Feature enhancement-based industrial internet of things semantic data transmission method and system

Abstract

The application discloses a feature enhancement-based industrial Internet of Things semantic data transmission method and system. Through the structural design of an encoder and a decoder, the encoder model deployed in the industrial Internet of Things node extracts semantic information, and the decoder model deployed in the edge node restores data; and under different signal-to-noise ratio conditions, the reconstruction capability is adaptively allocated to the key semantic area, so that the recovery accuracy and structural consistency are maintained in a complex wireless environment; through feature enhancement training, the industrial Internet of Things semantic data transmission under a small sample is realized. The application effectively reduces the exposure of industrial data, reduces data redundancy, and improves transmission efficiency while efficiently transmitting and extracting semantic information.

Description

Technical Field

[0001] This invention belongs to the field of semantic data transmission in the Industrial Internet of Things (IIoT), and relates to a semantic data transmission system and method for industrial environments based on feature enhancement. Background Technology

[0002] Industrial data is collected in industrial settings by deploying automated equipment such as sensors and robotic arms. Traditional industrial data transmission transmits raw collected data without differentiation, which consumes a lot of energy and also causes a lot of data redundancy. When it is necessary to extract valuable information from massive amounts of industrial data, a lot of redundant data received from the receiving end must be analyzed, which consumes a lot of manpower and resources.

[0003] Semantic communication extracts feature values ​​from raw data and transmits the semantics of the data, thereby reducing the amount of data transmitted and energy consumption. At the same time, semantic communication can determine the feature information extracted by the sending end based on the downstream industrial tasks, thus transmitting data with higher information entropy and higher value.

[0004] Existing research focuses on semantic communication studies using datasets such as CIFAR10 and CIFAR100, designing decoder models that can reproduce the original data as accurately as possible. However, there is a lack of training on real industrial data. Furthermore, industrial data is highly sensitive, and industrial environments are vulnerable to attacks. The availability of industrial data for model training is limited, resulting in poor training performance. In addition, industrial networks are heterogeneous, with diverse downstream tasks. Small industrial data samples cannot adaptively adjust model parameters, failing to meet the current requirements for semantic information data transmission in industrial environments. Xu et al. proposed ADJSCC, a leading model in the field of semantic transmission for the Industrial Internet of Things. However, the image processing workflow cannot directly handle the resolutions of current mainstream industrial data. Mainstream image resolution adaptation schemes include spatial compression convolution (pixel), large kernel overlapping convolution (conv), and progressively deep separable convolution (ds). These three adaptation schemes lack certain local feature modeling capabilities, easily lose image structural details, and are unable to preserve subtle defects or texture information in industrial images, thus failing to effectively process industrial image data.

[0005] Therefore, there is an urgent need to design an industrial IoT semantic data transmission system and method based on feature enhancement, which can accurately extract and restore semantic features based on a small amount of industrial data, while adapting to the heterogeneous characteristics of industrial networks and the dynamic changes of channels, and meeting the dual requirements of industrial data security and transmission efficiency. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for semantic data transmission in the industrial Internet of Things based on feature enhancement.

[0007] An industrial IoT semantic data transmission method based on feature enhancement is proposed. This method standardizes any raw image to be transmitted collected by an industrial IoT node and then inputs it into an encoder model for semantic feature extraction. The industrial IoT node then sends these semantic features to an edge node. Upon receiving the semantic features, the edge node inputs them into the encoder model for image reconstruction, thus completing the industrial IoT semantic data transmission. The details are as follows:

[0008] The encoder model projects the input image data into the feature space to obtain an initial feature representation. It then employs a layer-by-layer downsampling structure to perform spatial downsampling, normalization compression, and residual local refinement mapping. This process progressively transforms the initial feature map into a low-resolution, high-semantic, and high-density abstract feature representation. The abstract feature representation is then concatenated with the wireless channel signal-to-noise ratio (SNR) and input into a fully connected layer to generate a stable semantic representation. Finally, a lightweight local window self-attention mechanism is used to project the multi-level downsampled output feature map using a linear mapping operator to obtain the final semantic representation for transmission.

[0009] The decoder model embeds the acquired semantic information into a high-dimensional feature space to reconstruct features; it adopts a layer-by-layer upsampling structure to perform spatial upsampling, feature depth extraction, and high- and low-frequency information fusion feature enhancement, gradually transforming the high-dimensional feature space reconstructed features into an initial feature map that is restored to high resolution, low semantics, and high detail; and generates the final reconstructed image result based on the initial feature map after multi-level upsampling.

[0010] Furthermore, the encoder model includes an initial feature construction module, a multi-level downsampling module, and a context enhancement module.

[0011] The initial feature construction module uses convolutional layers to project the input image data into the feature space to obtain an initial feature representation. The multi-level downsampling module adopts a layer-by-layer downsampling structure, performing spatial downsampling, normalization compression, and residual local refinement mapping to gradually transform the initial feature map output by the initial feature construction module into a low-resolution, high-semantic, and high-density abstract feature representation. The context enhancement module processes the output feature map of the multi-level downsampling, employing a lightweight local window self-attention mechanism to enhance numerical stability under low signal-to-noise ratio conditions. It projects the output feature map of the multi-level downsampling through a linear mapping operator to obtain the final semantic representation used for transmission. Through the synergistic effect of these modules, the encoder is adapted for small-sample training in industrial IoT scenarios.

[0012] Furthermore, the decoder model includes a semantic feature mapping reconstruction module, a multi-level upsampling module, and an output mapping reconstruction module.

[0013] The semantic feature mapping reconstruction module embeds the acquired semantic information into the high-dimensional feature space to reconstruct features through convolutional layers; the multi-level upsampling module adopts a layer-by-layer upsampling structure, performs spatial upsampling and feature depth extraction and high- and low-frequency information fusion feature enhancement, and gradually transforms the high-dimensional feature space reconstruction features output by the semantic feature mapping reconstruction module into an initial feature map with high resolution, low semantics, and high detail; the output mapping reconstruction module is used to process the output features of multi-level upsampling, and generates the final reconstructed image result by superimposing the sigmoid function on the convolutional layers.

[0014] Furthermore, the encoder and decoder models are trained using a small sample training set. Specifically, feature enhancement iterative learning is performed. In each training round, a small batch of data is sampled from the training data to perform data augmentation. Random signal-to-noise ratio is sampled, and the augmented data and signal-to-noise ratio are input into the encoder. Semantic features are extracted through the encoder structure module, and the image is reconstructed after being received by the decoder. The loss between the reconstructed image and the original image is calculated, and the model parameters are optimized by combining the observation of peak signal-to-noise ratio and structural similarity index.

[0015] Furthermore, each layer-by-layer downsampling includes a depth-generalized division normalization module, a nonlinear deactivation module, and an attention fusion module. The depth-generalized division normalization module halves the width and height of the initial feature map, expanding the receptive field and effectively removing redundant information in the image. The nonlinear deactivation module achieves efficient feature nonlinear transformation and local detail refinement without introducing complex nonlinear activation functions. While maintaining the same spatial size, it transforms the feature map output by the depth-generalized division normalization module to output a feature map with enhanced feature representation capabilities. The attention fusion module processes the importance weights of the feature map output by the nonlinear deactivation module at different channels or spatial locations, outputting a feature map with amplified important features and suppressing background noise.

[0016] Furthermore, each layer-by-layer upsampling includes a nonlinear inactivation-free module, an attention fusion module, and a depthwise inverse generalized division normalization module. The nonlinear inactivation-free module performs a depth extraction on the current low-resolution features to obtain a feature map of the same size. The attention fusion module processes the feature map output by the nonlinear inactivation-free module to obtain an enhanced feature map that integrates high and low frequency information. The depthwise inverse generalized division normalization module processes the output feature map of the attention fusion module and doubles the width and height of the feature map through transpose convolution to obtain a feature map with a larger size and a reduced number of channels.

[0017] Furthermore, the depth generalized division normalization module superimposed convolutional layers and generalized division normalization layers introduces a channel-by-channel normalization mechanism in the channel dimension.

[0018] Furthermore, the nonlinear non-activation module introduces a channel-wise normalization mechanism in the channel dimension to suppress amplitude instability and enhance energy distribution robustness. By combining convolutional layers and simple gating stacking, and utilizing spatial channel attention, the network adaptively focuses on key channel features related to industrial semantics while ensuring computational efficiency, thereby improving the semantic extraction accuracy based on small sample data in complex industrial scenarios. Overfitting is prevented by discarding layers and layer scaling, thus enhancing the stability of training.

[0019] Furthermore, the attention fusion module concatenates adaptive average pooling with signal-to-noise ratio (SNR) and inputs it into the fully connected layer, matching the encoder output with the wireless channel state, so that the semantic features of the output can dynamically adjust their semantic information according to the channel conditions.

[0020] Furthermore, the depthwise inverse generalized division normalization module consists of convolutional layers, subpixel convolutional layers, and generalized division normalization layers. By transposing convolution and convolution, the width and height of the feature map are doubled, so that the feature distribution is readjusted to the needs of image reconstruction and the accuracy of image reconstruction is improved.

[0021] Furthermore, the standardization process includes image resolution setting, channel conversion, size scaling, training sample partitioning, data integrity verification, and metadata file generation.

[0022] To achieve the above method, this invention also proposes an industrial IoT semantic data transmission system based on feature enhancement. The system includes industrial IoT nodes, wireless Gaussian noise (AWGN) channels, and edge nodes. The industrial IoT nodes are responsible for raw data acquisition and semantic feature extraction, while the edge nodes complete semantic decoding and accuracy evaluation.

[0023] Industrial IoT nodes, acting as data acquisition and semantic encoding endpoints, include sensor modules, encoder modules, and transmission modules. The sensor modules are responsible for acquiring raw image data from the industrial environment, providing the foundational data source for subsequent transmission. The encoder module preprocesses the acquired image data to ensure a consistent data format to meet encoding requirements; it also trains the encoder model based on feature enhancement, accurately extracting core semantic features from the image data to achieve semantic-level data compression. The transmission module adapts to the wireless AWGN channel, sending the semantic features to the edge nodes.

[0024] The edge node, serving as the semantic decoding and accuracy evaluation endpoint, includes a receiving module, a decoder module, and an accuracy evaluation module. The receiving module is matched and adapted to the wireless AWGN channel with the transmitting module to receive semantic features transmitted by industrial IoT nodes. The decoder module reconstructs the received semantic features into the initial data form and performs inverse normalization and pixel value range constraint operations on the reconstructed data, outputting final data that meets the requirements of actual applications. The accuracy evaluation module quantitatively outputs the transmission and reconstruction accuracy, providing a basis for system performance optimization.

[0025] The beneficial effects of this invention are as follows:

[0026] Through the structural design of the encoder and decoder, the system can adaptively allocate reconstruction capabilities to key semantic regions under different signal-to-noise ratio conditions, thereby maintaining high recovery accuracy and structural consistency in complex wireless environments and meeting the current needs of effective transmission of semantic data in the industrial Internet of Things.

[0027] Through feature enhancement training, based on a small amount of industrial data, and through phased and progressive feature enhancement processing, the semantic data transmission of industrial IoT under small sample conditions is realized. While training on real industrial data, it effectively overcomes the situation of poor training effect caused by a small number of training samples, and also prevents large-scale leakage of industrial data, ensuring a certain level of industrial data security. It efficiently transmits extracted semantic information while effectively reducing the exposure of industrial data.

[0028] By extracting semantic information using encoder models deployed in industrial IoT nodes and recovering data using decoder models deployed in edge nodes, the amount of data transmitted is reduced, thus decreasing data redundancy and improving transmission efficiency compared to transmitting raw data. Attached Figure Description

[0029] Figure 1 This is a system configuration diagram of the present invention;

[0030] Figure 2 This is a flowchart of the method steps described in this invention;

[0031] Figure 3 This is a schematic diagram of the nonlinear, non-activation module structure in the embodiment;

[0032] Figure 4 This is a schematic diagram of the attention fusion module structure in the embodiment;

[0033] Figure 5 The PSNR curves under different SNR conditions are shown in the examples.

[0034] Figure 6 The SSIM curves are shown under different SNR conditions in the examples. Detailed Implementation

[0035] The present invention will be further described below with reference to the accompanying drawings.

[0036] like Figure 1 As shown, the feature-enhanced industrial IoT semantic data transmission system includes industrial IoT nodes, wireless Gaussian noise (AWGN) channels, and edge nodes. The industrial IoT nodes are responsible for raw data acquisition and semantic feature extraction, while the edge nodes complete semantic decoding and accuracy evaluation.

[0037] Industrial IoT nodes, acting as data acquisition and semantic encoding endpoints, specifically include sensor modules, encoder modules, and transmission modules. The sensor module is responsible for acquiring raw image data of the industrial environment, providing the basic data source for subsequent transmission. The encoder module preprocesses the acquired image data to ensure a consistent data format to meet encoding requirements; it also trains the encoder model based on feature enhancement, accurately extracting core semantic features from the image data to achieve semantic-level data compression. The transmission module adapts to the wireless AWGN channel, sending the semantic features to the edge nodes.

[0038] The edge node, serving as the semantic decoding and accuracy evaluation endpoint, specifically includes a receiving module, a decoder module, and an accuracy evaluation module. The receiving module is matched and adapted to the wireless AWGN channel with the transmitting module to receive semantic features transmitted by industrial IoT nodes. The decoder module reconstructs the received semantic features into the initial data form and performs inverse normalization and pixel value range constraint operations on the reconstructed data, outputting final data that meets the requirements of actual applications. The accuracy evaluation module calculates the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) to quantitatively output transmission and recovery accuracy, providing a basis for system performance optimization.

[0039] As the core transmission medium connecting industrial IoT nodes and edge nodes, the wireless AWGN channel is responsible for transmitting semantic feature signals. Its characteristics, together with the adaptation design of the transmitting and receiving modules, ensure transmission stability.

[0040] Based on the above system, the following feature-enhanced semantic data transmission methods for industrial IoT are proposed: Figure 2 As shown, the specific steps include the following:

[0041] Step 1: Collect a dataset of real images of the industrial environment using a sensor module, and standardize the dataset to obtain a training set;

[0042] The standardization process in this embodiment is as follows:

[0043] Step 1.1: Obtain and configure the industrial image dataset, set the target image resolution, division ratio and random seed parameters, and perform unified initialization processing on all samples of multiple categories to ensure the reproducibility and structured consistency of subsequent operations.

[0044] Step 1.2: Perform standardized preprocessing operations on each original industrial image, including grayscale to RGB channel conversion, uniform scaling based on bicubic interpolation, and using nearest neighbor interpolation to maintain the binary characteristics of the segmented mask image to ensure the integrity of the image's semantic features.

[0045] Step 1.3: Implement stratified random partitioning for normal training samples of each category, fully preserving the test set structure, including all normal samples and defective samples, to ensure the authenticity and comprehensiveness of the evaluation scenario;

[0046] Step 1.4: Perform multi-dimensional data integrity verification, checking the completeness of the training set directory structure, the non-emptiness of samples in each category, and the records of failed processing files;

[0047] Step 1.5: Generate and save a metadata file containing version information, processing parameters, partitioning strategy, statistical indicators and reference specifications, and output the processed dataset directory structure and usage examples.

[0048] Step 2: Construct the encoder model in the industrial IoT node and the decoder model in the edge node:

[0049] The encoder model includes an initial feature construction module, a multi-level downsampling module, and a context enhancement module.

[0050] The initial feature construction module uses convolutional layers to project the input image data into the feature space to obtain the initial feature representation. ;in The channel mapping operator, representing the local receptive field, is used to construct the basic features for subsequent multi-scale semantic representations. To input industrial images, ;

[0051] In this embodiment, the multi-level downsampling module adopts a three-level layer-by-layer downsampling structure, performing spatial downsampling, normalization compression, and residual local refinement mapping to gradually transform the initial feature map output by the initial feature construction module into a low-resolution, high-semantic, and high-density abstract feature representation. Each downsampling level includes a deep generalized division normalization module, a nonlinear activation-free (NAF) module, and an attention-fusion (AF) module. The deep generalized division normalization module halves the width and height of the initial feature map, expanding the receptive field and effectively removing redundant information in the image. The nonlinear activation-free (NAF) module achieves efficient nonlinear feature transformation and local detail refinement without introducing complex nonlinear activation functions. While maintaining the same spatial size, it transforms the features output by the deep generalized division normalization module into a feature map with enhanced feature representation capabilities. The attention-fusion (AF) module processes the output features of the nonlinear activation-free (NAF) module, processing the importance weights of the output feature map in different channels or spatial locations, outputting a feature map that amplifies important features such as edges and textures, and suppressing background noise.

[0052] The deep generalized division normalization module is superimposed with convolutional layers and generalized division normalization (GDN) layers. It introduces a channel-by-channel normalization mechanism along the channel dimension to suppress amplitude instability and enhance energy distribution robustness. This maintains the advantages of normalization while reducing computational overhead and expanding the receptive field. Specifically:

[0053] In multi-level downsampling, the first class( ) Input features Perform the following operations to halve the feature map space size using a downsampling operator with a step size of 2, while simultaneously expanding the channel dimension to [missing information]. ,in For downsampling mapping operators;

[0054] like Figure 3 As shown, the nonlinear non-activation-free (NAF) module introduces a channel-wise normalization mechanism in the channel dimension to suppress amplitude instability and enhance energy distribution robustness. Combined with convolutional layers and simple gating stacking, and utilizing spatial channel attention, the network adaptively focuses on key channel features relevant to industrial semantics while ensuring computational efficiency, thus improving semantic extraction accuracy in complex industrial scenarios. Overfitting is prevented through layer dropping and layer scaling, enhancing training stability. Specifically:

[0055] Features after size reduction and channel dimension expansion Apply local semantic refinement operator ,in It represents a spatial mixing map (obtained by combining channel expansion, channel-wise convolution, and gating). Indicates channel blending mapping, , The learnable scaling parameter is initialized to zero and outputs a feature map with enhanced feature representation capabilities.

[0056] like Figure 4 As shown, the attention fusion (AF) module concatenates adaptive average pooling with signal-to-noise ratio (SNR) and inputs it to the fully connected layer. This matches the encoder output with the wireless channel state, enabling the semantic features of the output to dynamically adjust their semantic information according to channel conditions. Specifically:

[0057] The channel statistics vector is obtained by the spatial global average pooling operator. and the signal-to-noise ratio scalar provided externally. Integration to form a joint representation , For local semantic refinement operators;

[0058] To match the encoder output with the wireless channel conditions, a channel-aware channel modulation mechanism is introduced after each level of feature refinement, enabling the encoded features to dynamically adjust their semantic emphasis according to channel conditions. Channel weights. Obtained by nonlinear mapping ,in , For linear mapping operators, It is a non-linear activation function. The final weighted output of the sigmoid function is: .

[0059] The context enhancement module processes the output feature maps of multi-level downsampling. It employs a lightweight local window self-attention mechanism to enhance numerical stability under low signal-to-noise ratio conditions. A linear mapping operator projects the multi-level downsampling output feature maps to obtain the final semantic representation used for transmission. Specifically:

[0060] The feature map output from the multi-level downsampling is divided into several parts of size . Non-overlapping windows are used to perform self-attention operations on the feature sequences within each window. ,in Obtained by linear projection of features within the window, the attention output is returned and added before the backbone features, incorporating normalization and a zero-initialized scaling factor. , to obtain features ,in This represents the inverse rearrangement and linear projection operations for window attention. The initial value is set to zero; the bottleneck-enhanced features are passed through the linear mapping operator. Projecting onto the target semantic channels yields the final semantic representation field. in , The number of channels is used to represent semantics. This refers to the semantic information that the final encoder uses for transmission.

[0061] The decoder model includes a semantic feature mapping reconstruction module, a multi-level upsampling module, and an output mapping reconstruction module, which gradually restores the low-resolution, high-semantic features compressed by the encoder into high-resolution feature maps and images.

[0062] The semantic feature mapping and reconstruction module embeds the acquired semantic information into a high-dimensional feature space through convolutional layers to reconstruct features; specifically:

[0063] Semantic information transmitted by the sending end After passing through the wireless AWGN channel and incurring certain losses, the signal reaches the receiver. The semantic representation obtained by the receiver is as follows: The decoder embeds the features into a high-dimensional feature space through the semantic feature mapping reconstruction module to obtain the initial reconstructed features. ,in It is a local convolution mapping operator used to recover the channel dimensions and spatial structure that match the deepest semantic layer;

[0064] In this embodiment, the multi-level upsampling module adopts a three-level layer-by-layer upsampling structure. Each level of upsampling includes a Nonlinear Activation-Free (NAF) module, an Attention-Fusion (AF) module, and an Inverse Generalized Divide Normalization (IGDN) module. Before upsampling, the NAF module performs a depth extraction on the current low-resolution features to obtain a feature map of the same size. The Attention-Fusion (AF) module processes the features output by the NAF module to obtain an enhanced feature map that integrates high and low frequency information. The IGDN module processes the output feature map of the Attention-Fusion (AF) module and doubles the width and height of the feature map through transpose convolution to obtain a feature map with a larger size and a reduced number of channels.

[0065] The Nonlinear Non-Activation-Free (NAF) module and the Attention Fusion (AF) module have the same structure as those in the encoder. Specifically, the NAF module processes the initial reconstructed features... A local semantic refinement mapping is applied to enhance the consistency of its internal structure. This mapping is expressed in residual form as follows: ,in and These represent spatial blending mapping and channel blending mapping, respectively. , The learnable scaling parameter is initialized to zero to ensure numerical stability under low signal-to-noise ratio conditions; the AF module performs global statistics on the features and compares them with the signal-to-noise ratio parameter. Merge to generate channel weight vectors: ,in This represents the global average pooling operator. It is a nonlinear mapping function. The Sigmoid function outputs the following: ;

[0066] The Deep Inverse Generalized Division Normalization (IGDN) module consists of convolutional layers, subpixel convolutional layers, and generalized division normalization (GDN) layers. Through transposed convolution and convolution, it doubles the width and height of the feature map, allowing the feature distribution to readjust to the needs of image reconstruction and improving image reconstruction accuracy. Specifically:

[0067] For the first class( The feature map resolution is increased by a factor of 2 by a spatial upsampling operator, and combined with channel-wise inverse normalization mapping, the feature amplitude distribution is gradually restored to the pixel-domain scale. = ,in This represents a combination operator of upsampling mapping and inverse normalization; after completing the three-level upsampling, a feature representation with the same resolution as the input image is obtained. .

[0068] The output mapping reconstruction module is used to process multi-level upsampled output features. It uses convolutional layers with a sigmoid function to upsample the features after three levels of upsampling. The final reconstructed image result is generated by: using the output mapping operator to upsample the features after three levels. Generate final reconstruction results ,in For convolution mapping operators, The element-wise nonlinear compression function restricts the output to the range of valid pixel values, ultimately yielding the reconstructed image data. .

[0069] Step 3: Train the encoder and decoder models constructed in Step 2 using the training set from Step 1;

[0070] In this embodiment, the training process for the encoder model in the industrial IoT node and the decoder model in the edge node is as follows:

[0071] Step 2.1: Initialize the encoder and decoder structure framework, configure training hyperparameters, including learning rate, batch size, weight decay coefficient, and select AdamW optimizer and mean squared error loss function;

[0072] Step 2.2: Perform feature enhancement iterative learning. In each training round, sample a small batch of data from the training data, perform data augmentation, sample random signal-to-noise ratio, input the augmented data and signal-to-noise ratio into the encoder, extract semantic features through the encoder structure module, and reconstruct the image after receiving it through the decoder.

[0073] Step 2.3: Calculate the loss between the reconstructed image and the original image, and optimize the model parameters by observing the Peak Signal-to-Noise Ratio (PSNR) and Structure Similarity Index Measure (SSIM). Calculate the gradient through backpropagation, update the model parameters using the AdamW optimizer, and adjust the ReduceLROnPlateau learning rate scheduler to ensure the improvement of the system's peak signal-to-noise ratio performance.

[0074] The Peak Signal-to-Noise Ratio (PSNR) is used to quantitatively evaluate the error magnitude between the reconstructed image and the original image, and it is defined as follows: ;in, The maximum value of an image pixel (usually 1 after normalization) ), Mean Square Error (MSE) is calculated using the following formula: ;in For the original image tensor, To reconstruct the image tensor, For batch size, For the number of channels, and These represent the height and width of the image, respectively. A higher PSNR value indicates a smaller error between the reconstructed image and the original image, and higher accuracy in system transmission and restoration.

[0075] The Structure Similarity Index Measure (SSIM) is used to measure the similarity between two images in terms of structure, brightness, and contrast. It is defined as follows: ;in and Original and image respectively With reconstructed images The average pixel value, and These are the pixel variances of the two images, respectively. The covariance of the two images and SSIM is a stability constant used to avoid instability when the denominator approaches zero. The range of values ​​for SSIM is... The closer the value is This indicates that the more structurally similar the two images are, the better the visual fidelity of the image recovered by the system.

[0076] When the PSNR fails to improve for five consecutive rounds, the ReduceLROnPlateau learning rate scheduler can reduce the learning rate to 0.5 times its original value and apply gradient clipping to limit the gradient norm from exceeding a threshold. Training is terminated and the optimal model parameters are saved when the PSNR does not exceed the historical best value for 15 consecutive rounds. and This completes the feature learning and training process under small sample conditions.

[0077] Step 4: After standardizing any raw image collected by the industrial IoT node that needs to be transmitted, input it into the encoder model trained in Step 3 for semantic feature extraction. The industrial IoT node sends the semantic features to the edge node through the transmitting module and wireless channel. After receiving the semantic features through the receiving module, the edge node inputs the semantic features into the encoder model trained in Step 3 for image reconstruction, thus completing the industrial IoT semantic data transmission.

[0078] To demonstrate the effectiveness of the above method, this invention uses the ADJSCC model to run three different mainstream adaptation schemes and compares their performance with the method described in this invention for the same data transmission. Specifically, these include: spatial compression convolution (pixel), large kernel overlapping convolution (conv), and progressive depthwise separable convolution (ds). All models are trained and tested based on the same industrial dataset and training process to ensure the fairness and effectiveness of the comparison.

[0079] like Figure 5The figure shows the peak signal-to-noise ratio (PSNR) performance comparison curves of the method described in this invention and the comparison model under different signal-to-noise ratio (SNR) conditions; Table 1 shows the PSNR values ​​of the four methods under different SNR conditions. Table 1 shows the peak signal-to-noise ratio (PSNR) performance comparison between the present invention and the comparative model under different signal-to-noise ratio (SNR) conditions. When SNR=0dB (strong noise environment), the PSNR of the present invention reaches 34.1015dB, which is 9.5595dB higher than the 'ds' method (24.5420dB), 10.4546dB higher than the 'conv' method (23.6469dB), and 11.0252dB higher than the 'pixel' method (23.0763dB), effectively solving the pain point of excessive error in industrial data transmission under low signal-to-noise ratio. As SNR increases, the PSNR of the present invention shows a continuous and stable upward trend, reaching 39.4993dB when SNR=20dB (weak noise environment), which is 14.3539dB higher than the 'ds' method (25.1454dB), further expanding its advantages. The PSNR value of the comparison method increases slowly with SNR, only improving by 0.8~1.3dB from 0dB to 20dB. This indicates that its feature extraction and noise resistance mechanisms are not very adaptable to channel changes and it is difficult to make full use of channel resources to improve recovery accuracy.

[0080] Table 1

[0081]

[0082] from Figure 5 As can be seen from Table 1, the method described in this invention significantly outperforms the three comparative models in terms of PSNR values ​​across the entire SNR range, and its advantages are evident in signal interference environments commonly encountered in industrial settings.

[0083] like Figure 6The figure shows the performance comparison curves of the structural similarity index (SSIM) between the method of the present invention and the comparison model under different signal-to-noise ratios (SNR). Table 2 shows the comparison of SSIM values ​​of the four methods under different SNR conditions. Table 2 shows the performance comparison values ​​of the structural similarity index (SSIM) between the present invention and the comparison model under different signal-to-noise ratios (SNR). When SNR=0dB, the SSIM of the present invention is 0.9123, which is 0.3038 higher than the 'ds' method (0.6085) and 0.3449 higher than the 'pixel' method (0.5674). Even under strong noise interference, it can still accurately preserve the core structural information of industrial images. As SNR increases, the SSIM of the present invention steadily approaches 1, reaching 0.9729 when SNR=20dB. This means that the structural consistency between the reconstructed image and the original image reaches a high level, which can meet the requirements of industrial scenarios with strict detail requirements, such as mechanical parts inspection and assembly line monitoring. The SSIM value of the comparison model remained at a low level and the increase was small (the maximum increase was only 0.0485), reflecting its insufficient ability to extract semantic features of industrial images and its inability to effectively restore key structural information of the images.

[0084] Table 2

[0085]

[0086] from Figure 6 As can be seen from Table 2, the SSIM value of the present invention is significantly better than that of the three comparative models in the entire SNR range, and remains above 0.9 in the entire SNR range, which is significantly higher than that of the comparative models (not exceeding 0.62).

[0087] Comparing the three adaptation schemes, the 'ds' method, due to its advantage in balancing feature representation and computational complexity through depthwise separable convolution, slightly outperforms the 'conv' and 'pixel' methods. However, all three lack adaptation designs and channel-aware mechanisms for industrial small-sample scenarios, making it difficult to balance transmission efficiency and recovery accuracy in industrial environments. This invention improves data transmission efficiency and recovery accuracy through synergistic innovation of feature enhancement and channel-aware technologies.

[0088] In summary, the feature-enhanced industrial IoT semantic data transmission system proposed in this invention significantly outperforms existing comparative models in both PSNR and SSIM, with its advantages being particularly pronounced under low signal-to-noise ratio conditions. This fully verifies the effectiveness of the system in achieving efficient and secure semantic data transmission in environments with limited industrial data samples and dynamically changing channels. This invention provides a reliable technical solution for semantic communication in the industrial IoT and is suitable for semantic data transmission needs in various industrial scenarios.

Claims

1. A semantic data transmission method for the Industrial Internet of Things (IIoT) based on feature enhancement, wherein any original image to be transmitted collected by an IIoT node is standardized and then input into an encoder model for semantic feature extraction; the IIoT node sends the semantic features to an edge node; the edge node receives the semantic features and inputs them into a decoder model for image reconstruction, thus completing the IIoT semantic data transmission; characterized in that, Specifically as follows: The encoder model projects the input image data into the feature space to obtain an initial feature representation. It then employs a layer-by-layer downsampling structure to perform spatial downsampling, normalization compression, and residual local refinement mapping. This process progressively transforms the initial feature map into a low-resolution, high-semantic, and high-density abstract feature representation. The abstract feature representation is then concatenated with the wireless channel signal-to-noise ratio (SNR) and input into a fully connected layer to generate a stable semantic representation. Finally, a lightweight local window self-attention mechanism is used to project the multi-level downsampled output feature map using a linear mapping operator to obtain the final semantic representation for transmission. The decoder model embeds the acquired semantic information into a high-dimensional feature space to reconstruct features; A layer-by-layer upsampling structure is adopted to perform spatial upsampling, feature depth extraction, and high- and low-frequency information fusion feature enhancement. The high-dimensional feature space reconstructed features are gradually transformed into initial feature maps with high resolution, low semantics, and high detail. The final reconstructed image result is generated based on the initial feature maps after multi-level upsampling. The layer-by-layer downsampling includes a depth generalized division normalization module, a nonlinear activation-free module, and an attention fusion module. The depth generalized division normalization module halves the width and height of the initial feature map, expands the receptive field, and effectively removes redundant information in the image. The nonlinear activation-free module achieves efficient feature nonlinear transformation and local detail refinement without introducing complex nonlinear activation functions. While keeping the spatial size unchanged, it outputs a feature map with enhanced feature representation capabilities from the feature transformation output by the depth generalized division normalization module. The attention fusion module processes the importance weights of the feature map output by the nonlinear inactivation module at different channels or spatial locations, outputs a feature map with amplified important features, and suppresses background noise. Through the synergistic effect of each module, the encoder is adapted to small-sample training in industrial IoT scenarios. Each layer-by-layer upsampling includes a nonlinear inactivation-free module, an attention fusion module, and a deep inverse generalized division normalization module. The nonlinear non-activation module performs a depth extraction on the current low-resolution features to obtain a feature map of the same size. The attention fusion module processes the feature map output by the nonlinear inactivation module to obtain an enhanced feature map that incorporates high and low frequency information. The deep inverse generalized division normalization module processes the output feature map of the attention fusion module and doubles the width and height of the feature map through transpose convolution to obtain a feature map with larger size and fewer channels.

2. The industrial IoT semantic data transmission method based on feature enhancement as described in claim 1, characterized in that, The depth-generalized division normalization module superimposed with convolutional layers and generalized division normalization layers introduces a channel-by-channel normalization mechanism in the channel dimension.

3. The industrial IoT semantic data transmission method based on feature enhancement as described in claim 1, characterized in that, In the nonlinear non-activation module, layer normalization introduces a channel-wise normalization mechanism in the channel dimension to suppress amplitude instability and enhance energy distribution robustness. By combining convolutional layers and simple gating stacking, and utilizing spatial channel attention, the network adaptively focuses on key channel features related to industrial semantics while ensuring computational efficiency, thereby improving the semantic extraction accuracy based on small sample data in complex industrial scenarios. Overfitting is prevented by discarding layers and layer scaling, thus enhancing the stability of training.

4. The industrial IoT semantic data transmission method based on feature enhancement as described in claim 1, characterized in that, The attention fusion module concatenates adaptive average pooling with signal-to-noise ratio (SNR) and inputs it into the fully connected layer. This matches the encoder output with the wireless channel state, enabling the semantic features of the output to dynamically adjust their semantic information according to the channel conditions.

5. The industrial IoT semantic data transmission method based on feature enhancement as described in claim 1, characterized in that, The depthwise inverse generalized division normalization module consists of convolutional layers, subpixel convolutional layers, and generalized division normalization layers. By transposing convolution and convolution, it doubles the width and height of the feature map, making the feature distribution adapt to the needs of image reconstruction and improving the accuracy of image reconstruction.

6. The industrial IoT semantic data transmission method based on feature enhancement as described in claim 1, characterized in that, The standardization process includes image resolution setting, channel conversion, size scaling, training sample partitioning, data integrity verification, and metadata file generation.

7. The industrial IoT semantic data transmission method based on feature enhancement as described in claim 1, characterized in that, The encoder and decoder training process is as follows: performing feature enhancement iterative learning, sampling small batches of data from the training data in each round of training, performing data enhancement, sampling random signal-to-noise ratio, inputting the enhanced data and signal-to-noise ratio into the encoder, extracting semantic features through the encoder structure module, and performing image reconstruction after being received by the decoder; calculating the loss between the reconstructed image and the original image, and optimizing the model parameters by combining observation of peak signal-to-noise ratio and structural similarity index.

8. An industrial IoT semantic data transmission system for implementing the method of claim 1, characterized in that, It includes industrial IoT nodes, wireless Gaussian noise channels, and edge nodes. Industrial IoT nodes are responsible for raw data acquisition and semantic feature extraction, while edge nodes complete semantic decoding and accuracy evaluation. Industrial IoT nodes serve as data acquisition and semantic encoding endpoints, comprising a sensor module, an encoder module, and a transmission module. The sensor module is responsible for acquiring raw image data of the industrial environment, providing a basic data source for subsequent transmission. The encoder module preprocesses the acquired image data to ensure that the data format is uniform and adapts to encoding requirements. It also trains the encoder model based on feature enhancement, accurately extracting the core semantic features from the image data and achieving semantic compression of the data. The transmission module adapts to wireless Gaussian noise channels and sends the semantic features to the edge nodes. The edge node serves as the semantic decoding and accuracy evaluation end, including a receiving module, a decoder module, and an accuracy evaluation module; the receiving module is matched and adapted to the wireless Gaussian noise channel to receive semantic features sent by industrial IoT nodes; The decoder module reverse-engineers the received semantic features into the initial data form, and performs inverse normalization and pixel value range constraint operations on the reconstructed data to output the final data that meets the actual application requirements. The accuracy evaluation module calculates the peak signal-to-noise ratio and structural similarity index to quantitatively output the transmission and recovery accuracy, providing a basis for system performance optimization.

Images