Wet-bulb globe temperature detection method combined with clothing correction and electronic device
By collecting environmental parameters and clothing images in real time, and using image recognition algorithms to automatically identify clothing types and update correction values, the problem of misjudgment of clothing correction values in existing wet-bulb black-bulb temperature detection methods has been solved. This has enabled accurate and timely assessment of thermal stress and reduced the risk of occupational heatstroke.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHANGSHU INSTITUTE OF TECHNOLOGY
- Filing Date
- 2026-05-08
- Publication Date
- 2026-07-24
AI Technical Summary
Existing wet-bulb spherical temperature detection methods cannot accurately obtain the clothing correction values for workers, leading to misjudgments of the thermal stress index. This is especially problematic in dynamic work scenarios where timely updates are not possible, posing a risk of occupational heatstroke.
By employing computer vision technology, environmental parameters and clothing images are collected in real time. The clothing type is identified using image recognition algorithms, and the wet-bulb and black-bulb temperatures are corrected based on a clothing correction value mapping table, thereby achieving automatic updates to the clothing correction values.
To ensure the objectivity and rigor of thermal stress safety assessments, prevent underestimation of the thermal stress index due to neglecting clothing factors, protect worker safety, and prevent occupational heatstroke accidents.
Smart Images
Figure CN122454549A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of heat measurement technology, and in particular to a wet-bulb sphere temperature detection method and electronic device that incorporates clothing calibration. Background Technology
[0002] The wet bulb globe temperature (WBGT) index is currently recognized internationally as the most effective indicator for assessing heat stress in high-temperature working environments.
[0003] Current wet-bulb temperature (WBGT) measurements rely on obtaining environmental physical parameters to calculate a baseline WBGT value. The ACGIH standard explicitly requires that the assessment of workers' actual heat load must incorporate a Clothing Adjustment Value (CAV). Furthermore, when workers wear non-standard workwear (such as double-layered chemical protective suits or SMS polypropylene bodysuits), the ambient physical temperature alone cannot accurately reflect the body's heat load. Existing detection methods cannot detect the workers' clothing conditions.
[0004] In the improved existing technology, when facing high-standard safety assessments, the acquisition of CAV values relies entirely on the inspectors' manual observation of the work site and manual input of CAV values by referring to complex standard forms (such as distinguishing between "summer work clothes", "double-layer woven clothing", "SMS polypropylene coveralls", etc.).
[0005] However, protective clothing is becoming increasingly specialized, with vastly different CAV values resulting from variations in the breathability of different materials (e.g., SMS polypropylene coveralls with a CAV of +0.5℃, microporous membrane coveralls with a CAV of +1.0℃, and vapor barrier coveralls with a CAV as high as +11.0℃). Yet, they often appear highly similar (e.g., all are white coveralls). Inspectors find it difficult to accurately distinguish between similar-looking materials with different breathability by visual inspection. Selecting the wrong CAV value can lead to a significant deviation in the final thermal stress index. Furthermore, in dynamic work environments, parameters are often not automatically updated when workers change protective equipment, causing measurement results to become disconnected from actual risks. This can easily lead to occupational heatstroke accidents due to underestimating heat risks. Summary of the Invention
[0006] To address the shortcomings of the existing technology, this invention provides a wet-bulb spherical temperature (CAV) detection method incorporating clothing correction, solving the problems of high misjudgment risk and difficulty in adapting to changes in work attire when relying on manual judgment to obtain CAV parameters. This invention also provides an electronic device for implementing this wet-bulb spherical temperature (CAV) detection method incorporating clothing correction.
[0007] The technical solution of the present invention is as follows:
[0008] A wet-bulb sphere temperature detection method incorporating clothing correction includes:
[0009] Real-time acquisition of environmental parameters and acquisition of clothing image information of workers in the target area to be measured. The environmental parameters include at least natural wet-bulb temperature, black-bulb temperature and dry-bulb temperature.
[0010] The image recognition algorithm is used to extract features from the clothing image information and identify the clothing type. The image recognition algorithm includes firstly locating the target and extracting the region of interest through the target detection model, detecting the human target in the image, and cropping the region of interest of the torso. Then, the region of interest is input into the classification network. The global average pooling module is replaced by a texture coding layer at the end of the backbone network of the classification network.
[0011] The basic WBGT value is calculated based on environmental parameters. The corresponding clothing correction value is found through a pre-stored clothing type and clothing correction value mapping table based on the determined clothing type. The effective wet-bulb black-bulb temperature is obtained by correcting the basic WBGT value with the clothing correction value.
[0012] Furthermore, the texture encoding layer includes a visual dictionary comprising a set of visual words and a learnable smoothing factor. The texture encoding layer calculates residual vectors by combining local feature descriptors from the high-dimensional semantic features extracted by the backbone network with visual words from the set of visual words, and calculates weights for the residual vectors based on the corresponding smoothing factor. Based on the weights, all residual vectors are weighted and aggregated in the spatial dimension to obtain an aggregated feature vector. The aggregated feature vectors of all visual words are concatenated to obtain a high-dimensional material texture feature vector for classification. The smoothing factor is obtained by a set of learnable unconstrained parameters through a positive value mapping function. The unconstrained parameters are updated during training as the gradient of the loss function is backpropagated.
[0013] Furthermore, when identifying clothing type, the confidence level of clothing type is output, and when the confidence level of clothing type is lower than a preset threshold, multiple candidate clothing types are output for user confirmation; if no user confirmation result is received, the clothing type is set to the type with the highest clothing correction value among the multiple candidate clothing types.
[0014] Furthermore, the clothing type is identified by using an image recognition algorithm to identify no fewer than three consecutive frames of images. When the identification results of each frame are consistent and the confidence level of the clothing type is not lower than a preset threshold, the identification result is taken as the determined clothing type.
[0015] Furthermore, when multiple workers are present in the target area, personnel identifiers are established for each worker, the region of interest (ROI) of each worker's torso is extracted, and the corresponding clothing type is identified. Based on each worker's clothing type, a corresponding clothing correction value is determined. The effective wet-bulb spherical temperature is calculated separately for each worker, and the maximum effective wet-bulb spherical temperature among the multiple workers is taken as the effective wet-bulb spherical temperature of the target area. This avoids underestimating the area's thermal risk due to different workers wearing different protective clothing.
[0016] Furthermore, the mapping table between clothing type and clothing calibration value is updated via local configuration or remote update. When relevant thermal stress evaluation standards or enterprise safety management rules are adjusted, the clothing calibration values corresponding to different clothing types can be updated without changing the main detection process of the wet-bulb black bulb temperature detection device.
[0017] Furthermore, the types of clothing include at least: ordinary work clothes, double-layer woven clothes, SMS polypropylene bodysuits, microporous coated bodysuits, and vapor barrier bodysuits.
[0018] Furthermore, when the mapping table of garment types and garment correction values is configured according to the ACGIH standard, the garment correction values are as follows: 0°C for ordinary work clothes, +3.0°C for double-layer woven garments, +0.5°C for SMS polypropylene coveralls, +1.0°C for microporous membrane coveralls, and +11.0°C for vapor barrier coveralls.
[0019] Furthermore, the effective wet-bulb black-bulb temperature obtained by correcting the base WBGT value with the clothing correction value is the sum of the base WBGT value and the clothing correction value.
[0020] Furthermore, the method includes the step of: based on the effective wet-bulb black bulb temperature, retrieving and outputting suggested work / rest time allocation schemes for different labor intensity levels through a pre-stored "work / rest time limit table".
[0021] Another technical solution of the present invention is: an electronic device, the electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor;
[0022] The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform the aforementioned wet-bulb black bulb temperature detection method combined with clothing correction.
[0023] Compared with the prior art, the present invention has the following advantages:
[0024] This invention introduces computer vision technology into the field of thermal stress monitoring. Through algorithmic quantitative analysis of clothing material characteristics, it eliminates safety hazards caused by human error and ensures the objectivity and rigor of thermal stress safety assessments. In dynamic work scenarios, when workers change protective equipment, the CAV parameters can be automatically identified and updated, allowing for timely assessment of thermal risks. This prevents the thermal stress index from being underestimated due to neglecting clothing factors, better protecting frontline workers and preventing occupational heatstroke accidents. Attached Figure Description
[0025] Figure 1 This is a schematic flowchart of a wet-bulb black sphere temperature detection method incorporating clothing correction, as an example.
[0026] Figure 2 This is a flowchart of an image processing algorithm.
[0027] Figure 3 This is a schematic diagram of the modules of an electronic device as an example. Detailed Implementation
[0028] The present invention will be further described below with reference to embodiments, but these are not intended to limit the scope of the invention.
[0029] Please combine Figure 1 As shown in the figure, the wet-bulb sphere temperature detection method combined with clothing correction in this embodiment is as follows:
[0030] Step 1: Collect environmental parameters in real time and calculate the basic WBGT value according to the ISO 7243 formula: Environmental parameters include natural wet-bulb temperature. Black ball temperature and dry bulb temperature .
[0031] Indoor / No solar radiation: ;
[0032] Outdoors / With solar radiation: .
[0033] Step 2: Real-time acquisition of clothing image information of workers in the target area for intelligent clothing type recognition:
[0034] To accurately distinguish between different protective suits specified in the ACGIH standard (such as work clothes, double-layer woven clothes, SMS polypropylene coversalls, microporous coversalls / polyolefin coversalls, vapor-barrier coversalls, etc.), a lightweight convolutional neural network algorithm is used for identification. In the implementation of the image processing algorithm, this invention employs a cascaded deep learning architecture, the main process of which is as follows: Figure 2 As shown, it is divided into the following two stages.
[0035] Phase 1: Target localization and Region of Interest (ROI) extraction.
[0036] To meet the stringent real-time requirements of embedded devices (e.g., inference latency less than 50ms), this embodiment preferably employs the single-stage object detection model YOLOv8-Nano. This model is based on an anchor-free architecture and has an extremely low parameter count (approximately 3.2M parameters). Alternatively, the EfficientDet-Lite model proposed by Tan et al. (CVPR 2020) can be used, leveraging its BiFPN feature fusion structure to maintain detection stability under complex lighting conditions. Using either the YOLOv8-Nano or EfficientDet-Lite model, human targets in the image are detected, and the region of interest (ROI) of the torso is cropped, eliminating interference from background heat sources or devices.
[0037] Phase 2: Fine-grained material analysis and clothing type output.
[0038] The cropped ROI image is input into a lightweight classification network. The lightweight classification network includes a feature extraction backbone network, a material feature enhancement module, and a classification result output module.
[0039] The feature extraction backbone network in this embodiment preferably adopts the MobileNetV3 architecture (Howard et al., ICCV 2019) designed specifically for mobile devices. Furthermore, when deploying to edge computing chips with extremely low power consumption and limited computing power, the backbone network is preferably the MobileNetV3-Small model. In the MobileNetV3-based feature extraction backbone network, stacked inverted residual bottleneck modules and depthwise separable convolutions are used as basic feature extraction units. Based on this efficient underlying structure, low-level semantic features, including clothing edge contours and basic color gradients, can be extracted in real time under the limited computing power of embedded devices. Then, utilizing the SE-Block (Squeeze-and-Excitation) channel attention mechanism integrated within the architecture, the weights of feature channels are adaptively learned and allocated during the feature extraction process. This effectively suppresses invalid feature channels related to background environment and lighting changes, and significantly amplifies high-frequency feature channels related to fabric material, outputting a weighted and optimized high-dimensional semantic feature map. In addition, it uses the quantization-friendly h-swish activation function to ensure fixed-point computation accuracy on NPU chips such as Rockchip RV1106.
[0040] At the end of a conventional feature extraction backbone network (such as the standard MobileNetV3), Global Average Pooling (GAP) is typically used to directly reduce the high-dimensional feature map to a one-dimensional vector. However, the spatial smoothing property of GAP inevitably obliterates the high-frequency local microscopic features of the clothing surface. In the identification scenario of thermal stress protective clothing, these obliterated features include the microscopic texture of geometric shape (such as the distribution of fine pressing points unique to SMS nonwoven fabric), and the core criterion reflecting the material's air-barrier performance—the surface optical reflectance characteristics (i.e., gloss difference, for example: the significant difference between the sharp brightness gradient and specular highlight response produced by the dense coating on the surface of a completely moisture-proof vapor barrier coverall under illumination and the soft diffuse reflection micro-shadow produced by a breathable microporous membrane coverall). Since special protective clothing with extremely different breathability often has highly similar appearance contours (such as all being white coveralls), relying on GAP for simple spatial averaging will lead to serious confusion of key features of different materials.
[0041] To address this issue, this invention eliminates the conventional GAP module at the end of the backbone network and replaces it with a Texture Encoding Layer to form a material feature enhancement module.
[0042] In this embodiment, the texture coding layer further performs fine-grained feature coding of the material based on the high-dimensional semantic feature map output by the feature extraction backbone network. Its composition and specific feature processing are as follows:
[0043] (1) Initialization and visual dictionary construction:
[0044] The texture coding layer internally maintains a visual dictionary, which consists of a set of K D-dimensional visual words C={c1,c2,...,c...} K} and the corresponding set of learnable smoothing factors s={s1, s2, ..., s} K The system consists of a set of visual words, C, uniformly distributed during the system initialization phase. Random initialization is performed within the range. The smoothing factor s is explained in detail in step (3).
[0045] (2) Residual calculation and end-to-end update:
[0046] Suppose that the feature map output by the backbone network contains N local feature descriptors X={x1, x2, ..., x...} N For each local feature descriptor x i And each visual word c k Calculate the residual vector r between them ik The formula is:
[0047]
[0048] During the supervised training phase, the visual dictionary uses the backpropagation algorithm to adaptively learn and update parameters based on the data distribution of input image features, thereby obtaining the optimal visual word representation for fine-grained classification of protective clothing materials without the need for offline pre-learning.
[0049] (3) Calculate the smoothing distribution weights:
[0050] To preserve the high-frequency local microscopic features of the material (such as differences in gloss and bonding points), a soft assignment mechanism with learning parameters is employed to calculate the descriptor x. i Belongs to the visual word c k weight a ik The formula is as follows:
[0051]
[0052] Among them, s k and s jLet be a learnable smoothing factor corresponding to the cluster centers, and it be a positive parameter. To ensure that the soft assignment weights decay as the residual distance increases, the smoothing factor can be determined by a set of learnable, unconstrained parameters u = {u1, u2, ..., u...}. K} is obtained through the positive value mapping function, i.e., s k = φ(u k ) + ε, where φ(·) is the mapping function for outputting non-negative or positive values, and ε is a constant greater than 0. Positive value mapping functions include, but are not limited to, the softplus function, the exponential function, or the square function. During training, the parameter u... k The smoothing factor s is updated by backpropagation of the gradient of the loss function, thereby affecting the corresponding smoothing factor s. k Adaptive adjustment. This design allows the network to automatically learn and match different feature distribution scales for different material micro-features (such as sharp specular response or broad diffuse texture), thereby improving the accuracy of weight allocation.
[0053] (4) Residual aggregation and normalization:
[0054] Combined with the calculated allocation weight a ik We perform weighted aggregation on all residual vectors in the spatial dimension to obtain the result for the visual word c. k aggregated feature vector e k :
[0055]
[0056] Finally, the aggregated feature vectors of all K visual words are concatenated and L2 normalized to output a high-dimensional material texture feature vector with scale invariance, which is then fed into the final fully connected layer for classification.
[0057] Furthermore, considering the limitations of memory bandwidth and computing power of edge processors in industrial settings, to prevent dimensionality explosion during high-dimensional feature aggregation, this embodiment inserts a 1×1 transitional convolutional layer between the MobileNetV3 backbone network and the texture coding layer. This layer is used to linearly compress and reduce the dimensionality of the high-dimensional feature map output by the backbone network (576 dimensions) to D=128 dimensions. Based on this, the number of visual words K in the visual dictionary within the texture coding layer is preferably set between 16 and 64. In this embodiment, K=16 is configured. This value provides sufficient representation capacity to cover the typical micro-textures and optical feature clusters of five types of protective clothing, while ensuring that the dimension of the material texture feature vector output after residual aggregation (K×D = 16×128 = 2048 dimensions) is within a reasonable computational range, thus balancing high recognition accuracy and low inference latency on embedded devices.
[0058] Finally, the feature vector enters the classification result output module, where it is fed into the fully connected layer (FC layer) at the end for linear dimension mapping, and combined with the Softmax classifier to output the final clothing category probability.
[0059] Furthermore, to prevent recognition jumps or misjudgments caused by rapid personnel movement, occlusion, low lighting, etc., a confidence smoothing mechanism based on a time window and a risk conservative confirmation mechanism are introduced. Specifically, a sliding recognition window is established for N consecutive frames of images of the same worker, where N is an integer not less than 3. When the clothing type recognition results are consistent in N consecutive frames, and the confidence score of the clothing type corresponding to each frame is not lower than a preset threshold, such as 0.7, the clothing type is locked as the current worker's confirmed clothing type. When the recognition results of N consecutive frames are inconsistent, or the highest clothing type confidence score in any frame is lower than the preset threshold, a manual assistance mode is entered. The display interface outputs several candidate clothing types with the highest confidence scores, such as the top 3 candidate clothing types, prompting the user for confirmation. If the user does not confirm within a preset time, the device defaults to the clothing type with the highest clothing correction value among the candidate clothing types.
[0060] In this embodiment, the cascaded model training strategy of the image recognition algorithm is as follows:
[0061] (1) Training data and parameters of the first-stage object detection model:
[0062] Since the first stage only needs to locate the "human target," this invention directly uses the YOLOv8-Nano model weights pre-trained on a general large-scale visual dataset (MS COCO dataset) (which can be downloaded directly from the official website), without needing to retrain for specific clothing. This model can stably output the bounding box of the human torso, enabling the cropping and extraction of the region of interest (ROI).
[0063] (2) Training data and parameters of the lightweight classification network in stage 2:
[0064] This embodiment utilizes MobileNetV3-Small, pre-trained on the general large-scale dataset (ImageNet), as the feature extraction base. This design ensures that the model possesses deep visual perception capabilities from the early stages of training. However, since conventional open-source datasets lack fine-grained classification of industrial special protective clothing, the classification network in this embodiment still requires fine-tuning through custom dataset construction and transfer learning.
[0065] Training Data Acquisition: For the identified M types of clothing (e.g., M=5), images of each type of clothing are collected under different lighting conditions, shooting angles, shooting distances, background environments, clothing wrinkle states, soiled states, and occlusion states. Ideally, at least 1000 original images should be collected for each type of clothing, and each category should include multiple samples from different brands, models, batches, or actual items to avoid the model learning only the appearance features of specific samples. Data augmentation techniques such as random cropping, horizontal flipping (50% probability), and color dithering are used to expand the dataset, which is then divided into training, validation, and test sets in an 8:1:1 ratio. Preferably, the data is grouped according to the clothing sample number, collection scene, or collection date to ensure that images of the same sample or from the same collection scene do not appear simultaneously in the training and test sets. All images are uniformly scaled to a preset resolution and normalized by mean before being input into the network.
[0066] Training hyperparameter settings: The MobileNetV3 network with texture coding layers is optimized end-to-end using the Stochastic Gradient Descent (SGD) optimizer, AdamW optimizer, or other gradient optimization algorithms. As a preferred embodiment, the SGD optimizer is used, with an initial learning rate of 0.01, which decays to one-tenth of its original value when the validation set error stops decreasing; the momentum parameter is set to 0.9; the weight decay coefficient is set to 0.0001; and the batch size is set to 64. The model is trained until the validation set loss converges, or until a preset early stopping condition is met.
[0067] After training, the model's accuracy, recall, and confusion rate for various clothing categories are evaluated using an independent test set. The evaluation focuses on cases where clothing categories with higher correction values are misidentified as categories with lower correction values. The clothing type confidence threshold can be calibrated based on the recognition results on the validation set, allowing low-confidence, category confusion, or high-risk category misjudgments to be handled in either a manual assistance mode or a conservative clothing correction value calculation mode.
[0068] Step 3: Matching garment correction values.
[0069] The database stores the mapping relationship between clothing types and Clothing Correction Values (CAVs). This mapping relationship can be updated, allowing you to find the corresponding clothing correction value based on the clothing type determined in step 2. Taking the ACGIH standard as an example, the mapping relationship between clothing types and clothing correction values is as follows:
[0070] Regular work clothes (lightweight summer workwear) -> CAV = 0℃;
[0071] Double-layer woven garment -> CAV = +3.0℃;
[0072] SMS polypropylene bodysuit -> CAV = +0.5℃;
[0073] Microporous membrane-coated bodysuit -> CAV = +1.0℃;
[0074] Steam barrier bodysuit -> CAV = +11.0℃.
[0075] The values above are only a standard configuration example. In actual applications, the mapping table can be updated according to the current applicable standards, enterprise security specifications or regulatory requirements.
[0076] Step 4: Calculation of effective wet-bulb black bulb temperature.
[0077] Calculate the effective wet-bulb temperature value: .
[0078] In some embodiments, multiple workers may be present simultaneously within the target area to be measured. In this case, the effective wet-bulb temperature (MBT) for each worker can be calculated separately, i.e.:
[0079]
[0080] Where p represents the operator identifier, CAV p This represents the clothing correction value corresponding to the worker. Simultaneously, the device can select the maximum value among multiple clothing correction values for different workers as the area clothing correction value, i.e.: And calculate the effective wet-bulb black-bulb temperature of the target area:
[0081]
[0082] This results in a more conservative assessment of regional thermal risk. In other words, the method described above uses the maximum effective wet-bulb black-bulb temperature of the target area as the value of the effective wet-bulb black-bulb temperature corresponding to multiple operators.
[0083] Step 5: Assessment and output of work and rest time.
[0084] The effective wet-bulb black bulb temperature calculated in step S4 Based on the pre-stored "Work / Rest Time Limit Table" in the ACGIH standard, it automatically retrieves and outputs suggested work / rest time allocation schemes for different labor intensity levels (mild / moderate / severe / extremely severe).
[0085] In some embodiments, step 2 can be performed by a cloud server. For example, in embodiments with wireless communication capabilities, the device can upload the acquired clothing images to the cloud server via Wi-Fi or a 4G / 5G module. The cloud server uses a larger-scale model to perform fine-grained identification of clothing types and distributes the identification results and the latest industry standards (such as the latest version of the CAV adjustment released by ACGIH), enabling the method to remotely upgrade the clothing feature library via OTA.
[0086] It is readily understood that the implementation of the methods described in the above embodiments can be based on a computer program, which is a set of instructions that can be executed by a computer (i.e., by a processor). When executed by the processor, the computer program implements a wet-bulb sphere temperature detection method incorporating clothing correction.
[0087] Furthermore, at least some of the computer programs associated with the methods of the embodiments can be distributed in a computer program product including a computer-readable storage medium carrying computer-usable instructions for one or more processors. This computer-readable storage medium can be provided in various forms, including non-transitory forms, such as, but not limited to, one or more disks, optical discs, magnetic tapes, chips, and magnetic and electronic storage. Further, the computer program can also be stored in a memory, which is part of an electronic device that also has a processor electrically connected to the memory. The computer program stored in the memory can be executed by the processor, thereby implementing a wet-bulb black bulb temperature detection method combined with clothing correction.
[0088] Please combine Figure 3 As shown, a specific embodiment of an electronic device includes the following parts:
[0089] Processor: A system-on-a-chip (SoC) with an integrated neural network acceleration unit (NPU) is used as the main control chip. An embedded AI chip based on the ARM Cortex-A series or RISC-V architecture is preferred, such as the Rockchip RV1106 or other processors with equivalent performance.
[0090] Environmental parameter sensor group: including natural wet-bulb temperature sensor, black bulb temperature sensor and dry-bulb temperature sensor, used to collect environmental physical parameters;
[0091] Image acquisition module: A wide-angle visible light camera (such as a GC2053 or IMX307 camera module) with a field of view (FOV) of not less than 90 degrees, used to cover the work site. In another embodiment, the image acquisition module also includes an optional low-resolution infrared thermal imaging module (such as an MLX90640) to assist in determining the thermal insulation performance of the clothing surface;
[0092] Storage module: eMMC is used as the main storage solution to store the preset clothing type database and its corresponding clothing calibration values (CAV), and also to store the measured environmental parameters;
[0093] Human-Computer Interaction Module: This module includes a display unit and an input unit. The display unit uses a color LCD screen to show real-time measurement data, identified clothing types, and evaluation results. The input unit has physical buttons (such as "Power," "Menu / Confirm," "Scroll Up," and "Scroll Down") for users to perform power on / off operations, system settings, and various menu operations.
[0094] Power Management Module: Includes a built-in rechargeable polymer lithium battery pack to provide continuous power to the system; equipped with a power management chip (PMIC) and a low dropout linear regulator (LDO) to convert the battery voltage into stable voltages such as 3.3V / 1.8V required by the core control unit, sensor group and camera; and features a USB Type-C charging and data interface for battery charging and firmware upgrades / data export.
Claims
1. A method for detecting wet-bulb spherical temperature using clothing correction, characterized in that, include: Real-time acquisition of environmental parameters and acquisition of clothing image information of at least one worker in the target area to be measured, wherein the environmental parameters include at least natural wet-bulb temperature, black-bulb temperature and dry-bulb temperature; The image recognition algorithm is used to extract features from the clothing image information and identify the clothing type. The image recognition algorithm includes firstly locating the target and extracting the region of interest through the target detection model, detecting the human target in the image, and cropping the region of interest of the torso. Then, the region of interest is input into the classification network. The global average pooling module is replaced by a texture coding layer at the end of the backbone network of the classification network. The basic WBGT value is calculated based on environmental parameters. The corresponding clothing correction value is found through a pre-stored clothing type and clothing correction value mapping table based on the determined clothing type. The effective wet-bulb black-bulb temperature is obtained by correcting the basic WBGT value with the clothing correction value.
2. The wet-bulb spherical temperature detection method combined with clothing correction according to claim 1, characterized in that, The texture encoding layer includes a visual dictionary comprising a set of visual words and a learnable smoothing factor. The texture encoding layer calculates residual vectors by combining local feature descriptors from the high-dimensional semantic features extracted by the backbone network with visual words from the set of visual words. It then calculates weights for the residual vectors based on the corresponding smoothing factor. Based on these weights, all residual vectors are weighted and aggregated in the spatial dimension to obtain an aggregated feature vector. The aggregated feature vectors of all visual words are concatenated to obtain a high-dimensional material texture feature vector for classification. The smoothing factor is obtained from a set of learnable unconstrained parameters through a positive value mapping function. These unconstrained parameters are updated during training as the gradient of the loss function is backpropagated.
3. The wet-bulb spherical temperature detection method combined with clothing correction according to claim 1, characterized in that, When identifying clothing type, output the clothing type confidence score. If the clothing type confidence score is lower than a preset threshold, output multiple candidate clothing types for user confirmation. If no user confirmation is received, set the clothing type to the type with the highest clothing correction value among the multiple candidate clothing types.
4. The wet-bulb spherical temperature detection method combined with clothing correction according to claim 1, characterized in that, To identify clothing types, an image recognition algorithm is used to identify at least three consecutive frames of images. When the identification results of each frame are consistent and the confidence level of the clothing type is not lower than a preset threshold, the identification result is taken as the determined clothing type.
5. The wet-bulb spherical temperature detection method combined with clothing correction according to claim 1, characterized in that, When there are multiple workers in the target area to be measured, personnel identifiers are established for each worker, the region of interest of the torso of each worker is extracted and the corresponding clothing type is identified, and the corresponding clothing correction value is determined based on the clothing type of each worker. The effective wet-bulb black-bulb temperature is calculated separately for each worker, and the maximum value of the effective wet-bulb black-bulb temperature of the multiple workers is taken as the effective wet-bulb black-bulb temperature of the target area to be measured.
6. The wet-bulb spherical temperature detection method combined with clothing correction according to claim 1, characterized in that, The mapping table between clothing type and clothing correction value is updated through local configuration or remote update.
7. The wet-bulb spherical temperature detection method combined with clothing correction according to claim 1, characterized in that, The types of clothing include at least: ordinary work clothes, double-layer woven clothes, SMS polypropylene bodysuits, microporous coated bodysuits, and vapor barrier bodysuits.
8. The wet-bulb spherical temperature detection method combined with clothing correction according to claim 7, characterized in that, When the mapping table of garment types and garment correction values is configured according to the ACGIH standard, the garment correction values are as follows: 0°C for ordinary work clothes, +3.0°C for double-layer woven garments, +0.5°C for SMS polypropylene coveralls, +1.0°C for microporous membrane coveralls, and +11.0°C for vapor barrier coveralls.
9. The wet-bulb sphere temperature detection method combined with clothing correction according to claim 1, characterized in that, The effective wet-bulb black-bulb temperature is obtained by correcting the base WBGT value with the clothing correction value. It is the sum of the base WBGT value and the clothing correction value.
10. The wet-bulb spherical temperature detection method combined with clothing correction according to any one of claims 1 to 9, characterized in that, The steps include: based on the effective wet-bulb black bulb temperature, using the pre-stored "Work / Rest Time Limit Table", retrieving and outputting suggested work / rest time allocation schemes for different labor intensity levels.
11. An electronic device, characterized in that, The electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; The memory stores a computer program executable by the at least one processor, which is then executed by the at least one processor to enable the at least one processor to perform the wet-bulb black-bulb temperature detection method combined with clothing correction as described in any one of claims 1 to 10.