Comparison method and device for live working shielding clothes and storage medium

Abstract

The invention provides a live working shielding clothes comparison method and device and a storage medium, and the method comprises the steps: tracking the posture data of a user at each time point when the user wears the shielding clothes for live working; loading a classifier and a job detection network; inputting the plurality of attitude data into a classifier for classification to obtain action data; respectively converting the attitude data into an attitude matrix and converting the action data into an action matrix; inputting the attitude matrix into a first backbone network to extract target attitude features; inputting the action matrix into a second backbone network to extract target action features; inputting the target posture feature and the target action feature into an encoder to detect the fitness of the user wearing the shielding clothes for live working; and comparing the fitness of various shielding clothes to generate technical optimization information for the shielding clothes. According to the embodiment of the invention, the shielding clothes continuously converge to the requirement of live-line field operation in the technology evolution process, and the live-line operation efficiency of a user is improved.

Description

Technical Field

[0001] The embodiments of the present application belong to the technical field of deep learning, and in particular, relate to a comparison method, device and storage medium for live working shielding clothing. Background Art

[0002] When workers carry out live work in seasons with higher temperatures, such as the operation and maintenance of ultra-high voltage transmission lines in the summer, they often face a working environment of "three highs" - high voltage, high altitude and high temperature. In order to ensure personal safety during live work, workers wear shielding clothing made of special shielding materials to complete their work.

[0003] Under the premise of maintaining the appearance requirements of the workers, the shielding suit can effectively dissipate heat and cool down through the research and development and selection of fabric materials. The materials and process strategies that are compatible with the passive cooling performance of the original characteristics of the shielding suit such as electromagnetic shielding, durability, safety, and anti-static performance are developed to achieve effective passive cooling of ultra-ultra-high voltage live workers in high temperature environments.

[0004] In addition, the design of the passive cooling optical metamaterial live working shielding suit will also take into account the feasibility and comfort of practical applications, so that it has good air permeability, moisture permeability and softness to ensure the comfort of workers when wearing it for a long time in a high temperature environment. At the same time, its wear resistance, UV aging resistance and washing cycle are strictly tested to ensure its stable performance in a variety of extreme environments.

[0005] Therefore, there are many technical routes for shielding suits, and these technical routes mainly rely on the research and experiments of R&D personnel on operators to gradually evolve, and more reflect the subjective understanding of R&D personnel, which makes there is a certain deviation between shielding suits and the needs of live on-site operations, reducing the efficiency of users' live operations. Summary of the invention

[0006] In view of this, the embodiments of the present application provide a comparison method, device and storage medium for live working shielding clothing, so as to improve the adaptability of the shielding clothing to on-site operations.

[0007] A first aspect of an embodiment of the present application provides a comparison method for live working shielding clothing, comprising:

[0008] Tracking posture data of the user at various time points when the user is wearing a shielding suit and performing live work;

[0009] Loading a classifier and a job detection network; the job detection network includes a first backbone network, a second backbone network and an encoder;

[0010] Inputting a plurality of the posture data into the classifier for classification to obtain action data;

[0011] Respectively converting the posture data into a posture matrix and converting the motion data into a motion matrix;

[0012] Inputting the posture matrix into the first backbone network to extract target posture features;

[0013] Inputting the action matrix into the second backbone network to extract target action features;

[0014] Inputting the target posture feature and the target motion feature into the encoder to detect the adaptability of the user to wear the shielding suit for live-line work;

[0015] The fitness of various shielding suits is compared to generate technical optimization information for the shielding suit.

[0016] A second aspect of an embodiment of the present application provides a comparison device for live working shielding clothing, including:

[0017] A posture data tracking module, used to track the posture data of the user at various time points when the user is wearing a shielding suit to perform live work;

[0018] A network loading module, used to load a classifier and a job detection network; the job detection network includes a first backbone network, a second backbone network and an encoder;

[0019] An action data classification module, used for inputting a plurality of the posture data into the classifier for classification to obtain action data;

[0020] A matrix conversion module, used to convert the posture data into a posture matrix and convert the motion data into a motion matrix;

[0021] A target posture feature extraction module, used for inputting the posture matrix into the first backbone network to extract target posture features;

[0022] A target action feature extraction module, used for inputting the action matrix into the second backbone network to extract target action features;

[0023] A fitness detection module, used for inputting the target posture feature and the target motion feature into the encoder to detect the fitness of the user wearing the shielding suit to perform live-line work;

[0024] The technical optimization information generating module is used to compare the fitness of various shielding suits to generate technical optimization information for the shielding suits.

[0025] A third aspect of an embodiment of the present application provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for comparing live working shielding clothing as described in the first aspect above is implemented.

[0026] A fourth aspect of an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for comparing live working shielding clothing as described in the first aspect above is implemented.

[0027] A fifth aspect of the embodiments of the present application provides a computer program product. When the computer program product is run on a computer, the computer is enabled to execute the method for comparing live working shielding clothing described in the first aspect.

[0028] In this embodiment, when a user wears a shielding suit to perform live-line work, the user's posture data is tracked at various time points; a classifier and a work detection network are loaded; the work detection network includes a first backbone network, a second backbone network and an encoder; multiple posture data are input into the classifier for classification to obtain action data; the posture data are converted into a posture matrix and the action data are converted into an action matrix respectively; the posture matrix is ​​input into the first backbone network to extract the target posture feature; the action matrix is ​​input into the second backbone network to extract the target action feature; the target posture feature and the target action feature are input into the encoder to detect the user's fitness for live-line work in a shielding suit; the fitness of various shielding suits is compared to generate technical optimization information for the shielding suit. This embodiment evaluates the realization of users wearing shielding suits for live-line work from the dimensions of posture, action, etc., so as to quantify the degree of adaptation of shielding suits and live-line work, provide objective data reference for the technical evolution of shielding suits, reduce the subjective influence of R&D personnel, and make shielding suits constantly converge to the needs of live-line on-site work in the process of technical evolution, and improve the efficiency of users' live-line work. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments or prior art descriptions. Obviously, the drawings described below are only some embodiments of the present application, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0030] Figure 1 It is a schematic diagram of a comparison method of live working shielding clothing provided in an embodiment of the present application;

[0031] Figure 2is a schematic diagram of a classifier and a job detection network provided in an embodiment of the present application;

[0032] Figure 3 It is a schematic diagram of another comparison method of live working shielding clothing provided in an embodiment of the present application;

[0033] Figure 4 is a schematic diagram of a comparison device of a live working shielding suit provided in an embodiment of the present application;

[0034] Figure 5 It is a schematic diagram of a terminal device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0035] In the following description, specific details such as specific system structures, technologies, etc. are proposed for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from hindering the description of the present application.

[0036] The technical solution of the present application is described below through specific embodiments.

[0037] Reference Figure 1 , showing a schematic diagram of a comparison method of live working shielding clothing provided in an embodiment of the present application, which may specifically include the following steps:

[0038] Step 101: When a user wears a shielding suit to perform live-line work, the user's posture data is tracked at various time points.

[0039] At the UHV live working site, edge computing nodes can be deployed and cameras can be installed. The cameras are connected to the edge computing nodes using wired methods.

[0040] Generally speaking, edge computing nodes are terminal devices with strong computing capabilities, such as computers, servers or embedded devices, which are equipped with graphics processing units (GPUs) or embedded neural network processors (NPUs).

[0041] Edge computing nodes refer to a new business platform built near the site of ultra-high voltage live operations (i.e., the edge of the network), providing storage, computing, network and other resources, and sinking some key business applications (i.e., detecting heat radiation risks for users) to the edge of the access network to reduce the width and delay losses caused by network transmission and multi-level forwarding. The edge computing node is located between the user and the cloud (server), and is closer to the user (data source) than the traditional cloud. It has the characteristics of miniaturization, distribution and closer to the user. Massive data (such as audio data) no longer needs to be uploaded to the cloud for processing, and data processing is realized at the edge of the network, reducing request response time, reducing network bandwidth, and ensuring data security and privacy.

[0042] When the user is wearing a shielding suit to perform live work, the camera is called at each time point to track the user and collect multiple frames of image data. Each frame of image data is input into a preset posture detection network for human posture estimation (Human Pose Estimation), such as CPM (Convolutional Pose Machine), HRNet (High-Resolution Network), etc., to obtain the user's posture data at each time point.

[0043] Generally, 33 key points can be selected to represent the posture data of the human body to meet the accuracy requirements. Then, the posture data at each moment is a 33-dimensional vector.

[0044] In actual applications, the technical route of shielding clothing has many branches. For example, phase change cold storage materials are used for cooling, which requires prior freezing and is heavy in weight. The setting of the cold insulation layer and the skin-friendly layer leads to limited cooling effect. In addition, the temperature of the contact part of the cooling clothing designed in this way with the human body surface is difficult to control, and it is difficult to reconcile comfort and cooling effect. The compressor principle is used for cooling, and the cooling clothing consists of basic clothing and a refrigeration system, including: air-cooled and liquid-cooled. Air-cooled cooling clothing uses air duct interlayers and fans for cooling, and uses power systems such as fans or pumps. Liquid-cooled cooling clothing needs to be equipped with compressors, refrigerants, etc. Both types of cooling clothing need to be equipped with power supplies, and are usually large in mass and volume, which will cause a large burden on the staff. In addition, the weight distribution of the clothes is uneven, which can easily affect the body balance of the staff, causing physical incoordination of the staff, and it is difficult to perform corresponding work or operations, etc.

[0045] For the operating personnel, they are not concerned about the technical route of the shielding suit, but are more concerned about whether they can complete the live-line work smoothly while wearing the shielding suit.

[0046] Then, this embodiment observes the posture data of the user wearing the shielding suit while performing live-line work, and compares the external performance of the user's work, so as to evaluate whether the technical route of the shielding suit is suitable.

[0047] Step 102: Load the classifier and the job detection network.

[0048] In this embodiment, a multi-classification classifier and operation detection network can be constructed in advance based on machine learning or deep learning, and the classifier and operation detection network are started when comparing and evaluating the technical routes of the shielding suit.

[0049] Among them, Figure 2 As shown, the operation detection network includes a first backbone network Backbone_1, a second backbone network Backbone_2 and an encoder Encoder.

[0050] Step 103: Input the multiple posture data into a classifier for classification to obtain motion data.

[0051] In this embodiment, a sliding window can be added to the posture data. The window slides with a preset step size, which is smaller than the width of the window. During each slide, multiple posture data in the window are input into the classifier for multi-classification to obtain action data.

[0052] Among them, the motion data includes normal working movements, such as standing, raising hands, etc., as well as abnormal working movements, such as falling, holding feet, etc. These abnormal working movements may be caused by the complex environment of live working and the inconvenience caused by shielding clothing.

[0053] In one design, Figure 2 As shown, the classifier includes a gated recurrent unit (Gated Recurrent Units, GRU) and a multilayer perceptron (Multilayer Perceptron, MLP). Then, multiple posture data Pose are input into the gated recurrent unit GRU to extract posture sequence features, and the posture sequence features are input into the multilayer perceptron MLP to be mapped into action data Action.

[0054] In this design, the classifier has a simple structure, meets the requirements of multi-classification in specific scenarios (i.e., detecting the actions of users wearing shielding clothing to perform live work), and has low computational complexity and fast time consumption.

[0055] Step 104: convert the posture data into a posture matrix and convert the motion data into a motion matrix.

[0056] Due to the complex environment of live-line working and the inconvenience caused by shielding suits, the posture data and motion data of users performing live-line working in shielding suits are usually long-time series and slowly changing one-dimensional data, and some small fluctuations have limited impact on the local area.

[0057] In this embodiment, converting one-dimensional posture data into a two-dimensional posture matrix and converting one-dimensional motion data into a two-dimensional motion matrix can amplify the impact of instantaneous changes on local areas and improve the quality of features.

[0058] In the specific implementation, Figure 2 As shown, the posture matrix includes multiple matrices of different scales, namely a first sub-matrix Matrix_1, a second sub-matrix Matrix_2 and a third sub-matrix Matrix_3.

[0059] On the one hand, a first sliding window, a second sliding window and a third sliding window can be added to the posture data Pose respectively; wherein the first sliding window, the second sliding window and the third sliding window are all sliding windows, the width of the first sliding window is smaller than the width of the second sliding window, and the width of the second sliding window is smaller than the width of the third sliding window. By adding windows of different scales, short-term fluctuations and long-term trend fluctuations in the posture data can be captured from different receptive fields.

[0060] In addition, the first sliding window slides according to a preset first step length, the second sliding window slides according to a preset second step length, and the third sliding window slides according to a preset third step length, the first step length is smaller than the second step length, and the second step length is smaller than the third step length.

[0061] Each time the first sliding window slides, in the first sliding window, the transpose of the posture data Pose is multiplied by the posture data Pose to obtain the first sub-matrix Matrix_1, that is, Matrix_1 = Pose T *Pose.

[0062] Each time the second sliding window slides, in the second sliding window, the transpose of the posture data Pose is multiplied by the posture data Pose to obtain the second sub-matrix Matrix_2, that is, Matrix_2 = Pose T *Pose.

[0063] Each time the third sliding window slides, in the third sliding window, the transpose of the posture data Pose is multiplied by the posture data Pose to obtain the third sub-matrix Matrix_3, that is, Matrix_3 = Pose T *Pose.

[0064] On the other hand, a fourth sliding window is added to the action data Action, and the fourth sliding window is a slidable window, which slides according to a preset fourth step length.

[0065] Each time the fourth sliding window slides, in the fourth sliding window, the transposition of the action data Action is equal to the action data Action, and the action matrix Matrix_4 is obtained, that is, Matrix_4=Action T *Action.

[0066] Step 105: Input the posture matrix into the first backbone network to extract the target posture features.

[0067] In this embodiment, if Figure 2 As shown, the posture matrix is ​​input into the first backbone network Backbone_1 to extract features, which are recorded as target posture features.

[0068] In one design, Figure 2 As shown, the first backbone network includes a first convolutional layer Conv_1, a second convolutional layer Conv_2, a third convolutional layer Conv_3, a fourth convolutional layer Conv_4 and a fifth convolutional layer Conv_5.

[0069] When the posture matrix includes a first sub-matrix, a second sub-matrix and a third sub-matrix, the first sub-matrix can be input into the first convolution layer Conv_1 to extract the first candidate posture features, the second sub-matrix can be input into the second convolution layer Conv_2 to extract the second candidate posture features, and the third sub-matrix can be input into the third convolution layer Conv_3 to extract the third candidate posture features.

[0070] Concatenate the first candidate posture feature and the second candidate posture feature into a fourth candidate posture feature;

[0071] The fourth candidate posture feature is input into the fourth convolutional layer Conv_4 to extract the fifth candidate posture feature.

[0072] The fifth candidate posture feature and the third candidate posture feature are concatenated to obtain a sixth candidate posture feature.

[0073] The sixth candidate posture feature is input into the fifth convolutional layer Conv_5 to extract the target posture feature.

[0074] Step 106: Input the action matrix into the second backbone network to extract target action features.

[0075] In this embodiment, if Figure 2 As shown, the action matrix Matrix_4 is input into the second backbone network Backbone_2 to extract features, which are recorded as target action features.

[0076] In one design, Figure 2 As shown, the second backbone network includes a sixth convolutional layer Conv_6, a seventh convolutional layer Conv_7, a first fully connected layer FC_1, a second fully connected layer FC_2 and a third fully connected layer FC_3.

[0077] The action matrix Matrix_4 is input into the sixth convolutional layer Conv_6 to extract the first candidate action feature.

[0078] The first candidate action feature is input into the first fully connected layer FC_1 and mapped into the second candidate action feature.

[0079] Input the first candidate action feature into the second fully connected layer FC_2 and map it into the third candidate action feature;

[0080] Input the first candidate action feature into the third fully connected layer FC_3 and map it into the fourth candidate action feature;

[0081] The first candidate action feature, the second candidate action feature, the third candidate action feature and the fourth candidate action feature are concatenated to form a fifth candidate action feature.

[0082] The fifth candidate action feature is input into the seventh convolutional layer Conv_7 to extract the target action feature.

[0083] Step 107: Input the target posture features and the target motion features into the encoder to detect the user's adaptability to performing live-line work while wearing the shielding suit.

[0084] In this embodiment, if Figure 2 As shown, the target posture features and the target motion features are input into the encoder to detect the user's fitness for live-line working in the shielding suit, where the fitness indicates the degree of adaptation of the shielding suit to the environment and requirements of the live-line working under the established technical route.

[0085] In one design, Figure 2 As shown, the encoder Encoder includes an attention layer Attention, a first depth-separable convolution layer SeparableConv_1, a second depth-separable convolution layer SeparableConv_2 and a fourth fully connected layer FC_4.

[0086] Among them, the first depth-wise separable convolution layer SeparableConv_1 and the second depth-wise separable convolution layer SeparableConv_2 can be decomposed into a depthwise convolution and a 1×1 pointwise convolution. The depthwise convolution extracts the information within the feature map, and the pointwise convolution extracts the features between the feature maps. The two can work together to learn the temporal patterns of posture data and motion data from local to global, effectively improving the quality of features, thereby improving the accuracy of detecting the adaptability of users wearing shielding clothing for live work.

[0087] In this design, the target posture feature and the target action feature are input into the attention layer Attention and fused into the first multimodal feature. In the attention mechanism provided in the attention layer Attention, the target posture feature is taken as the K (query) matrix, the target action feature is taken as the Q (key) matrix and the V (value) matrix, and the attention weight between the target posture feature and the target action feature is calculated. The target action feature is fused into the target posture feature according to the attention weight to obtain the first multimodal feature.

[0088] The first multimodal features are input into the first depthwise separable convolutional layer SeparableConv_1 to extract the second multimodal features.

[0089] Concatenate the first multimodal feature and the second multimodal feature into a third multimodal feature.

[0090] The third multimodal feature is input into the second depth-wise separable convolutional layer SeparableConv_2 to extract the fourth multimodal feature.

[0091] Concatenate the first multimodal feature and the fourth multimodal feature into a fifth multimodal feature;

[0092] Input the fifth multimodal feature into the fourth fully connected layer FC_4 and map it into the sixth multimodal feature;

[0093] The sixth multimodal feature is activated using activation functions such as Sogmoid (S-type function) to obtain the user's fitness for live-line work in shielding clothing.

[0094] Step 108: Compare the fitness of various shielding suits to generate technical optimization information for the shielding suits.

[0095] In this embodiment, the adaptability of shielding suits under various technical routes can be compared, and the advantages and disadvantages of shielding suits under various technical routes in user operation performance can be analyzed in conjunction with data mining algorithms such as frequent itemsets and Apriori. On this basis, in conjunction with NLP (Natural Language Processing) technology, such as LLM (Large Language Model), etc., technical optimization information is generated for the shielding suit for reference by R&D personnel, providing a data basis for their technological evolution.

[0096] Of course, the adaptability of shielding clothing under various technical routes can also be summarized and concluded to obtain a technical report as a reference for R&D personnel as a stage of successful technological evolution. With the cooperation of R&D personnel, technical optimization information for shielding clothing can be generated. This embodiment does not limit this.

[0097] In this embodiment, when a user wears a shielding suit to perform live-line work, the user's posture data is tracked at various time points; a classifier and a work detection network are loaded; the work detection network includes a first backbone network, a second backbone network and an encoder; multiple posture data are input into the classifier for classification to obtain action data; the posture data are converted into a posture matrix and the action data are converted into an action matrix respectively; the posture matrix is ​​input into the first backbone network to extract the target posture feature; the action matrix is ​​input into the second backbone network to extract the target action feature; the target posture feature and the target action feature are input into the encoder to detect the user's fitness for live-line work in a shielding suit; the fitness of various shielding suits is compared to generate technical optimization information for the shielding suit. This embodiment evaluates the realization of users wearing shielding suits for live-line work from the dimensions of posture, action, etc., so as to quantify the degree of adaptation of shielding suits and live-line work, provide objective data reference for the technical evolution of shielding suits, reduce the subjective influence of R&D personnel, and make shielding suits constantly converge to the needs of live-line on-site work in the process of technical evolution, and improve the efficiency of users' live-line work.

[0098] Reference Figure 3 , shows a schematic diagram of another comparison method of live working shielding clothing provided in an embodiment of the present application, which may specifically include the following steps:

[0099] Step 201: When a user wears a shielding suit to perform live-line work, the user's posture data is tracked at various time points.

[0100] Step 202: Load the classifier and the job detection network.

[0101] The operation detection network includes a first backbone network, a second backbone network and an encoder.

[0102] Step 203: Input the plurality of posture data into a classifier for classification to obtain motion data.

[0103] Step 204: convert the posture data into a posture matrix and convert the motion data into a motion matrix.

[0104] Step 205: Input the posture matrix into the first backbone network to extract the target posture features.

[0105] Step 206: Input the action matrix into the second backbone network to extract target action features.

[0106] Step 207: Input the target posture feature and the target motion feature into the encoder to detect the user's adaptability to live-line work while wearing the shielding suit.

[0107] Step 208: Compare the fitness of various shielding suits to generate technical optimization information for the shielding suits.

[0108] Step 209: Correct the fitness and use it as a label.

[0109] In actual applications, there are fewer samples of users wearing shielding clothing to perform live-line work, and the performance of the initially trained work detection network may be insufficient.

[0110] Then, in the initial stage of deploying the job detection network, the job detection network can be continuously updated to improve the performance of the job detection network.

[0111] When continuing to update the work detection network, considering that the pairwise loss is more sensitive to noise, the fitness of the user wearing shielding clothing for live work output by the work detection network can be corrected.

[0112] In one correction method, the user may correct the adaptability based on his / her own experience of wearing the shielding suit while performing live-line work.

[0113] In another correction method, experts can correct the fitness based on the user's performance of wearing shielding clothing while performing live-line work.

[0114] When the corrected fitness is received, it is labeled as Label.

[0115] Step 210: Select two sets of posture data to form sample pairs based on the user and the shielding suit.

[0116] Since the user's adaptability to live-line work while wearing a shielding suit is subjective and the evaluation criteria are relatively vague, we can select two groups of posture data that are as close as possible as samples and form sample pairs based on the similarity of the user's feelings and the technical route of the shielding suit for easy comparison.

[0117] In actual applications, multiple attributes of the user and multiple technical parameters of the shielding suit can be queried separately. The user's attributes include place of origin, age, height, weight, years of working with electrical equipment, etc. The technical parameters of the shielding suit are parameters that characterize the technical route, such as the type of each layer of material, the thickness of each layer of material, the size of each joint, etc.

[0118] Use K-means (K-means) and other clustering algorithms to perform unsupervised clustering of users based on multiple attributes to obtain multiple first clusters.

[0119] Clustering algorithms such as K-means are used to perform unsupervised clustering of shielding suits based on multiple technical parameters to obtain multiple second clusters.

[0120] Compare any two sets of posture data with the comparison conditions, and when the two sets of posture data meet the comparison conditions, the two sets of posture data are combined into a sample pair.

[0121] The comparison conditions include live working from users in the same first cluster wearing shielding clothing in the same second cluster, and the values ​​of the labels are not equal (ie, one label is greater than the other label).

[0122] Exemplarily, user A and user B are included in the same first cluster, and shielding suit A and shielding suit B are included in the same second cluster. If the fitness of user A when wearing shielding suit A to perform live working (i.e., label Label) is different from the fitness of user B when wearing shielding suit B to perform live working (i.e., label Label), then the posture data of user A when wearing shielding suit A to perform live working and the posture data of user B when wearing shielding suit B to perform live working can be combined into a sample pair.

[0123] Step 211: construct a pairwise loss value based on the fitness and label of the sample pair.

[0124] In this embodiment, the corrected fitness (i.e., label Label) corresponding to the two sets of posture data in each sample pair is substituted into the pairwise loss function (Pairwise Loss) to obtain the pairwise loss value. For example, the pairwise loss function includes SortNet, RankNet, FFRNet, Margin, and the like.

[0125] The paired loss function requires knowing whether the fitness is good or bad (i.e., the numerical relationship between the fitness), but does not require knowing how much better the fitness is or how few bad errors it has. This is conducive to training the job detection network, forming a reference-free quality evaluation mechanism, and ranking the fitness of users when wearing shielding clothing for live work.

[0126] Step 212: Train the operation detection network based on the paired loss values.

[0127] In this embodiment, fine-tuning and other techniques can be used to continue training the job detection network based on the paired loss values, update the parameters in the job detection network, and continuously improve the performance of the job network.

[0128] It should be noted that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0129] Reference Figure 4 , shows a schematic diagram of a comparison device of a live working shielding suit provided in an embodiment of the present application, which may specifically include the following modules:

[0130] The posture data tracking module 301 is used to track the posture data of the user at various time points when the user wears the shielding suit to perform live work;

[0131] A network loading module 302 is used to load a classifier and a job detection network; the job detection network includes a first backbone network, a second backbone network and an encoder;

[0132] The motion data classification module 303 is used to input the plurality of posture data into the classifier for classification to obtain motion data;

[0133] A matrix conversion module 304, used to convert the posture data into a posture matrix and convert the motion data into a motion matrix;

[0134] A target posture feature extraction module 305, used for inputting the posture matrix into the first backbone network to extract target posture features;

[0135] A target action feature extraction module 306, used for inputting the action matrix into the second backbone network to extract target action features;

[0136] A fitness detection module 307, used for inputting the target posture feature and the target motion feature into the encoder to detect the fitness of the user wearing the shielding suit to perform live-line work;

[0137] The technical optimization information generating module 308 is used to compare the fitness of various shielding suits to generate technical optimization information for the shielding suits.

[0138] In a preferred embodiment of the present application, the classifier includes a gated recurrent unit and a multilayer perceptron;

[0139] The action data classification module 303 is also used for:

[0140] Inputting a plurality of the posture data into the gated recurrent unit to extract posture sequence features;

[0141] The posture sequence features are input into the multi-layer perceptron and mapped into action data.

[0142] In a preferred embodiment of the present application, the posture matrix includes a first sub-matrix, a second sub-matrix and a third sub-matrix;

[0143] The matrix conversion module 304 is also used for:

[0144] Adding a first sliding window, a second sliding window and a third sliding window to the posture data respectively; the width of the first sliding window is smaller than the width of the second sliding window, and the width of the second sliding window is smaller than the width of the third sliding window;

[0145] In the first sliding window, multiplying the transpose of the posture data by the posture data to obtain a first sub-matrix;

[0146] In the second sliding window, multiplying the transpose of the posture data by the posture data to obtain a second sub-matrix;

[0147] In the third sliding window, multiplying the transpose of the posture data by the posture data to obtain a third sub-matrix;

[0148] adding a fourth sliding window to the action data;

[0149] In the fourth sliding window, the transpose of the action data is equal to the action data to obtain an action matrix.

[0150] In a preferred embodiment of the present application, the first backbone network includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer and a fifth convolutional layer;

[0151] The target posture feature extraction module 305 is also used for:

[0152] Inputting the first sub-matrix into the first convolutional layer to extract a first candidate posture feature;

[0153] Inputting the second sub-matrix into the second convolutional layer to extract second candidate posture features;

[0154] Inputting the third sub-matrix into the third convolutional layer to extract the third candidate posture feature;

[0155] splicing the first candidate posture feature and the second candidate posture feature into a fourth candidate posture feature;

[0156] Inputting the fourth candidate posture feature into the fourth convolutional layer to extract the fifth candidate posture feature;

[0157] splicing the fifth candidate posture feature and the third candidate posture feature into a sixth candidate posture feature;

[0158] The sixth candidate posture feature is input into the fifth convolutional layer to extract the target posture feature.

[0159] In a preferred embodiment of the present application, the second backbone network includes a sixth convolutional layer, a seventh convolutional layer, a first fully connected layer, a second fully connected layer and a third fully connected layer;

[0160] The target action feature extraction module 306 is also used for:

[0161] Inputting the action matrix into the sixth convolutional layer to extract the first candidate action feature;

[0162] Inputting the first candidate action feature into the first fully connected layer to map it into a second candidate action feature;

[0163] Inputting the first candidate action feature into the second fully connected layer to map it into a third candidate action feature;

[0164] Inputting the first candidate action feature into the third fully connected layer to map it into a fourth candidate action feature;

[0165] splicing the first candidate action feature, the second candidate action feature, the third candidate action feature and the fourth candidate action feature into a fifth candidate action feature;

[0166] The fifth candidate action feature is input into the seventh convolutional layer to extract the target action feature.

[0167] In a preferred embodiment of the present application, the encoder includes an attention layer, a first depth-separable convolutional layer, a second depth-separable convolutional layer and a fourth fully connected layer;

[0168] The fitness detection module 307 is also used for:

[0169] Inputting the target posture feature and the target action feature into the attention layer to fuse them into a first multimodal feature;

[0170] Inputting the first multimodal feature into the first depthwise separable convolutional layer to extract a second multimodal feature;

[0171] splicing the first multimodal feature and the second multimodal feature into a third multimodal feature;

[0172] Inputting the third multimodal feature into the second depthwise separable convolutional layer to extract a fourth multimodal feature;

[0173] splicing the first multimodal feature and the fourth multimodal feature into a fifth multimodal feature;

[0174] Inputting the fifth multimodal feature into the fourth fully connected layer to map it into a sixth multimodal feature;

[0175] The sixth multimodal feature is activated to obtain the adaptability of the user to perform live-line work while wearing the shielding suit.

[0176] In a preferred embodiment of the present application, the device further comprises:

[0177] A label marking module, used for correcting the fitness as a label;

[0178] A sample pair composition module, used for selecting two groups of posture data to form a sample pair according to the user and the shielding suit;

[0179] A paired loss value calculation module, used to construct a paired loss value according to the fitness of the sample pair and the label;

[0180] The operation detection network training module is used to train the operation detection network according to the paired loss values.

[0181] In a preferred embodiment of the present application, the sample pair composition module is further used for:

[0182] respectively querying a plurality of attributes of the user and a plurality of technical parameters of the shielding suit;

[0183] Performing unsupervised clustering on the users according to the multiple attributes to obtain multiple first clusters;

[0184] Performing unsupervised clustering on the shielding suit according to the plurality of technical parameters to obtain a plurality of second clusters;

[0185] When the two groups of posture data meet the comparison condition, the two groups of posture data are combined into a sample pair;

[0186] The comparison condition includes live working performed by the users in the same first cluster wearing the shielding suits in the same second cluster, and the values ​​of the tags are not equal.

[0187] An embodiment of the present application provides a comparison device for live working shielding clothing. By using the device, each step in the aforementioned method embodiments can be implemented.

[0188] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiment part.

[0189] Reference Figure 5 , shows a schematic diagram of a terminal device provided in an embodiment of the present application. Figure 5 As shown, the terminal device 400 in the embodiment of the present application includes: a processor 410, a memory 420, and a computer program 421 stored in the memory 420 and executable on the processor 410. When the processor 410 executes the computer program 421, the steps in each embodiment of the comparison method of live working shielding clothing are implemented. Alternatively, when the processor 410 executes the computer program 421, the functions of each module / unit in the above-mentioned device embodiments are implemented.

[0190] Exemplarily, the computer program 421 may be divided into one or more modules / units, which are stored in the memory 420 and executed by the processor 410 to complete the present application. The one or more modules / units may be a series of computer program instruction segments capable of completing specific functions, which may be used to describe the execution process of the computer program 421 in the terminal device 400.

[0191] The terminal device 400 may include, but is not limited to, a processor 410 and a memory 420. Those skilled in the art will appreciate that Figure 5 It is only an example of the terminal device 400 and does not constitute a limitation on the terminal device 400. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the terminal device 400 may also include input and output devices, network access devices, buses, etc.

[0192] The processor 410 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0193] The memory 420 may be an internal storage unit of the terminal device 400, such as a hard disk or memory of the terminal device 400. The memory 420 may also be an external storage device of the terminal device 400, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the terminal device 400. Further, the memory 420 may also include both an internal storage unit of the terminal device 400 and an external storage device. The memory 420 is used to store the computer program 421 and other programs and data required by the terminal device 400. The memory 420 may also be used to temporarily store data that has been output or is to be output.

[0194] An embodiment of the present application also discloses a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the comparison method of live working shielding clothing as described in the aforementioned embodiments is implemented.

[0195] The embodiment of the present application further discloses a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the comparison method of live working shielding clothing as described in the above-mentioned embodiments is implemented.

[0196] The embodiment of the present application further discloses a computer program product. When the computer program product is run on a computer, the computer is enabled to execute the comparison method of live working shielding clothing described in the above-mentioned embodiments.

[0197] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application is described in detail with reference to the above-mentioned embodiments, a person skilled in the art should understand that the technical solutions described in the above-mentioned embodiments can still be modified, or some of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A comparison method for live working shielding clothing, characterized in that: include: Tracking posture data of the user at various time points when the user is wearing a shielding suit and performing live work; Loading a classifier and a job detection network; the job detection network includes a first backbone network, a second backbone network and an encoder; Inputting a plurality of the posture data into the classifier for classification to obtain action data; Respectively converting the posture data into a posture matrix and converting the motion data into a motion matrix; Inputting the posture matrix into the first backbone network to extract target posture features; Inputting the action matrix into the second backbone network to extract target action features; Inputting the target posture feature and the target motion feature into the encoder to detect the adaptability of the user to wear the shielding suit for live-line work; The fitness of various shielding suits is compared to generate technical optimization information for the shielding suit.

2. The method according to claim 1, characterized in that The classifier includes a gated recurrent unit and a multi-layer perceptron; The step of inputting the plurality of posture data into the classifier for classification to obtain the action data comprises: Inputting a plurality of the posture data into the gated recurrent unit to extract posture sequence features; The posture sequence features are input into the multi-layer perceptron and mapped into action data.

3. The method according to claim 1, characterized in that The posture matrix includes a first sub-matrix, a second sub-matrix and a third sub-matrix; The step of converting the posture data into a posture matrix and converting the motion data into a motion matrix respectively comprises: Adding a first sliding window, a second sliding window and a third sliding window to the posture data respectively; the width of the first sliding window is smaller than the width of the second sliding window, and the width of the second sliding window is smaller than the width of the third sliding window; In the first sliding window, multiplying the transpose of the posture data by the posture data to obtain a first sub-matrix; In the second sliding window, multiplying the transpose of the posture data by the posture data to obtain a second sub-matrix; In the third sliding window, multiplying the transpose of the posture data by the posture data to obtain a third sub-matrix; adding a fourth sliding window to the action data; In the fourth sliding window, the transpose of the action data is equal to the action data to obtain an action matrix.

4. The method according to claim 3, characterized in that The first backbone network includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer and a fifth convolutional layer; The step of inputting the posture matrix into the first backbone network to extract the target posture feature comprises: Inputting the first sub-matrix into the first convolutional layer to extract a first candidate posture feature; Inputting the second sub-matrix into the second convolutional layer to extract second candidate posture features; Inputting the third sub-matrix into the third convolutional layer to extract the third candidate posture feature; splicing the first candidate posture feature and the second candidate posture feature into a fourth candidate posture feature; Inputting the fourth candidate posture feature into the fourth convolutional layer to extract the fifth candidate posture feature; splicing the fifth candidate posture feature and the third candidate posture feature into a sixth candidate posture feature; The sixth candidate posture feature is input into the fifth convolutional layer to extract the target posture feature.

5. The method according to claim 4, characterized in that The second backbone network includes a sixth convolutional layer, a seventh convolutional layer, a first fully connected layer, a second fully connected layer and a third fully connected layer; The step of inputting the action matrix into the second backbone network to extract target action features comprises: Inputting the action matrix into the sixth convolutional layer to extract the first candidate action feature; Inputting the first candidate action feature into the first fully connected layer to map it into a second candidate action feature; Inputting the first candidate action feature into the second fully connected layer to map it into a third candidate action feature; Inputting the first candidate action feature into the third fully connected layer to map it into a fourth candidate action feature; splicing the first candidate action feature, the second candidate action feature, the third candidate action feature and the fourth candidate action feature into a fifth candidate action feature; The fifth candidate action feature is input into the seventh convolutional layer to extract the target action feature.

6. The method according to claim 5, characterized in that The encoder includes an attention layer, a first depth-separable convolutional layer, a second depth-separable convolutional layer and a fourth fully connected layer; The step of inputting the target posture feature and the target motion feature into the encoder to detect the adaptability of the user wearing the shielding suit to perform live-line work includes: Inputting the target posture feature and the target action feature into the attention layer to fuse them into a first multimodal feature; Inputting the first multimodal feature into the first depthwise separable convolutional layer to extract a second multimodal feature; splicing the first multimodal feature and the second multimodal feature into a third multimodal feature; Inputting the third multimodal feature into the second depthwise separable convolutional layer to extract a fourth multimodal feature; splicing the first multimodal feature and the fourth multimodal feature into a fifth multimodal feature; Inputting the fifth multimodal feature into the fourth fully connected layer to map it into a sixth multimodal feature; The sixth multimodal feature is activated to obtain the adaptability of the user to perform live-line work while wearing the shielding suit.

7. The method according to any one of claims 1 to 6, characterized in that Also includes: The fitness is modified as a label; Selecting two groups of the posture data to form a sample pair according to the user and the shielding suit; Constructing a pairwise loss value according to the fitness of the sample pair and the label; The operation detection network is trained according to the pairwise loss values.

8. The method according to claim 7, characterized in that The selecting two groups of the posture data to form a sample pair according to the user and the shielding suit includes: respectively querying a plurality of attributes of the user and a plurality of technical parameters of the shielding suit; Performing unsupervised clustering on the users according to the multiple attributes to obtain multiple first clusters; Performing unsupervised clustering on the shielding suit according to the plurality of technical parameters to obtain a plurality of second clusters; When the two groups of posture data meet the comparison condition, the two groups of posture data are combined into a sample pair; The comparison condition includes live working performed by the users in the same first cluster wearing the shielding suits in the same second cluster, and the values ​​of the tags are not equal.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the comparison method of the live working shielding clothing as described in any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the comparison method of live working shielding clothing as described in any one of claims 1 to 8 is implemented.

Images