neural network. The ImageNet pre-training process freezes the global l-weights θ_shared; (to accelerate training and prevent overfitting). Each independent BN layer inference stage is routed via a 2-bit index, without increasing ROM.
4. The cleaning and inspection method for elevator maintenance scenarios according to claim 3, characterized in that, include: The classification head GlobalAvgPool→Dropout(0.2)→FC(1024→2) is constructed for the convolutional neural network based on global average pooling.
5. The cleaning and inspection method for elevator maintenance scenarios according to claim 1, characterized in that, include: The base loss of the loss function is L_cls = 0.7 CrossEntropy + 0.3 FocalLoss(γ=2) is used to alleviate the situation of excessive clean labels in the target dataset.
6. The cleaning and inspection method for elevator maintenance scenarios according to claim 1, characterized in that, The method further includes: Gamma correction is performed on the image to be identified in order to improve the contrast of dark areas in the image.
7. The cleaning and inspection method for elevator maintenance scenarios according to claim 6, characterized in that, include: Contrast limiting and the number of block grid divisions are applied to the image to be identified in order to suppress the reflection effect; The contrast ratio is limited to CLAHE clipLimit=2.0, and the number of block grid divisions is limited to tileGridSize=8×8.
8. The cleaning and inspection method for elevator maintenance scenarios according to claim 1, characterized in that, The step of using the logic-guided model to perform traffic splitting and determination on the image to be recognized includes: The logical guidance model performs GPS field parsing and region indexing on the image to be identified; The image to be recognized is fed into a separate BN layer by forward computation; The classification head outputs the cleanliness probability corresponding to the image to be identified. If the cleaning probability is higher than the preset cleaning threshold, then the area indicated by the image to be identified is determined to meet the cleaning specifications.
9. A cleaning and inspection method for elevator maintenance scenarios according to claim 8, characterized in that, The method further includes: The logic-guided model identifies and determines whether the image to be identified matches any preset region label. If it does not meet the requirements, the image to be identified is determined to be inconsistent with the input specifications.
10. A cleaning and inspection system for elevator maintenance scenarios, characterized in that, include: The data acquisition and annotation unit is used to acquire a target dataset, which includes several visible light images labeled with region labels and status labels. The model building unit is used to introduce the corresponding independent BN layer in the convolutional neural network for each visible light image based on its region label, and to build a logically guided model. The model optimization unit is used to optimize the logic-guided model using a loss function; The detection and determination unit is used to perform a flow division determination on the image to be recognized using the logic-guided model.