An industrial production behavior recognition method based on pose DSMWC

Abstract

The application discloses an industrial production behavior recognition method based on PoseDSMWC, comprising the following steps: S1, collecting complete operation video data of workers on a production line, extracting time series of human body skeleton key point coordinates and confidence, and constructing a labeled industrial production behavior skeleton key point sequence dataset; S2, constructing a lightweight PoseDSMWC network model, and configuring an I3Dhead classification head at the end of the network; S3, generating a four-dimensional heat map by using the constructed skeleton sequence dataset, and performing supervised training on the PoseDSMWC network model; S4, extracting a skeleton key point sequence from a production operation video to be recognized in real time, generating a four-dimensional heat map, inputting the trained PoseDSMWC network model, outputting an action category and confidence through the I3Dhead classification head, and completing the recognition. The application reduces the model parameter quantity and the calculation complexity through a depth separable convolution, enhances the feature expression capability by combining a multi-dimensional weighted collaborative attention mechanism, and realizes the lightweight, high-precision and real-time recognition of industrial production behaviors.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, specifically relating to an industrial production behavior recognition method based on PoseDSMWC. Background Technology

[0002] Industrial production behavior recognition, as a core component of intelligent manufacturing systems, aims to automatically monitor and identify the operational behaviors of workers on the production line through computer vision and deep learning technologies. This ensures that the production process meets the requirements of standard operating procedures (SOPs), thereby improving production efficiency, ensuring product quality, and protecting worker safety. Traditional industrial behavior monitoring mainly relies on manual inspection, which is not only inefficient and costly, but also prone to missed or false detections due to human fatigue or distraction, making it difficult to meet the needs of large-scale, real-time industrial monitoring.

[0003] With the widespread application of deep learning technology in computer vision, video-based behavior recognition methods have gradually become a research hotspot in industrial monitoring. Existing methods are mainly divided into two categories: one is based on two-dimensional convolutional neural networks (2D CNNs) using video frames, which achieves temporal modeling through frame-by-frame processing or stacking of time dimensions, but struggles to effectively capture the temporal dynamics of actions; the other is based on three-dimensional convolutional neural networks (3D CNNs), which simultaneously model spatial and temporal features through three-dimensional convolutional kernels, demonstrating superior performance in behavior recognition tasks. Among these, PoseC3D, a behavior recognition method based on skeletal point sequences, encodes key points of the human skeleton into heatmaps and utilizes a 3D convolutional network for end-to-end learning, achieving excellent performance on multiple public datasets.

[0004] However, existing skeleton-based behavior recognition methods still face many challenges in industrial production scenarios. First, the industrial production environment is complex and variable, with challenging conditions such as changing lighting, occlusion interference, and multi-person collaboration, placing higher demands on the accuracy and robustness of skeleton point extraction. Second, industrial production lines have high real-time requirements, necessitating algorithms that maintain recognition accuracy while possessing low inference latency. Third, industrial production behaviors are highly specialized and temporal; subtle differences in actions can lead to product quality issues, requiring algorithms with fine-grained behavior recognition capabilities. Summary of the Invention

[0005] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and propose an industrial production behavior recognition method based on PoseDSMWC.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] An industrial production behavior recognition method based on PoseDSMWC includes the following steps:

[0008] S1. Collect complete operation video data of production line workers, divide the video into temporal actions according to standard operating procedures, and use a pose estimation network to extract the time series of human skeleton key point coordinates and confidence, and construct a labeled industrial production behavior skeleton key point sequence dataset.

[0009] S2. Using PoseC3D as the baseline, replace Bottleneck in Backbone with Bottleneck-DS based on depthwise separable convolution. At the same time, embed a multidimensional weighted collaborative attention module in each Bottleneck module to build a lightweight PoseDSMWC network model, and configure the I3Dhead classification head at the end of the network.

[0010] S3. Use the skeletal sequence dataset constructed in step S1 to generate a four-dimensional heatmap, and perform supervised training on the constructed PoseDSMWC network model to obtain the trained PoseDSMWC network model.

[0011] S4. Extract the skeletal key point sequence from the production operation video to be identified in real time and generate a four-dimensional heat map. Input the sequence into the trained PoseDSMWC network model and output the action category and confidence score through the I3Dhead classification head to complete the recognition.

[0012] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0013] 1. This invention significantly reduces the number of model parameters and computational complexity by replacing standard 3D convolution with depthwise separable convolution, thus meeting the real-time requirements of industrial production scenarios. The embedded multidimensional weighted collaborative attention module can simultaneously enhance features in the three dimensions of channel, time and space, effectively capturing fine-grained differences in industrial production behavior and improving the ability to recognize fine-grained behaviors.

[0014] 2. This invention uses heatmap encoding to represent skeletal point sequences, preserving the spatial location information and temporal continuity of skeletal points, which is more robust than the direct input of coordinates. Based on FineGym pre-trained weights, fine-tuning can make full use of the prior knowledge of large-scale skeletal point behavior datasets, improving the convergence speed and final performance of the model on industrial production datasets.

[0015] 3. This invention effectively alleviates the class imbalance problem in industrial production behavior datasets by using a Cosine Annealing learning rate scheduling strategy and a Focal Loss loss function, thereby improving the recognition accuracy of minority class behaviors. Attached Figure Description

[0016] Figure 1This is a flowchart of the method of the present invention.

[0017] Figure 2 This is a schematic diagram of the structure of the PoseDSMWC network model in the embodiment.

[0018] Figure 3 This is a schematic diagram of the original Bottleneck and the improved Bottleneck-DS in the embodiments.

[0019] Figure 4 This is a schematic diagram of the multidimensional weighted collaborative attention module in the embodiment.

[0020] Figure 5 This is a schematic diagram of the Bottleneck-DSMWC module and the Bottleneck-MWC module in the embodiment. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0022] Explanation of relevant terms:

[0023] PoseC3D: A 3D convolutional behavior recognition network based on human pose heatmaps.

[0024] DSC: Depthwise Separable Convolution.

[0025] Bottleneck module: A residual network unit that uses a three-layer convolutional structure in the residual network, which first reduces dimensionality, then extracts features, and finally increases dimensionality.

[0026] MWC: Multi-dimensional Weighted Collaboration.

[0027] Cosine Annealing learning rate scheduling strategy: A dynamic strategy that achieves smooth decay of the learning rate according to the cosine function curve.

[0028] Examples; such as Figure 1 As shown, an industrial production behavior recognition method based on PoseDSMWC includes the following steps:

[0029] S1. Collect complete operation video data of production line workers, divide the video into temporal actions according to standard operating procedures, and use a pose estimation network to extract the time series of human skeleton key point coordinates and confidence, and construct a labeled industrial production behavior skeleton key point sequence dataset.

[0030] In this embodiment, complete video data of the production line workers' operations is collected. An industrial-grade camera is used for video acquisition, installed above and to the side of the production line to ensure clear capture of the workers' upper body movements. The video resolution is 1920×1080 pixels, and the frame rate is 30fps. To cover different production scenarios and individual worker differences, the data collection process is conducted across multiple production line shifts, including day and night shifts, to obtain data under varying lighting conditions. Simultaneously, workers of different heights and body types are recruited to participate in data collection, enhancing the diversity of the dataset.

[0031] After video data acquisition is completed, the video is divided into time-series actions according to the standard operating procedure (SOP), specifically as follows:

[0032] The annotation is completed independently by two or more professional annotators. They manually watch the video and annotate the start and end time frames of each action. For samples with inconsistent annotations, the final annotation result is determined through discussion and negotiation. The duration of each action segment is determined according to the specific operation type, generally between 0.5 seconds and 5 seconds. After the annotation is completed, the annotation data is saved in a standard format. Each record includes the video file name, action category label, start frame number, and end frame number.

[0033] The time series of human skeleton keypoint coordinates and confidence scores are extracted using a pose estimation network, specifically as follows:

[0034] First, an object detection network (such as YOLOv5) is used to detect the workers in the video frames and obtain the bounding box ROI of the workers;

[0035] The ROI is input into a pose estimation network (such as HRNet or ViTPose) to obtain the coordinates and confidence of 17 COCO keypoints; the pose estimation network uses weights pre-trained on the COCO dataset and fine-tunes them on industrial scene data to improve the extraction accuracy in specific production environments.

[0036] The 17 COCO key points include: nose, left eye, right eye, left ear, right ear, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left hip, right hip, left knee, right knee, left ankle, and right ankle.

[0037] S2. Using PoseC3D as the baseline, replace Bottleneck in Backbone with Bottleneck-DS based on depthwise separable convolution. At the same time, embed a multidimensional weighted collaborative attention module in each Bottleneck module to build a lightweight PoseDSMWC network model, and configure the I3Dhead classification head at the end of the network.

[0038] In this embodiment, the first 1×1×1 convolution in Bottleneck is changed to a 3×1×1 convolution, and the original 3×3×3 convolution is changed to a spatial-temporal depthwise convolution. Finally, a linear projection along the C dimension is performed using a 1×1×1 pointwise convolution kernel to complete the cross-channel information integration; Figure 3 The image shows a schematic diagram of the original Bottleneck and the improved Bottleneck-DS.

[0039] Depthwise separable convolution, for input tensor First, a 3×1×1 convolutional layer is used to perform preliminary spatial-temporal feature extraction and channel dimension adjustment on the input. Then, the channel dimension C and the temporal dimension T are merged into C×T, and 3×3 grouped convolutions are performed, with the number of groups equal to C×T (i.e., each group contains only a single channel), completing the spatial-temporal convolution on the H×W plane. After convolution, the merged dimensions are restored to C and T. Finally, cross-channel information is integrated through 1×1×1 pointwise convolutions. Compared with standard 3D convolution, depthwise separable convolution reduces the number of parameters by about 8-9 times, significantly reducing the computational cost.

[0040] In this embodiment, the multidimensional weighted collaborative attention module includes a channel-temporal collaborative attention branch, a spatial attention branch, and an adaptive fusion unit;

[0041] The multidimensional weighted collaborative attention module is embedded inside the Bottleneck module and is located after the depthwise separable convolution output and before being added to the skip layer residual. It performs channel-temporal-spatial triple enhancement on the intermediate feature map before participating in residual fusion.

[0042] Among them, the channel-temporal collaborative attention branch, for input features Global average pooling is performed on both the channel and time dimensions to extract global feature representations at the channel and time levels, resulting in channel-level features (dimension C) and time-level features (dimension T).

[0043] Channel-level features are sequentially passed through two fully connected layers and a ReLU activation function, and finally channel attention weights are generated using a sigmoid function. Temporal features are sequentially passed through two fully connected layers and a ReLU activation function, and finally, temporal attention weights are generated using a Sigmoid function. ;

[0044] The channel attention weights and temporal attention weights are multiplied together using a broadcast mechanism to obtain the channel-temporal collaborative attention weights. , is represented as:

[0045]

[0046] Spatial attention branch, for input features Processed independently at each time step; global average pooling and global max pooling are performed along the channel axis respectively to generate two H×W spatial description maps; the two spatial description maps are concatenated along the channel dimension and input into a 7×7 convolutional layer, and the spatial attention weight Ms is generated by the Sigmoid activation function, as follows:

[0047]

[0048] Adaptive fusion unit, through learnable parameter vectors Channel-time co-attention weights Spatial attention weights Perform weighted fusion, fusion weight From the learnable parameter vector The final output is the fused enhanced feature map obtained through Softmax normalization:

[0049]

[0050] in, .

[0051] like Figure 4 The diagram shown is a schematic of the multidimensional weighted collaborative attention module.

[0052] In this embodiment, the backbone of the PoseDSMWC network model adopts a four-level progressive architecture; the first layer is a 3D convolutional stem, which upscales the K-channel heatmap to 32 channels; then, three stages are deployed sequentially, with each stage stacking several Bottleneck-DSMWC modules, resulting in output channels of 128, 256, and 512 respectively, achieving progressive scaling of spatiotemporal and channel-scale features; for example... Figure 2 The diagram shown is a schematic representation of the PoseDSMWC network model.

[0053] The four-level progressive architecture is as follows:

[0054] Stem layer: contains a 3D convolutional layer and a max pooling layer; the 3D convolutional layer upscales the input K×T×H×W four-dimensional heatmap (K=17, T=32, H=W=64) to 32 channels, with a kernel size of 1×7×7, a stride of 1×1×1, padding of 0×3×3, and an activation function of ReLU; the max pooling layer has a kernel size of 1×3×3, a stride of 1×1×1, and padding of 0×1×1.

[0055] Phase 1: Contains 4 Bottleneck-DSMWC modules. Each module first reduces the dimensionality to 32 channels through 1×1×1 convolutions, then extracts spatiotemporal features through 3×3 depthwise separable convolutions, and finally restores the channel count to 128 through 1×1×1 convolutions. A multidimensional weighted collaborative attention module is then embedded after the depthwise separable convolutions. The output channel count is 128. Spatial downsampling in this phase is achieved by setting a stride of 2×2 in the 3×3 convolution of the first Bottleneck module.

[0056] Phase 2: Contains 6 Bottleneck-DSMWC modules with 256 output channels; temporal downsampling in this phase is achieved by setting a max pooling layer with a kernel size of 2×1×1 and a stride of 2×1×1 before the first Bottleneck, and spatial downsampling is achieved by setting a stride of 2×2 in the 3×3 convolution of the first Bottleneck module.

[0057] Phase 3: Contains one Bottleneck-MWC module and two Bottleneck-DSMWC modules, with a total of 512 output channels; the spatiotemporal downsampling in this phase is achieved by setting a stride of 2×2×2 in the 1×3×3 convolution of the first Bottleneck-MWC module.

[0058] All Bottleneck modules employ a residual connection structure, where input and output features are added through skip-layer connections. Specifically, the Bottleneck-DSMWC module is a residual network unit obtained by embedding a multi-dimensional weighted collaborative attention module after the depthwise separable convolutional output of the Bottleneck-DS module and before adding it to the skip-layer residual; the Bottleneck-MWC module is a residual network unit obtained by embedding a multi-dimensional weighted collaborative attention module after the up-dimensional output of the original Bottleneck module and before adding it to the skip-layer residual; for example... Figure 5 The diagram shown is a schematic of the Bottleneck-DSMWC module and the Bottleneck-MWC module.

[0059] The I3Dhead classification head is located at the end of the backbone and receives the high-dimensional feature map output from the last stage. The I3Dhead classification head includes a global average pooling layer, a Dropout layer, and a fully connected layer connected in sequence.

[0060] The global average pooling layer performs global average pooling on the input feature map in both the temporal and spatial dimensions, and outputs a global feature vector, thereby compressing spatiotemporal information and eliminating position-sensitive bias.

[0061] The Dropout layer randomly discards some neurons at a preset dropout rate to enhance the model's generalization ability;

[0062] Fully connected layers linearly map global feature vectors to the behavior category space, thus obtaining the categories. N represents the number of industrial production behavior categories; subsequently, the probability distribution of each behavior category is output after Softmax normalization. The final identification result is the behavior category with the highest probability.

[0063] S3. Using the skeletal sequence dataset constructed in step S1, generate a four-dimensional heatmap and perform supervised training on the constructed PoseDSMWC network model to obtain the trained PoseDSMWC network model.

[0064] In this embodiment, the method for generating the four-dimensional heatmap is as follows:

[0065] Based on the skeleton point coordinates, normalized coordinates are mapped to a fixed resolution 64×64 grid for each frame;

[0066] Visible keypoints with a confidence level greater than 0.1 A single-frame heatmap is generated using a two-dimensional Gaussian kernel, with kernel size k = 2σ + 1, σ = 1.5, and a peak value of 1.

[0067]

[0068] For invisible points with a confidence level less than 0.1, maintain an all-zero heatmap;

[0069] The frames are stacked along the time dimension T=32 to form a four-dimensional tensor of 17×32×64×64;

[0070] All heatmaps are normalized to ensure that the sum of pixels in the visible area of ​​a single frame is 1, thus guaranteeing the stability of the input values.

[0071] In this embodiment, the training of the PoseDSMWC network model uses Focal Loss as the optimization objective (loss function); the formula for calculating Focal Loss is:

[0072]

[0073] in, Output the probability of each behavior category for the model, where γ is the focusing parameter (set to 2). A weighting balancing factor related to the category;

[0074] The optimizer uses stochastic gradient descent (SGD) with momentum, and the specific hyperparameters are set as follows: initial learning rate 0.2, momentum 0.9, and weight decay 0.0003.

[0075] To prevent gradient explosion, gradient clipping is enabled, and the upper limit of the L2 norm is set to 40. The learning rate is scheduled using a Cosine Annealing strategy, where the learning rate is annealed from its initial value to 0 using a cosine curve, and updated according to the iteration step size. The update formula is as follows:

[0076]

[0077] in, Let be the learning rate in round t, and T be the total number of training steps. The initial maximum learning rate is set to 0.2. The minimum learning rate is set (to 0).

[0078] The model weights were initialized using PoseC3D weights pre-trained on the FineGym dataset. Fine-tuning was performed on a constructed labeled dataset of skeletal keypoint sequences of industrial production behavior. Data augmentation strategies were employed during training, including random horizontal flipping, random scaling, and aspect ratio changes.

[0079] S4. Extract the skeletal key point sequence from the production operation video to be identified in real time and generate a four-dimensional heat map. Input the sequence into the trained PoseDSMWC network model and output the action category and confidence score through the I3Dhead classification head to complete the recognition.

[0080] In this embodiment, the trained PoseDSMWC network model is deployed in an industrial production environment for real-time behavior recognition. The video stream of the production operation to be recognized first passes through an object detection network to extract the worker's ROI, and then through a pose estimation network to obtain the skeletal keypoint sequence. The skeletal keypoint sequence is converted into a four-dimensional heatmap in real time and input into the trained PoseDSMWC network model. After forward inference, the output layer uses a softmax function to give the probability distribution of each behavior category, and the final recognition result is the behavior category with the highest probability.

[0081] It should also be noted that, in this specification, terms such as "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0082] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for recognizing industrial production behavior based on PoseDSMWC, characterized in that, Includes the following steps: S1. Collect complete operation video data of production line workers, divide the video into temporal actions according to standard operating procedures, and use a pose estimation network to extract the time series of human skeleton key point coordinates and confidence, and construct a labeled industrial production behavior skeleton key point sequence dataset. S2. Using PoseC3D as the baseline, replace Bottleneck in Backbone with Bottleneck-DS based on depthwise separable convolution. At the same time, embed a multidimensional weighted collaborative attention module in each Bottleneck module to build a lightweight PoseDSMWC network model, and configure the I3Dhead classification head at the end of the network. S3. Use the skeletal sequence dataset constructed in step S1 to generate a four-dimensional heatmap, and perform supervised training on the constructed PoseDSMWC network model to obtain the trained PoseDSMWC network model. S4. Extract the skeletal key point sequence from the production operation video to be identified in real time and generate a four-dimensional heat map. Input the sequence into the trained PoseDSMWC network model and output the action category and confidence score through the I3Dhead classification head to complete the recognition.

2. The industrial production behavior recognition method based on PoseDSMWC according to claim 1, characterized in that, In step S1, the video is divided into time-series actions according to the standard operating procedure, specifically as follows: Two or more professional annotators manually watch the video and annotate the start and end time frames of each action; for samples with inconsistent annotations, the final annotation result is determined through discussion and negotiation; the duration of each action segment is determined according to the specific operation type; after the annotation is completed, the annotation data is saved in a standard format, and each record includes the video file name, action category label, start frame number and end frame number.

3. The industrial production behavior recognition method based on PoseDSMWC according to claim 1, characterized in that, In step S1, the time series of human skeleton keypoint coordinates and confidence scores are extracted using a pose estimation network, specifically as follows: First, an object detection network is used to detect the workers in the video frames and obtain the bounding boxes (ROIs) of the workers. The ROI is input into the pose estimation network to obtain the coordinates and confidence scores of 17 COCO keypoints. The pose estimation network uses weights pre-trained on the COCO dataset and is fine-tuned on industrial scene data to improve the extraction accuracy in specific production environments. The 17 COCO key points include: nose, left eye, right eye, left ear, right ear, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left hip, right hip, left knee, right knee, left ankle, and right ankle.

4. The industrial production behavior recognition method based on PoseDSMWC according to claim 1, characterized in that, In step S2, Bottleneck in Backbone is replaced with Bottleneck-DS based on depthwise separable convolution, specifically as follows: The first 1×1×1 convolution in Bottleneck is changed to a 3×1×1 convolution, the original 3×3×3 convolution is changed to a spatial-temporal depth convolution, and finally a linear projection along the C dimension is made through a 1×1×1 pointwise convolution kernel to complete the cross-channel information integration. Bottleneck-DS, based on depthwise separable convolution, for input tensors First, a 3×1×1 convolutional layer is used to perform preliminary spatial-temporal feature extraction and channel dimension adjustment on the input. Then, the channel dimension C and the time dimension T are merged into C×T, and 3×3 grouped convolution is performed, with the number of groups equal to C×T, that is, each group contains only a single channel, to complete the spatial-temporal convolution on the H×W plane. After the convolution is completed, the merged dimensions are restored to C and T. Finally, cross-channel information is integrated through 1×1×1 pointwise convolution.

5. The industrial production behavior recognition method based on PoseDSMWC according to claim 4, characterized in that, In step S2, the multidimensional weighted collaborative attention module includes a channel-temporal collaborative attention branch, a spatial attention branch, and an adaptive fusion unit; The multidimensional weighted collaborative attention module is embedded inside the Bottleneck module and is located after the depthwise separable convolution output and before being added to the skip layer residual. It performs channel-temporal-spatial triple enhancement on the intermediate feature map before participating in residual fusion.

6. The industrial production behavior recognition method based on PoseDSMWC according to claim 5, characterized in that, Channel-time co-attention branch includes the following steps: Global average pooling is performed on the input features in both the channel and time dimensions to extract global feature representations at the channel and time levels, resulting in channel-level features and time-level features. Channel-level features are sequentially passed through two fully connected layers and a ReLU activation function, and finally channel attention weights are generated using a sigmoid function. Temporal features are sequentially passed through two fully connected layers and a ReLU activation function, and finally, temporal attention weights are generated using a Sigmoid function. ; The channel attention weights and temporal attention weights are multiplied together using a broadcast mechanism to obtain the channel-temporal collaborative attention weights. , is represented as: The spatial attention branch processes the input features independently at each time step; it performs global average pooling and global max pooling along the channel axis to generate two spatial description maps; these two spatial description maps are concatenated along the channel dimension and input into a 7×7 convolutional layer, where the spatial attention weights Ms are generated by the Sigmoid activation function, as follows: Adaptive fusion unit, through learnable parameter vectors Channel-time co-attention weights Spatial attention weights Perform weighted fusion, fusion weight From the learnable parameter vector The final output is the fused enhanced feature map obtained through Softmax normalization: in, .

7. The industrial production behavior recognition method based on PoseDSMWC according to claim 6, characterized in that, The backbone of the PoseDSMWC network model adopts a four-level progressive architecture. The first layer is a 3D convolutional stem that upscales the K-channel heatmap to 32 channels. Then, three stages are deployed in sequence, with several Bottleneck-DSMWC modules stacked in each stage. The number of output channels is 128, 256, and 512, respectively, realizing the progressive expansion of spatiotemporal-channel multi-scale features. The four-level progressive architecture is as follows: Stem layer: contains a 3D convolutional layer and a max pooling layer; the 3D convolutional layer upscales the input K×T×H×W four-dimensional heatmap to 32 channels, with a kernel size of 1×7×7, a stride of 1×1×1, padding of 0×3×3, and an activation function of ReLU; the max pooling layer has a kernel size of 1×3×3, a stride of 1×1×1, and padding of 0×1×1. Phase 1: Contains 4 Bottleneck-DSMWC modules. Each module first reduces the dimensionality to 32 channels through 1×1×1 convolutions, then extracts spatiotemporal features through 3×3 depthwise separable convolutions, and finally restores the channel count to 128 through 1×1×1 convolutions. A multidimensional weighted collaborative attention module is then embedded after the depthwise separable convolutions. The output channel count is 128. Spatial downsampling in this phase is achieved by setting a stride of 2×2 in the 3×3 convolution of the first Bottleneck module. Phase 2: Contains 6 Bottleneck-DSMWC modules with 256 output channels; temporal downsampling in this phase is achieved by setting a max pooling layer with a kernel size of 2×1×1 and a stride of 2×1×1 before the first Bottleneck, and spatial downsampling is achieved by setting a stride of 2×2 in the 3×3 convolution of the first Bottleneck module. Phase 3: Contains one Bottleneck-MWC module and two Bottleneck-DSMWC modules, with a total of 512 output channels; the spatiotemporal downsampling in this phase is achieved by setting a stride of 2×2×2 in the 1×3×3 convolution of the first Bottleneck-MWC module. All Bottleneck modules employ a residual connection structure, where input and output features are added through skip connections. The Bottleneck-DSMWC module is a residual network unit obtained by embedding a multidimensional weighted collaborative attention module after the depthwise separable convolutional output of the Bottleneck-DS module and before adding it to the skip residual. The Bottleneck-MWC module is a residual network unit obtained by embedding a multidimensional weighted collaborative attention module after the up-dimensional output of the original Bottleneck module and before adding it to the skip residual.

8. The industrial production behavior recognition method based on PoseDSMWC according to claim 7, characterized in that, The I3Dhead classification head is located at the end of the backbone and receives the high-dimensional feature map output from the last stage. The I3Dhead classification head includes a global average pooling layer, a Dropout layer, and a fully connected layer connected in sequence. The global average pooling layer performs global average pooling on the input feature map in both the temporal and spatial dimensions, and outputs a global feature vector, thereby compressing spatiotemporal information and eliminating position-sensitive bias. The Dropout layer randomly discards some neurons at a preset dropout rate to enhance the model's generalization ability; Fully connected layers linearly map global feature vectors to the behavior category space, thus obtaining the categories. N represents the number of industrial production behavior categories; subsequently, the probability distribution of each behavior category is output after Softmax normalization. The final identification result is the behavior category with the highest probability.

9. The industrial production behavior recognition method based on PoseDSMWC according to claim 1, characterized in that, In step S3, the training of the PoseDSMWC network model uses Focal Loss as the optimization objective; the formula for calculating Focal Loss is: in, The model outputs the probability of each behavior category, where γ is the focusing parameter. A weighting balancing factor related to the category; The optimizer employs stochastic gradient descent with momentum and enables gradient clipping; the learning rate scheduling uses a cosine annealing strategy, where the learning rate is cosine-annealed from its initial value to 0, and updated according to the iteration step size. The update formula is as follows: in, Let be the learning rate in round t, and T be the total number of training steps. The initial maximum learning rate, The minimum learning rate set; The model weights were initialized using PoseC3D weights pre-trained on the FineGym dataset. Fine-tuning was performed on a constructed labeled dataset of skeletal keypoint sequences of industrial production behavior. Data augmentation strategies were employed during training, including random horizontal flipping, random scaling, and aspect ratio changes.

10. The industrial production behavior recognition method based on PoseDSMWC according to claim 1, characterized in that, The method for generating a four-dimensional heatmap is as follows: Based on the skeleton point coordinates, normalized coordinates are mapped to a grid with a fixed resolution of H×W for each frame; Visible keypoints with a confidence level greater than 0.1 A single-frame heatmap is generated using a two-dimensional Gaussian kernel, with kernel size k = 2σ + 1, σ = 1.5, and a peak value of 1. For invisible points with a confidence level less than 0.1, maintain an all-zero heatmap; The frames are stacked along the time dimension T to form a four-dimensional tensor of K×T×H×W, where K is the number of joints and T is the temporal window. All heatmaps are normalized so that the sum of pixels in the visible area of ​​a single frame is 1, ensuring the stability of the input values.

Images