Method for improving image recognition precision of new energy station

Abstract

The invention provides a method for improving the image recognition precision of a new energy station, and relates to the technical field of image recognition, and the method comprises the steps: obtaining monitoring image data and environment parameter data, and carrying out the image preprocessing based on an environment impact factor and an image degradation score; extracting image features by using a deep convolutional neural network, and extracting time sequence correlation features by using a recurrent neural network; and feature optimization is realized through feature stability evaluation and dynamic calibration, and finally the working state and abnormal early warning information of the equipment are output. According to the method, the precision and robustness of new energy station image recognition can be effectively improved, and the interference of environmental factors on a recognition result is reduced.

Description

Technical Field

[0001] The present invention relates to image recognition technology, and in particular to a method for improving the image recognition accuracy of new energy stations. Background Art

[0002] The demand for intelligent operation, maintenance and management of new energy stations is growing. Image recognition technology, as an important means of intelligent station, can realize automatic monitoring and fault diagnosis of new energy power generation equipment. The existing image recognition technology for new energy stations has the following shortcomings: New energy stations are usually located in outdoor environments and are significantly affected by environmental factors such as weather and light, resulting in unstable monitoring image quality. Existing technologies lack systematic consideration of environmental factors and cannot dynamically adjust image processing strategies according to actual environmental conditions, affecting recognition accuracy and reliability.

[0003] The working status of renewable energy power generation equipment has the characteristics of continuity and time correlation, but the existing recognition methods are mainly based on single-frame image analysis, ignoring the time sequence information contained in the image sequence, and it is difficult to accurately capture the dynamic changes of the equipment status.

[0004] Existing technologies have limitations in feature extraction and optimization, lack effective feature stability evaluation and dynamic calibration mechanisms, and are easily affected by image noise and interference, resulting in unstable and unreliable recognition results, especially poor recognition performance under complex environmental conditions. Summary of the invention

[0005] The embodiment of the present invention provides a method for improving the image recognition accuracy of new energy stations, which can solve the problems in the prior art. It includes: Acquire monitoring image data of new energy stations; collect environmental parameter data based on the sensor array arranged according to the Voronoi rule, and calculate image degradation scores and environmental impact factors through the dynamic coupling relationship of environmental parameters; Build a deep learning model including preprocessing module, feature extraction module and feature optimization module; The preprocessing module dynamically adjusts image enhancement parameters according to the environmental influencing factors, selects a noise removal strategy based on the image degradation score, and performs adaptive histogram equalization and image normalization processing; The feature extraction module uses a deep convolutional neural network to extract features from the preprocessed monitoring image data to obtain an image feature vector; The feature optimization module processes the image feature vector to obtain the recognition result of the new energy power generation equipment in the new energy station, wherein the time series feature extraction unit uses a recurrent neural network to extract the time series correlation features of continuous frame images, the feature stability evaluation unit calculates the stability score of the feature based on the multi-dimensional similarity measurement and the optimal transmission distance, and the dynamic calibration unit adaptively adjusts the current feature through the Squeeze-and-Excitation modulation layer and the residual nonlinear transformation based on the stability score; the recognition result includes equipment working status information and abnormal warning information.

[0006] In an optional embodiment, The sensor array arranged based on the Voronoi rule collects environmental parameter data, and the image degradation score and environmental impact factors are calculated through the dynamic coupling relationship of environmental parameters, including: The new energy stations are divided based on the Voronoi partitioning rule, and an environmental sensor array is arranged at the center of the divided area. The environmental sensor array collects light intensity, temperature, humidity, air quality and visibility data; the collected environmental parameter data is preprocessed to obtain smooth data; Calculate the spatial distribution estimation result and the time series prediction result based on the smoothed data and combine them to form an environmental state matrix; Establishing an associative mapping relationship between the data in the environmental state matrix and the image quality degradation type respectively, including: calculating the dynamic coupling relationship between environmental parameters, calculating the coupling strength between the light intensity, temperature, humidity, air quality and visibility parameters based on parameter similarity, change rate difference and spatial adjacency, and forming a dynamic coupling matrix; calculating the image quality index by a time attenuation weighting method based on the current environmental parameters and the dynamic coupling matrix, and performing time series smoothing and calibration on the image quality index to obtain an image blur index and an image contrast index; A degradation type vector is constructed based on the image blur index and the image contrast index, wherein the degradation type vector includes a blur component, a contrast component and a noise component; the degradation type vector is weightedly combined with a preset weight coefficient to obtain an image degradation score; and the environmental parameters in the environmental state matrix are combined with a preset influence function and a weight coefficient to calculate an environmental influence factor.

[0007] In an optional embodiment, Dynamically adjusting image enhancement parameters according to the environmental influencing factors, selecting a noise removal strategy based on the image degradation score, and performing adaptive histogram equalization and image normalization processing include: Establishing a mapping relationship between environmental impact factors and image enhancement parameters, wherein the coefficients of the mapping relationship are dynamically adjusted based on the environmental state matrix; constructing an enhancement operator library including a contrast enhancement operator and a sharpening enhancement operator, wherein the parameters of the contrast enhancement operator are adaptively adjusted according to a mapping rule between light intensity and contrast, and the parameters of the sharpening enhancement operator are dynamically adjusted according to an image blur index; Constructing an image quality evaluation function, the evaluation function including a peak signal-to-noise ratio index, a structural similarity index and a no-reference image quality evaluation index, the weight of the evaluation function being adaptively adjusted according to environmental influencing factors, and optimizing image enhancement parameters based on the evaluation function; Based on the image degradation score, noise type is identified, and a corresponding noise removal strategy is selected according to the noise type, including multi-scale noise analysis and adaptive filtering processing, and parameters of the filtering processing are dynamically adjusted according to noise characteristics; Perform adaptive histogram equalization on the filtering results, including adaptive blocking based on local entropy value and histogram equalization parameter adjustment. The histogram equalization parameters include equalization strength and block fusion weight. The histogram equalization parameters are dynamically adjusted according to environmental influencing factors. The equalization result is normalized based on the time series statistical features, wherein the time series statistical features include the mean distribution and the variance distribution of the image sequence, and the scaling parameters and the translation parameters of the normalization are dynamically adjusted according to the time series prediction result.

[0008] In an optional embodiment, The deep convolutional neural network of the feature extraction module includes: A main network and an auxiliary branch network, wherein the main network adopts a dense connection structure, including four dense blocks, and the number of convolution kernels in each dense block is adaptively adjusted according to the complexity of the image; the auxiliary branch network adopts a lightweight structure and receives an environment state matrix as input; A channel attention module is set after each dense block of the backbone network, and the weight coefficient of the channel attention module is modulated by the environmental feature vector output by the auxiliary branch network; a feature recalibration unit is set between adjacent dense blocks, and the feature recalibration unit dynamically adjusts the scale parameter of the feature map based on the image quality score of the local area; At the end of the backbone network, the global average pooling layer and the fully connected layer are connected in series to output the image feature vector.

[0009] In an optional embodiment, The temporal feature extraction unit uses a recurrent neural network to extract the temporal correlation features of continuous frame images. The feature stability evaluation unit calculates the stability score of the feature based on multi-dimensional similarity measurement and optimal transmission distance. The dynamic calibration unit adaptively adjusts the current feature based on the stability score through the Squeeze-and-Excitation modulation layer and residual nonlinear transformation, including: The image feature vector is constructed as a feature sequence of continuous frame images, and the feature sequence is input into a bidirectional long short-term memory network, wherein the output features of the forward long short-term memory unit and the backward long short-term memory unit are concatenated to obtain a time series correlation feature; Calculate the local stability score and the global stability score of the time series correlation feature, the local stability score is obtained by constructing a directed graph structure and calculating the edge weight based on a multi-dimensional similarity metric and an optimal transmission distance, the global stability score is calculated by an entropy complexity index of a multi-scale feature sequence, and the local stability score and the global stability score are fused through an adaptive weight network to obtain a feature stability score; The time series correlation feature is calibrated based on the stability score, wherein the calibration generates attention weights through a Squeeze-and-Excitation modulation layer in a dynamic calibration unit, and the modulated features are calibrated using a residual nonlinear transformation, and an optimal historical feature sequence is selected based on dynamic programming, and a comparison learning sample pair is constructed between the enhanced view of the time series correlation feature and the enhanced view of the historical feature, and a calibration feature is obtained through multi-objective optimization; Performing exponential sliding average on the calibration feature to obtain a smooth feature, wherein the smoothing coefficient of the exponential sliding average is dynamically adjusted according to the stability score; inputting the smooth feature into a classifier to obtain device working status information; The Euclidean distance between the smoothing feature and the time series correlation feature is calculated, the Euclidean distance is multiplied by the complementary value of the stability score to obtain an abnormality score, an adaptive threshold is constructed based on the statistical characteristics of the abnormality score, and abnormal warning information is generated when the abnormality score exceeds the adaptive threshold.

[0010] In an optional embodiment, The local stability score and the global stability score of the time series correlation feature are calculated, the local stability score is obtained by constructing a directed graph structure and calculating the edge weight based on a multi-dimensional similarity metric and an optimal transmission distance, the global stability score is calculated by an entropy complexity index of a multi-scale feature sequence, and the feature stability score is obtained by fusing the local stability score and the global stability score through an adaptive weight network, including: The temporal correlation features within the time window are constructed into a directed graph structure, and the edge weights of the directed graph structure are calculated by multi-dimensional similarity measurement; The multi-dimensional similarity measurement includes: constructing a multi-level cost matrix between feature distributions, the multi-level cost matrix includes a point-to-point cost matrix based on local features and a structural cost matrix based on global features; solving the optimal transmission problem based on the Sinkhorn algorithm to obtain a transmission scheme, the transmission scheme satisfies the first-order moment matching constraint and the second-order moment correlation constraint; calculating the multi-scale optimal transmission distance and obtaining the transmission distance through a scale-adaptive fusion network; and performing a weighted combination of the transmission distance and the mutual information through a parameterized sigmoid mapping function to obtain the edge weight; A transfer matrix is ​​constructed based on the edge weights, the importance of nodes in the directed graph structure is obtained through iterative calculation, and the product of the node importance and the edge weight is normalized to obtain a local stability score; Constructing a multi-scale feature sequence of the time series correlation feature, calculating the sample entropy of the multi-scale feature sequence, constructing an entropy complexity plane based on the sample entropy to obtain a complexity index, and calculating a global stability score based on the complexity index; A feature state vector including the local stability score, the global stability score and their time difference is constructed, the feature state vector is input into an adaptive weight network with a Softmax activation function to obtain a fusion weight, and the local stability score and the global stability score are combined by the fusion weight to obtain a stability score of the feature.

[0011] In an optional embodiment, The time series correlation feature is calibrated based on the stability score. The calibration generates attention weights through a Squeeze-and-Excitation modulation layer in a dynamic calibration unit, and the modulated features are calibrated using a residual nonlinear transformation. The optimal historical feature sequence is selected based on dynamic programming, and the enhanced view of the time series correlation feature and the enhanced view of the historical feature are used to construct a comparative learning sample pair. The calibration features obtained through multi-objective optimization include: A Squeeze-and-Excitation modulation layer is constructed in the dynamic calibration unit, global average pooling is performed on the temporal correlation features, and attention weights are generated through a feedforward network, the attention weights are adjusted based on the stability score, and the temporal correlation features are modulated with the attention weights to obtain modulation features; The modulation feature is subjected to a residual nonlinear transformation to obtain an initial calibration feature, wherein the residual nonlinear transformation includes passing the modulation feature through a first transformation layer with a ReLU activation function and a second transformation layer with a tanh activation function, multiplying the transformation result with a sigmoid output of a feature stability score and then adding the result to the modulation feature; Construct a self-similarity matrix of the time series correlation feature, the self-similarity matrix is ​​calculated by combining the cosine similarity and Gaussian kernel similarity of the feature vector, calculate the state score between the current moment and the historical moment, construct a state transfer equation based on the state score to calculate the optimal state value, the calculation range of the state transfer equation is N moments before the current moment, backtrack from the optimal state value to obtain the feature selection path and extract the corresponding historical features, where the formula for the state score is: ; Where t is the current moment, k is the historical moment, |tk| is the time interval between two moments, and f t is the feature vector at the current time t, f k is the feature vector of historical moment k, sim cos (f k， f t ) is the feature vector f k and f t The cosine similarity between ||f k -f t || 2 is the feature vector f k and f t The square of the Euclidean distance, σ is the bandwidth parameter of the Gaussian kernel, λ is the time decay factor, the larger the value, the faster the time decay; Generate enhanced views of the initial calibration feature and the historical feature respectively through random mask and Gaussian noise, construct different enhanced views of the same feature as a positive sample pair, construct the enhanced views of the initial calibration feature and the historical feature as a negative sample pair, and calculate the contrast loss through the InfoNCE loss function; The initial calibration feature is fused with the weighted average of the historical features to obtain the calibration feature, a multi-objective loss function including mean square error, contrast loss and time series smoothing regularization term is constructed, and the parameters of the dynamic calibration unit are optimized using a cosine annealing learning rate strategy.

[0012] In an optional embodiment, Establishing an associative mapping relationship between the data in the environment state matrix and the image quality degradation type includes: The dynamic coupling relationship between environmental parameters is calculated to obtain the dynamic coupling matrix C(t), which is calculated as follows: ; Among them, C ij (t) is the coupling strength between the i-th environmental parameter and the j-th environmental parameter at time t; E i (t) is the value of the i-th environmental parameter at time t; E j (t) is the jth environmental parameter value at time t; ψ(Ei (t), E j (t)) is the parameter similarity function; τ(t) is the time-varying attenuation coefficient, which controls the influence of the rate of change; ΔE ij (t) is the difference in parameter change rate, which is the difference in the speed of change of the two parameters; S ij (t) is the spatial correlation coefficient, which reflects the correlation degree of the parameters in spatial distribution; Based on the environmental parameters and the dynamic coupling matrix, the image quality index Q(t) is calculated, and its calculation formula is: ; Among them, t is the time stamp of the current moment; k is the historical moment index, from 1 to N, where N is the size of the historical time window considered; ω k (t) is the adaptive weight coefficient; E(t) is the environmental parameter vector at the current moment; C(t) is the dynamic coupling matrix at the current moment; f k is the mapping function between environmental parameters and quality indicators; λ is the time decay factor; |tk| is the time interval between the current moment and the historical moment; The image quality index is time-series smoothed and calibrated, and the statistical characteristics of the quality index at the latest N moments are calculated; the current quality index is weighted smoothed based on the statistical characteristics; the smoothing result is calibrated according to historical observation data to obtain a blur index and a contrast index.

[0013] The beneficial effects of the present invention are as follows: The present invention collects environmental parameter data through a sensor array arranged based on the Voronoi rule, calculates image degradation scores and environmental impact factors, and realizes accurate evaluation of the monitoring image quality of new energy stations, provides a reliable reference basis for subsequent image processing, and effectively improves the adaptability of image recognition.

[0014] The present invention designs a deep learning model including a preprocessing module, a feature extraction module and a feature optimization module, wherein the preprocessing module can dynamically adjust image enhancement parameters according to environmental influencing factors and select noise removal strategies based on image degradation scores, thereby ensuring the pertinence and effectiveness of image preprocessing and laying a good foundation for subsequent feature extraction.

[0015] The present invention innovatively introduces a time series feature extraction unit, a feature stability evaluation unit and a dynamic calibration unit in the feature optimization module, calculates the stability score of the feature through multi-dimensional similarity measurement and optimal transmission distance, and uses the Squeeze-and-Excitation modulation layer and residual nonlinear transformation to perform feature adaptive adjustment, which significantly improves the accuracy and reliability of new energy power generation equipment identification and provides strong support for equipment status monitoring and abnormal early warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A flow chart of a method for improving the image recognition accuracy of a new energy station according to an embodiment of the present invention; Figure 2 This is a schematic diagram of optimizing the layout of environmental sensors in new energy stations based on the Voronoi rule of the present invention; Figure 3 This is a comparison chart of the performance of different similarity measurement methods during feature changes; Figure 4 It is the evolution diagram of the dynamic coupling matrix of environmental parameters of the present invention over time. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0019] Figure 1 FIG. 1 is a flow chart of a method for improving the image recognition accuracy of a new energy station according to an embodiment of the present invention. Figure 1 As shown, the method includes: Acquire monitoring image data of new energy stations; collect environmental parameter data based on the sensor array arranged according to the Voronoi rule, and calculate image degradation scores and environmental impact factors through the dynamic coupling relationship of environmental parameters; Build a deep learning model including preprocessing module, feature extraction module and feature optimization module; The preprocessing module dynamically adjusts image enhancement parameters according to the environmental influencing factors, selects a noise removal strategy based on the image degradation score, and performs adaptive histogram equalization and image normalization processing; The feature extraction module uses a deep convolutional neural network to extract features from the preprocessed monitoring image data to obtain an image feature vector; The feature optimization module processes the image feature vector to obtain the recognition result of the new energy power generation equipment in the new energy station, wherein the time series feature extraction unit uses a recurrent neural network to extract the time series correlation features of continuous frame images, the feature stability evaluation unit calculates the stability score of the feature based on the multi-dimensional similarity measurement and the optimal transmission distance, and the dynamic calibration unit adaptively adjusts the current feature through the Squeeze-and-Excitation modulation layer and the residual nonlinear transformation based on the stability score; the recognition result includes equipment working status information and abnormal warning information.

[0020] In an optional implementation, the sensor array arranged based on the Voronoi rule collects environmental parameter data, and the image degradation score and environmental impact factor are calculated through the dynamic coupling relationship of the environmental parameters, including: The new energy stations are divided based on the Voronoi partitioning rule, and an environmental sensor array is arranged at the center of the divided area. The environmental sensor array collects light intensity, temperature, humidity, air quality and visibility data; the collected environmental parameter data is preprocessed to obtain smooth data; Calculate the spatial distribution estimation result and the time series prediction result based on the smoothed data and combine them to form an environmental state matrix; Establishing an associative mapping relationship between the data in the environmental state matrix and the image quality degradation type respectively, including: calculating the dynamic coupling relationship between environmental parameters, calculating the coupling strength between the light intensity, temperature, humidity, air quality and visibility parameters based on parameter similarity, change rate difference and spatial adjacency, and forming a dynamic coupling matrix; calculating the image quality index by a time attenuation weighting method based on the current environmental parameters and the dynamic coupling matrix, and performing time series smoothing and calibration on the image quality index to obtain an image blur index and an image contrast index; A degradation type vector is constructed based on the image blur index and the image contrast index, wherein the degradation type vector includes a blur component, a contrast component and a noise component; the degradation type vector is weightedly combined with a preset weight coefficient to obtain an image degradation score; and the environmental parameters in the environmental state matrix are combined with a preset influence function and a weight coefficient to calculate an environmental influence factor.

[0021] Exemplarily, the new energy station is subjected to Voronoi partitioning, and the partitioning rules are as follows: taking the station boundary as a constraint, selecting multiple initial points as seed points, and using Euclidean distance as a metric, the space is divided into several polygonal areas. The distance from any point in each area to the seed point of the area is less than the distance to the seed points of other areas. An environmental sensor array is arranged at the center of each Voronoi area, and each sensor array includes a light intensity sensor, a temperature and humidity sensor, an air quality sensor, and a visibility sensor. Figure 2As shown in the schematic diagram of the optimization of the layout of environmental sensors in new energy stations based on the Voronoi rule, the solution of the present invention reduces the number of sensors by 28%, increases the coverage by 10%, and significantly improves the boundary adaptability.

[0022] The collected environmental parameter data is preprocessed and the sliding average filtering method is used to eliminate outliers and noise. Specifically, the sliding window size is set to 5 minutes, and the average value of the data in the window is taken as the smoothing result at that moment. For example, if the temperature data collected by a sensor within 5 minutes is [25.3, 25.1, 25.4, 25.2, 25.3]℃, the smoothed temperature value is 25.26℃.

[0023] The spatial distribution and time series prediction are calculated based on the smoothed data. The spatial distribution uses the inverse distance weighted interpolation method, taking the sensor location as a known point to calculate the parameter value at any location. The time series prediction uses exponential smoothing prediction, and the prediction time is 30 minutes. The spatial distribution and time series prediction results are combined to form an environmental state matrix.

[0024] Calculate the dynamic coupling relationship between environmental parameters: Based on parameter similarity, calculate the Pearson correlation coefficient between any two parameters; based on the difference in change rate, calculate the Euclidean distance of the parameter change rate; based on the spatial adjacency relationship, calculate the physical distance between sensors. Combine the three indicators to calculate the coupling strength and form a dynamic coupling matrix.

[0025] The image quality index is calculated based on the environmental parameters and the dynamic coupling matrix. The time decay weighting method is adopted, and the weight of recent data is 0.6 and the weight of long-term data is 0.4. The calculation results are exponentially smoothed to obtain the image blur index and contrast index. For example, the calculated blur at a certain moment is 0.35 and the contrast is 0.72.

[0026] Construct a degradation type vector, including blur, contrast and noise components. Combine the degradation type vector with the preset weights [0.4, 0.4, 0.2] to get the image degradation score. When calculating the environmental impact factor, combine the environmental parameters with the impact function and weights to get a normalized result between 0 and 1.

[0027] The present invention realizes uniform collection of environmental parameters by arranging sensor arrays according to the Voronoi rule, improves the spatial representativeness and integrity of data, and provides a reliable data basis for subsequent analysis; calculates image degradation scores based on the dynamic coupling relationship of environmental parameters, takes into account the mutual influence between parameters, makes the scoring results more accurate and reasonable, and can truly reflect the degree of influence of the environment on image quality; uses time attenuation weighting and smoothing calibration methods to process data, improves the stability and reliability of calculation results, reduces the interference of abnormal fluctuations on evaluation results, and makes environmental impact assessment more objective and accurate.

[0028] In an optional implementation, dynamically adjusting image enhancement parameters according to the environmental impact factors, selecting a noise removal strategy based on an image degradation score, and performing adaptive histogram equalization and image normalization processing include: Establishing a mapping relationship between environmental impact factors and image enhancement parameters, wherein the coefficients of the mapping relationship are dynamically adjusted based on the environmental state matrix; constructing an enhancement operator library including a contrast enhancement operator and a sharpening enhancement operator, wherein the parameters of the contrast enhancement operator are adaptively adjusted according to a mapping rule between light intensity and contrast, and the parameters of the sharpening enhancement operator are dynamically adjusted according to an image blur index; Constructing an image quality evaluation function, the evaluation function including a peak signal-to-noise ratio index, a structural similarity index and a no-reference image quality evaluation index, the weight of the evaluation function being adaptively adjusted according to environmental influencing factors, and optimizing image enhancement parameters based on the evaluation function; Based on the image degradation score, noise type is identified, and a corresponding noise removal strategy is selected according to the noise type, including multi-scale noise analysis and adaptive filtering processing, and parameters of the filtering processing are dynamically adjusted according to noise characteristics; Perform adaptive histogram equalization on the filtering results, including adaptive blocking based on local entropy value and histogram equalization parameter adjustment. The histogram equalization parameters include equalization strength and block fusion weight. The histogram equalization parameters are dynamically adjusted according to environmental influencing factors. The equalization result is normalized based on the time series statistical features, wherein the time series statistical features include the mean distribution and the variance distribution of the image sequence, and the scaling parameters and the translation parameters of the normalization are dynamically adjusted according to the time series prediction result.

[0029] For example, a mapping relationship between environmental impact factors and image enhancement parameters is established. The mapping coefficients are dynamically adjusted through the environmental state matrix. For example, when the light intensity is 200 lux, the contrast enhancement coefficient is set to 1.2 and the sharpening enhancement coefficient is set to 0.8; when the light intensity drops to 50 lux, the contrast enhancement coefficient is adjusted to 1.5 and the sharpening enhancement coefficient is adjusted to 1.2.

[0030] Build an enhancement operator library, including contrast enhancement and sharpening enhancement operators. The contrast enhancement operator adaptively adjusts parameters according to the light intensity, such as setting the gamma value to 0.8 when the light intensity is 100 lux; the sharpening enhancement operator dynamically adjusts according to the blur index, and the sharpening radius is set to 3 pixels when the blur is 0.6.

[0031] An image quality evaluation function is established, including peak signal-to-noise ratio, structural similarity, and no-reference quality evaluation indicators. The weight is adjusted according to environmental factors, such as the structural similarity weight is 0.5 when there is sufficient light, and it is adjusted to 0.3 when there is insufficient light. The enhancement parameters are optimized based on the evaluation function to maximize the evaluation score.

[0032] The noise type is identified and the noise removal strategy is selected according to the degradation score. For example, Gaussian noise uses multi-scale analysis and adaptive Gaussian filtering, and salt and pepper noise uses median filtering. The filtering parameters are adjusted according to the noise characteristics. For example, when the noise variance is 0.02, the Gaussian filter window is set to 5×5.

[0033] Adaptive histogram equalization is performed on the filtering results. Adaptive block division is performed based on the local entropy value. The block size of the area with high entropy value is set to 16×16, and the area with low entropy value is set to 32×32. The equalization parameters are adjusted according to environmental factors. For example, when the lighting is uneven, the equalization intensity is 1.5 and the block fusion weight is 0.7.

[0034] Finally, normalization is performed based on the time series statistical features. The mean and variance distribution of the image sequence are analyzed to predict the statistical features of the next frame. The normalization parameters are dynamically adjusted according to the prediction results. For example, when the mean prediction value is 128, the scaling factor is set to 1.2 and the translation is set to 10.

[0035] The present invention dynamically adjusts enhancement parameters through environmental influencing factors, establishes an adaptive mapping relationship, improves the environmental adaptability and robustness of image enhancement, and makes the enhancement effect more stable and reliable; selects a noise removal strategy based on image degradation score, adopts multi-scale analysis and adaptive filtering, improves the accuracy of noise removal, and effectively maintains image detail information; performs adaptive histogram equalization and normalization processing, dynamically adjusts parameters in combination with time series statistical characteristics, improves image contrast and brightness uniformity, and enhances image visual quality.

[0036] In an optional implementation, the deep convolutional neural network of the feature extraction module includes: A main network and an auxiliary branch network, wherein the main network adopts a dense connection structure, including four dense blocks, and the number of convolution kernels in each dense block is adaptively adjusted according to the complexity of the image; the auxiliary branch network adopts a lightweight structure and receives an environment state matrix as input; A channel attention module is set after each dense block of the backbone network, and the weight coefficient of the channel attention module is modulated by the environmental feature vector output by the auxiliary branch network; a feature recalibration unit is set between adjacent dense blocks, and the feature recalibration unit dynamically adjusts the scale parameter of the feature map based on the image quality score of the local area; At the end of the backbone network, the global average pooling layer and the fully connected layer are connected in series to output the image feature vector.

[0037] Exemplarily, the feature extraction module adopts a deep convolutional neural network architecture, which includes two parts: a backbone network and an auxiliary branch network. The backbone network adopts a densely connected structure and consists of four dense blocks. Each dense block contains multiple convolutional layers, and dense connections are used between adjacent layers. Specifically, the input of each layer in a dense block contains the feature maps of all previous layers. The number of convolution kernels of the first dense block is set to 32, and the number of convolution kernels of subsequent dense blocks is adaptively adjusted according to the complexity of the input image, and the complexity is calculated by the gradient information and texture features of the image.

[0038] The auxiliary branch network adopts a lightweight structure, including 3 convolutional layers and 1 fully connected layer. The convolution kernel size of the first convolutional layer is 3×3, and the number of convolution kernels is 16; the convolution kernel size of the second convolutional layer is 3×3, and the number of convolution kernels is 32; the convolution kernel size of the third convolutional layer is 1×1, and the number of convolution kernels is 64. The auxiliary branch network receives a 256×256 environment state matrix as input and outputs a 128-dimensional environment feature vector.

[0039] A channel attention module is set after each dense block to modulate the channel weight of the feature map. The channel attention module first performs global average pooling on the feature map to obtain the channel descriptor, then fuses the channel descriptor with the environmental feature vector output by the auxiliary branch network, and finally obtains the weight coefficient of each channel through a fully connected layer. The weight coefficient ranges from 0 to 1 and is used to weight each channel of the feature map.

[0040] A feature recalibration unit is set between adjacent dense blocks. The unit first divides the feature map into multiple local regions and calculates the image quality score of each region. The quality score is calculated based on the clarity, contrast and other features in the region, and the value range is between 0 and 1. Then the scale parameter of the feature map is dynamically adjusted according to the quality score. The higher the quality score, the larger the scale parameter of the region.

[0041] The global average pooling layer and the fully connected layer are connected in series at the end of the backbone network. The global average pooling layer compresses the spatial dimension of the feature map and outputs a 512-dimensional feature vector. The fully connected layer maps the feature vector to a 256-dimensional image feature space. The final output 256-dimensional image feature vector contains the key visual information of the input image.

[0042] The present invention enhances the expressiveness of feature extraction through the design of dense connection structure and adaptive number of convolution kernels, introduces auxiliary branch network and channel attention mechanism, realizes dynamic modulation of feature extraction process by environmental information, enables the extracted features to better adapt to different environmental conditions, improves the environmental adaptability of the features, adopts feature recalibration unit to perform quality-aware dynamic adjustment of feature maps, enhances the feature expression of high-quality areas, suppresses interference in low-quality areas, and improves the accuracy and reliability of feature extraction.

[0043] In an optional implementation, the temporal feature extraction unit extracts temporal correlation features of continuous frame images using a recurrent neural network, the feature stability evaluation unit calculates the stability score of the feature based on a multi-dimensional similarity metric and an optimal transmission distance, and the dynamic calibration unit adaptively adjusts the current feature based on the stability score through a Squeeze-and-Excitation modulation layer and a residual nonlinear transformation, including: The image feature vector is constructed as a feature sequence of continuous frame images, and the feature sequence is input into a bidirectional long short-term memory network, wherein the output features of the forward long short-term memory unit and the backward long short-term memory unit are concatenated to obtain a time series correlation feature; Calculate the local stability score and the global stability score of the time series correlation feature, the local stability score is obtained by constructing a directed graph structure and calculating the edge weight based on a multi-dimensional similarity metric and an optimal transmission distance, the global stability score is calculated by an entropy complexity index of a multi-scale feature sequence, and the local stability score and the global stability score are fused through an adaptive weight network to obtain a feature stability score; The time series correlation feature is calibrated based on the stability score, wherein the calibration generates attention weights through a Squeeze-and-Excitation modulation layer in a dynamic calibration unit, and the modulated features are calibrated using a residual nonlinear transformation, and an optimal historical feature sequence is selected based on dynamic programming, and a comparison learning sample pair is constructed between the enhanced view of the time series correlation feature and the enhanced view of the historical feature, and a calibration feature is obtained through multi-objective optimization; Performing exponential sliding average on the calibration feature to obtain a smooth feature, wherein the smoothing coefficient of the exponential sliding average is dynamically adjusted according to the stability score; inputting the smooth feature into a classifier to obtain device working status information; The Euclidean distance between the smoothing feature and the time series correlation feature is calculated, the Euclidean distance is multiplied by the complementary value of the stability score to obtain an abnormality score, an adaptive threshold is constructed based on the statistical characteristics of the abnormality score, and abnormal warning information is generated when the abnormality score exceeds the adaptive threshold.

[0044] Exemplarily, a feature vector sequence of continuous frame images is obtained. For each frame of image, a 2048-dimensional feature vector is extracted by a pre-trained convolutional neural network. The feature vectors of 32 consecutive frames of images are arranged in time sequence to form a feature sequence. The feature sequence is input into a bidirectional long short-term memory network, which includes two layers of long short-term memory units and a hidden layer dimension of 512. The output features of the forward and backward long short-term memory units are concatenated to obtain a 1024-dimensional temporal correlation feature.

[0045] Evaluate the stability of time series correlation features. First, construct a directed graph structure in which nodes represent each feature vector in the feature sequence. Calculate the multi-dimensional similarity between adjacent nodes, including cosine similarity, Euclidean distance, and Pearson correlation coefficient. Construct an edge weight matrix based on the similarity measure. Use the Sinkhorn algorithm to calculate the optimal transmission distance between nodes and obtain a local stability score. At the same time, perform multi-scale decomposition on the feature sequence, calculate the sample entropy and approximate entropy of the sequence at different scales, and obtain a global stability score. Adaptively fuse the local and global stability scores through a three-layer fully connected network, and output a feature stability score between 0 and 1.

[0046] The temporal correlation features are calibrated based on the stability score. The squeeze-and-excitation modulation layer first performs global average pooling on the features to obtain the channel attention weights. The weights are nonlinearly transformed using a two-layer fully connected network to obtain the modulation coefficient. The modulation coefficient is multiplied by the original feature to achieve feature modulation. The modulated features are then nonlinearly transformed using a two-layer fully connected network with residual connections. The most similar features are selected as references from the historical feature sequence based on dynamic programming. Data enhancement is performed on the current features and the historical reference features respectively to construct contrastive learning sample pairs. The calibrated features are obtained by multi-objective optimization that minimizes the contrast loss and reconstruction loss.

[0047] The exponential sliding average is performed on the calibration features to obtain smooth features. The smoothing coefficient is dynamically adjusted according to the stability score. The higher the stability score, the larger the smoothing coefficient. The smoothed features are input into the fully connected classifier to output the probability distribution of the device working state.

[0048] Calculate the Euclidean distance between the smoothing feature and the time series correlation feature, and multiply the distance with the complement of the stability score to get the anomaly score. Build an adaptive threshold based on the mean and standard deviation of the anomaly score. When the anomaly score exceeds the adaptive threshold, generate an anomaly warning message.

[0049] The present invention extracts temporal correlation features through a bidirectional long short-term memory network, which can effectively capture the temporal dependency between consecutive frame images and improve the expressiveness of features; based on multi-dimensional similarity measurement and optimal transmission distance to evaluate feature stability, combined with local and global stability scores, it can comprehensively measure the reliability of features and provide a reliable basis for feature calibration; a dynamic calibration unit is used to adaptively adjust features, and the robustness of features is improved through comparative learning and multi-objective optimization, and anomalies are detected in time in combination with anomaly detection mechanism to ensure the reliability and security of the system.

[0050] In an optional implementation, a local stability score and a global stability score of the time series correlation feature are calculated, the local stability score is obtained by constructing a directed graph structure and calculating edge weights based on a multi-dimensional similarity metric and an optimal transmission distance, the global stability score is calculated by an entropy complexity index of a multi-scale feature sequence, and the feature stability score is obtained by fusing the local stability score and the global stability score through an adaptive weight network, including: The temporal correlation features within the time window are constructed into a directed graph structure, and the edge weights of the directed graph structure are calculated by multi-dimensional similarity measurement; The multi-dimensional similarity measurement includes: constructing a multi-level cost matrix between feature distributions, the multi-level cost matrix includes a point-to-point cost matrix based on local features and a structural cost matrix based on global features; solving the optimal transmission problem based on the Sinkhorn algorithm to obtain a transmission scheme, the transmission scheme satisfies the first-order moment matching constraint and the second-order moment correlation constraint; calculating the multi-scale optimal transmission distance and obtaining the transmission distance through a scale-adaptive fusion network; and performing a weighted combination of the transmission distance and the mutual information through a parameterized sigmoid mapping function to obtain the edge weight; A transfer matrix is ​​constructed based on the edge weights, the importance of nodes in the directed graph structure is obtained through iterative calculation, and the product of the node importance and the edge weight is normalized to obtain a local stability score; Constructing a multi-scale feature sequence of the time series correlation feature, calculating the sample entropy of the multi-scale feature sequence, constructing an entropy complexity plane based on the sample entropy to obtain a complexity index, and calculating a global stability score based on the complexity index; A feature state vector including the local stability score, the global stability score and their time difference is constructed, the feature state vector is input into an adaptive weight network with a Softmax activation function to obtain a fusion weight, and the local stability score and the global stability score are combined by the fusion weight to obtain a stability score of the feature.

[0051] Exemplarily, for the calculation of stability scores of time-series correlation features, a directed graph structure is constructed and the local stability score is calculated. The time-series correlation features within the time window are taken as nodes in the graph, and the edge weights between nodes are calculated by multi-dimensional similarity measurement. When calculating the edge weights, a multi-level cost matrix between feature distributions is constructed, including a point-to-point cost matrix based on local features and a structural cost matrix based on global features. The point-to-point cost matrix is ​​obtained by calculating the Euclidean distance between feature vectors, and the structural cost matrix is ​​obtained by calculating the distance between the covariance matrices of feature distributions. The Sinkhorn algorithm is used to solve the optimal transmission problem, and the transmission scheme must satisfy the first-order moment matching constraint and the second-order moment correlation constraint. In specific implementation, the transmission matrix is ​​first initialized, and then the row normalization and column normalization operations are iteratively performed until convergence. The optimal transmission distance is calculated at different scales, and the final transmission distance is obtained through a scale-adaptive fusion network. The transmission distance is weightedly combined with the mutual information through the sigmoid mapping function to obtain the edge weight. The kernel density estimation method is used when calculating the mutual information, and the Gaussian kernel function is used to estimate the probability distribution of the feature, and the amount of information between the marginal distribution and the joint distribution is calculated by a numerical method.

[0052] The transfer matrix is ​​constructed based on the edge weights, and the importance of the nodes is calculated by the power iteration method. The product of the node importance and the edge weight is normalized to obtain the local stability score. For example, for a directed graph containing 10 nodes, a stable node importance distribution can be obtained through 20 iterations.

[0053] For the calculation of the global stability score, a multi-scale feature sequence is first constructed. The feature sequence is extracted with different window sizes such as 60 seconds, 300 seconds, and 900 seconds using the sliding window method. The sample entropy of the feature sequence at each scale is calculated, and the entropy complexity plane is constructed based on the sample entropy. When calculating the sample entropy, the similarity pattern of the feature sequence under different template dimensions is mainly considered. In the specific implementation, the appropriate template dimension and similarity threshold are selected, the number of similar templates in the sequence is counted, and the sample entropy is calculated by the ratio of the number of similar templates under adjacent dimensions. The complexity index is obtained by calculating the area under the curve in the entropy complexity plane, and then the global stability score is calculated.

[0054] Construct a feature state vector, including the local stability score, the global stability score and their time difference. Input the feature state vector into the adaptive weight network with the Softmax activation function to obtain the fusion weight. Combine the local stability score and the global stability score through the fusion weight to obtain the final feature stability score. For the implementation of the adaptive weight network, a three-layer fully connected network structure is adopted. The input layer dimension is the same as the feature state vector dimension. The hidden layer uses the ReLU activation function. The output layer uses the Softmax activation function to obtain the fusion weight. The cross entropy loss function is used for network training. The initial value of the learning rate is set to 0.001, and it decays to the original 0.1 every 100 rounds.

[0055] The existing technologies generally have problems such as single feature similarity measurement, insufficient utilization of local-global correlation of time series features, and insufficient adaptability of evaluation results. The feature stability evaluation scheme of the present invention creatively introduces the optimal transmission theory into the feature similarity measurement, and designs a multi-level cost matrix to simultaneously consider local point pair similarity and global structural similarity, making feature comparison more comprehensive and accurate. An innovative two-layer evaluation mechanism is proposed, which organically combines the local stability evaluation based on graph structure and the global stability evaluation based on entropy complexity, so as to better capture the multi-scale dynamic characteristics of the features. At the same time, by introducing an adaptive weight network to dynamically adjust the fusion ratio of local and global stability scores, the adaptability of the evaluation results is significantly improved. Figure 3 The performance comparison of different similarity measurement methods in the process of feature change is demonstrated, and the advantages of the multi-dimensional similarity measurement method (calculated by optimal transmission distance) proposed in the present invention in feature similarity judgment are intuitively demonstrated, especially at the key distinguishing points (0.65-0.78), which show obvious sensitivity changes and can accurately distinguish normal feature fluctuations from abnormal changes, thus realizing the accurate similarity measurement required by the present invention of "obtaining the edge weight by weighted combination of parameterized sigmoid mapping function and mutual information".

[0056] In an optional implementation, the time series association feature is calibrated based on the stability score, the calibration generates attention weights through a Squeeze-and-Excitation modulation layer in a dynamic calibration unit, the modulated features are calibrated using a residual nonlinear transformation, the optimal historical feature sequence is selected based on dynamic programming, and a comparison learning sample pair is constructed between the enhanced view of the time series association feature and the enhanced view of the historical feature, and the calibration feature is obtained through multi-objective optimization, including: A Squeeze-and-Excitation modulation layer is constructed in the dynamic calibration unit, global average pooling is performed on the temporal correlation features, and attention weights are generated through a feedforward network, the attention weights are adjusted based on the stability score, and the temporal correlation features are modulated with the attention weights to obtain modulation features; The modulation feature is subjected to a residual nonlinear transformation to obtain an initial calibration feature, wherein the residual nonlinear transformation includes passing the modulation feature through a first transformation layer with a ReLU activation function and a second transformation layer with a tanh activation function, multiplying the transformation result with a sigmoid output of a feature stability score and then adding the result to the modulation feature; Construct a self-similarity matrix of the time series correlation feature, the self-similarity matrix is ​​calculated by combining the cosine similarity and Gaussian kernel similarity of the feature vector, calculate the state score between the current moment and the historical moment, construct a state transfer equation based on the state score to calculate the optimal state value, the calculation range of the state transfer equation is N moments before the current moment, backtrack from the optimal state value to obtain the feature selection path and extract the corresponding historical features, where the formula for the state score is: ; Where t is the current moment, k is the historical moment, |tk| is the time interval between two moments, and f t is the feature vector at the current time t, f k is the feature vector of historical moment k, sim cos (f k， f t ) is the feature vector f k and f t The cosine similarity between ||f k -f t || 2 is the feature vector f k and f t The square of the Euclidean distance, σ is the bandwidth parameter of the Gaussian kernel, λ is the time decay factor, the larger the value, the faster the time decay; Generate enhanced views of the initial calibration feature and the historical feature respectively through random mask and Gaussian noise, construct different enhanced views of the same feature as a positive sample pair, construct the enhanced views of the initial calibration feature and the historical feature as a negative sample pair, and calculate the contrast loss through the InfoNCE loss function; The initial calibration feature is fused with the weighted average of the historical features to obtain the calibration feature, a multi-objective loss function including mean square error, contrast loss and time series smoothing regularization term is constructed, and the parameters of the dynamic calibration unit are optimized using a cosine annealing learning rate strategy.

[0057] Exemplarily, the dynamic calibration unit first constructs a Squeeze-and-Excitation modulation layer and performs a global average pooling operation on the input temporal correlation features. Specifically, for the input 256-dimensional feature vector, a 1-dimensional feature representation is obtained by global average pooling. The feature is then input into a two-layer feedforward network. The first layer uses 16 hidden units, and the second layer outputs a 256-dimensional vector with the same dimension as the input feature as the attention weight. The attention weight is adjusted based on the stability score. The higher the stability score, the greater the attention weight. The adjusted attention weight is element-wise multiplied with the original feature to obtain the modulation feature.

[0058] The modulation features are transformed by residual nonlinearity. First, the 256-dimensional features are mapped to 512-dimensional space through a fully connected layer with ReLU activation, and then mapped back to 256 dimensions through a fully connected layer with tanh activation. The transformation result is multiplied by the sigmoid output of the feature stability score and added to the modulation feature to obtain the initial calibration feature.

[0059] The self-similarity matrix of the time series correlation features is constructed by calculating the cosine similarity and Gaussian kernel similarity between the feature vectors. For the feature vectors of any two moments, the cosine similarity is first calculated, and then the square of the Euclidean distance is calculated and converted into similarity through the Gaussian kernel function. The two similarities are weighted and combined, and the time decay factor is considered to obtain the state score. The state transition equation is constructed based on the state score, and dynamic programming is performed within the first 30 moments to select the optimal historical feature sequence.

[0060] Enhanced views are generated for the initial calibration features and historical features. 30% of the feature elements are randomly set to zero as mask enhancement, and Gaussian noise with a mean of 0 and a standard deviation of 0.1 is added as noise enhancement. Different enhanced views of the same feature are constructed as positive sample pairs, and enhanced views of calibration features and historical features are constructed as negative sample pairs. InfoNCE loss is used to calculate the contrast loss, and the temperature parameter is set to 0.07.

[0061] The final calibration features are obtained by weighted averaging the initial calibration features and historical features. The weight of the historical features is determined based on the similarity score. A multi-objective loss function is constructed, including the mean square error between the predicted value and the true value, the contrast loss, and the time series smoothing regularization term. The learning rate is adjusted using cosine annealing, with an initial learning rate of 0.001, a minimum learning rate of 0.00001, and a total of 100 training rounds.

[0062] Traditional feature calibration methods mainly focus on the optimization of single-frame features and fail to fully utilize the temporal correlation information of features. The dynamic feature calibration scheme proposed in this invention realizes the adaptive modulation of features through the Squeeze-and-Excitation structure, and innovatively introduces stability scores to dynamically adjust the attention weights, making the calibration process more targeted. A historical feature selection strategy based on dynamic programming is designed, which realizes the adaptive selection of the optimal historical feature sequence by comprehensively considering feature similarity and time decay factors. At the same time, contrastive learning is innovatively introduced into the calibration process, and the discriminative ability of features is improved by constructing multi-view positive and negative sample pairs.

[0063] The present invention forms a unified feature calibration framework by organically combining attention mechanism, dynamic programming and contrastive learning. This multi-level calibration mechanism can not only make full use of time series information, but also adaptively balance multiple performance indicators of features, providing a new idea for solving feature optimization problems in similar scenarios.

[0064] In an optional implementation, establishing an associative mapping relationship between the data in the environment state matrix and the image quality degradation type includes: The dynamic coupling relationship between environmental parameters is calculated to obtain the dynamic coupling matrix C(t), which is calculated as follows: ; Among them, C ij (t) is the coupling strength between the i-th environmental parameter and the j-th environmental parameter at time t; E i (t) is the value of the i-th environmental parameter at time t; E j (t) is the jth environmental parameter value at time t; ψ(E i (t), E j (t)) is the parameter similarity function; τ(t) is the time-varying attenuation coefficient, which controls the influence of the rate of change; ΔE ij (t) is the difference in parameter change rate, which is the difference in the speed of change of the two parameters; S ij (t) is the spatial correlation coefficient, which reflects the correlation degree of the parameters in spatial distribution; Based on the environmental parameters and the dynamic coupling matrix, the image quality index Q(t) is calculated, and its calculation formula is: ; Among them, t is the time stamp of the current moment; k is the historical moment index, from 1 to N, where N is the size of the historical time window considered; ω k (t) is the adaptive weight coefficient; E(t) is the environmental parameter vector at the current moment; C(t) is the dynamic coupling matrix at the current moment; f kis the mapping function between environmental parameters and quality indicators; λ is the time decay factor; |tk| is the time interval between the current moment and the historical moment; The image quality index is time-series smoothed and calibrated, and the statistical characteristics of the quality index at the latest N moments are calculated; the current quality index is weighted smoothed based on the statistical characteristics; the smoothing result is calibrated according to historical observation data to obtain a blur index and a contrast index.

[0065] Exemplarily, environmental state matrix data is obtained, including multiple environmental parameters such as temperature, humidity, and light. For each environmental parameter at each moment, the dynamic coupling relationship between the parameters is calculated. Specifically, the coupling strength between the parameters is characterized by calculating parameter similarity, change rate difference, and spatial correlation. Parameter similarity is evaluated based on the degree of closeness of the values, the change rate difference reflects the different speeds of parameter change, and the spatial correlation indicates the degree of correlation of the parameter distribution. The time-varying attenuation coefficient is introduced to adjust the influence weight of the change rate.

[0066] Taking temperature and humidity as an example, when the temperature is 25°C and the humidity is 60%, the normalized parameter similarity is first calculated to be 0.85; the temperature change rate is 2°C / hour, the humidity change rate is 5% / hour, and the change rate difference is 3; the spatial correlation coefficient is 0.7. Combined with the time-varying attenuation coefficient of 0.8, the final coupling strength between temperature and humidity is 0.65. Similar calculations are performed to obtain the coupling relationship between all parameters to form a dynamic coupling matrix.

[0067] The image quality index is calculated based on the environmental parameters and the dynamic coupling matrix. The environmental state within the historical time window is considered, and the influence of different moments is weighted by the adaptive weight coefficient. The weight coefficient decays exponentially with the time interval, and the decay factor is 0.9. For each historical moment, the quality index is calculated by the mapping function based on the environmental parameters and coupling matrix at that time. Taking 10 historical moments as an example, the quality index of the current moment is obtained by weighted accumulation.

[0068] Perform time series smoothing and calibration on the quality indicators. Count the mean, variance and other characteristics of the quality indicators for the last 10 moments. Smooth the current indicator based on the inverse of the variance as the weight. Establish a calibration model based on historical observation data and map the smoothing results to standardized ambiguity and contrast indicators. The calibrated ambiguity indicator ranges from 0 to 1, and the contrast indicator ranges from 0 to 100.

[0069] Traditional methods are difficult to accurately characterize the dynamic coupling relationship between multiple environmental parameters, and are unable to effectively establish a mapping relationship between environmental changes and image quality degradation. Figure 4This is a graph showing the evolution of the dynamic coupling matrix of environmental parameters of the present invention over time, which shows the 24-hour change process of the coupling strength of key parameter pairs (temperature-humidity, light-temperature, humidity-visibility) in the dynamic coupling matrix of environmental parameters. The black solid line dots represent the temperature-humidity coupling strength, the dark gray dotted line square dots represent the light-temperature coupling strength, and the light gray dotted line triangle dots represent the humidity-visibility coupling strength. As shown in the figure, the present invention can effectively identify environmental mutation events by accurately depicting the change pattern of the coupling relationship between parameters over time. The figure clearly shows the differences in the coupling characteristics of each parameter pair at different times of the day (such as early morning, noon, and evening), so that the system can adjust adaptively.

[0070] The present invention innovatively proposes a sensor layout scheme based on the Voronoi rule and an environmental parameter correlation analysis method of the dynamic coupling matrix. By comprehensively considering parameter similarity, rate of change differences and spatial correlation, accurate modeling of complex relationships between environmental parameters is achieved. At the same time, an adaptive weight mechanism with time decay is designed to improve the temporal continuity of the evaluation results and the ability to respond to environmental mutations. The present invention adopts a multi-level computing framework. A data preprocessing method based on sliding average is designed at the bottom layer to improve the reliability of the original data. In the middle layer, the complex correlation between environmental parameters is captured by the dynamic coupling matrix, and a complete description of the environmental state is constructed using the spatiotemporal interpolation method. In the upper layer, the final image degradation score is obtained by multi-index fusion, realizing end-to-end mapping from environmental monitoring to quality assessment.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for improving the image recognition accuracy of a new energy station, characterized in that: include: Acquire monitoring image data of new energy stations; collect environmental parameter data based on the sensor array arranged according to the Voronoi rule, and calculate image degradation scores and environmental impact factors through the dynamic coupling relationship of environmental parameters; Build a deep learning model including preprocessing module, feature extraction module and feature optimization module; The preprocessing module dynamically adjusts image enhancement parameters according to the environmental influencing factors, selects a noise removal strategy based on the image degradation score, and performs adaptive histogram equalization and image normalization processing; The feature extraction module uses a deep convolutional neural network to extract features from the preprocessed monitoring image data to obtain an image feature vector; The feature optimization module processes the image feature vector to obtain the recognition result of the new energy power generation equipment in the new energy station, wherein the time series feature extraction unit uses a recurrent neural network to extract the time series correlation features of continuous frame images, the feature stability evaluation unit calculates the stability score of the feature based on the multi-dimensional similarity measurement and the optimal transmission distance, and the dynamic calibration unit adaptively adjusts the current feature through the Squeeze-and-Excitation modulation layer and the residual nonlinear transformation based on the stability score; the recognition result includes equipment working status information and abnormal warning information.

2. The method according to claim 1, characterized in that The sensor array arranged based on the Voronoi rule collects environmental parameter data, and the image degradation score and environmental impact factors are calculated through the dynamic coupling relationship of environmental parameters, including: The new energy stations are divided based on the Voronoi partitioning rule, and an environmental sensor array is arranged at the center of the divided area. The environmental sensor array collects light intensity, temperature, humidity, air quality and visibility data; the collected environmental parameter data is preprocessed to obtain smooth data; Calculate the spatial distribution estimation result and the time series prediction result based on the smoothed data and combine them to form an environmental state matrix; Establishing an associative mapping relationship between the data in the environmental state matrix and the image quality degradation type respectively, including: calculating the dynamic coupling relationship between environmental parameters, calculating the coupling strength between the light intensity, temperature, humidity, air quality and visibility parameters based on parameter similarity, change rate difference and spatial adjacency, and forming a dynamic coupling matrix; calculating the image quality index by a time attenuation weighting method based on the current environmental parameters and the dynamic coupling matrix, and performing time series smoothing and calibration on the image quality index to obtain an image blur index and an image contrast index; A degradation type vector is constructed based on the image blur index and the image contrast index, wherein the degradation type vector includes a blur component, a contrast component and a noise component; the degradation type vector is weightedly combined with a preset weight coefficient to obtain an image degradation score; and the environmental parameters in the environmental state matrix are combined with a preset influence function and a weight coefficient to calculate an environmental influence factor.

3. The method according to claim 2, characterized in that Dynamically adjusting image enhancement parameters according to the environmental influencing factors, selecting a noise removal strategy based on the image degradation score, and performing adaptive histogram equalization and image normalization processing include: Establishing a mapping relationship between environmental impact factors and image enhancement parameters, wherein the coefficients of the mapping relationship are dynamically adjusted based on the environmental state matrix; constructing an enhancement operator library including a contrast enhancement operator and a sharpening enhancement operator, wherein the parameters of the contrast enhancement operator are adaptively adjusted according to a mapping rule between light intensity and contrast, and the parameters of the sharpening enhancement operator are dynamically adjusted according to an image blur index; Constructing an image quality evaluation function, the evaluation function including a peak signal-to-noise ratio index, a structural similarity index and a no-reference image quality evaluation index, the weight of the evaluation function being adaptively adjusted according to environmental influencing factors, and optimizing image enhancement parameters based on the evaluation function; Based on the image degradation score, noise type is identified, and a corresponding noise removal strategy is selected according to the noise type, including multi-scale noise analysis and adaptive filtering processing, and parameters of the filtering processing are dynamically adjusted according to noise characteristics; Perform adaptive histogram equalization on the filtering results, including adaptive blocking based on local entropy value and histogram equalization parameter adjustment. The histogram equalization parameters include equalization strength and block fusion weight. The histogram equalization parameters are dynamically adjusted according to environmental influencing factors. The equalization result is normalized based on the time series statistical features, wherein the time series statistical features include the mean distribution and the variance distribution of the image sequence, and the scaling parameters and the translation parameters of the normalization are dynamically adjusted according to the time series prediction result.

4. The method according to claim 1, characterized in that: The deep convolutional neural network of the feature extraction module includes: A main network and an auxiliary branch network, wherein the main network adopts a dense connection structure, including four dense blocks, and the number of convolution kernels in each dense block is adaptively adjusted according to the complexity of the image; the auxiliary branch network adopts a lightweight structure and receives an environment state matrix as input; A channel attention module is set after each dense block of the backbone network, and the weight coefficient of the channel attention module is modulated by the environmental feature vector output by the auxiliary branch network; a feature recalibration unit is set between adjacent dense blocks, and the feature recalibration unit dynamically adjusts the scale parameter of the feature map based on the image quality score of the local area; At the end of the backbone network, the global average pooling layer and the fully connected layer are connected in series to output the image feature vector.

5. The method according to claim 1, characterized in that The temporal feature extraction unit uses a recurrent neural network to extract the temporal correlation features of continuous frame images. The feature stability evaluation unit calculates the stability score of the feature based on multi-dimensional similarity measurement and optimal transmission distance. The dynamic calibration unit adaptively adjusts the current feature based on the stability score through the Squeeze-and-Excitation modulation layer and residual nonlinear transformation, including: The image feature vector is constructed as a feature sequence of continuous frame images, and the feature sequence is input into a bidirectional long short-term memory network, wherein the output features of the forward long short-term memory unit and the backward long short-term memory unit are concatenated to obtain a time series correlation feature; Calculate the local stability score and the global stability score of the time series correlation feature, the local stability score is obtained by constructing a directed graph structure and calculating the edge weight based on a multi-dimensional similarity metric and an optimal transmission distance, the global stability score is calculated by an entropy complexity index of a multi-scale feature sequence, and the local stability score and the global stability score are fused through an adaptive weight network to obtain a feature stability score; The time series correlation feature is calibrated based on the stability score, wherein the calibration generates attention weights through a Squeeze-and-Excitation modulation layer in a dynamic calibration unit, and the modulated features are calibrated using a residual nonlinear transformation, and an optimal historical feature sequence is selected based on dynamic programming, and a comparison learning sample pair is constructed between the enhanced view of the time series correlation feature and the enhanced view of the historical feature, and a calibration feature is obtained through multi-objective optimization; Performing exponential sliding average on the calibration feature to obtain a smooth feature, wherein the smoothing coefficient of the exponential sliding average is dynamically adjusted according to the stability score; inputting the smooth feature into a classifier to obtain device working status information; The Euclidean distance between the smoothing feature and the time series correlation feature is calculated, the Euclidean distance is multiplied by the complementary value of the stability score to obtain an abnormality score, an adaptive threshold is constructed based on the statistical characteristics of the abnormality score, and abnormal warning information is generated when the abnormality score exceeds the adaptive threshold.

6. The method according to claim 5, characterized in that The local stability score and the global stability score of the time series correlation feature are calculated, the local stability score is obtained by constructing a directed graph structure and calculating the edge weight based on a multi-dimensional similarity metric and an optimal transmission distance, the global stability score is calculated by an entropy complexity index of a multi-scale feature sequence, and the feature stability score is obtained by fusing the local stability score and the global stability score through an adaptive weight network, including: The temporal correlation features within the time window are constructed into a directed graph structure, and the edge weights of the directed graph structure are calculated by multi-dimensional similarity measurement; The multi-dimensional similarity measurement includes: constructing a multi-level cost matrix between feature distributions, the multi-level cost matrix includes a point-to-point cost matrix based on local features and a structural cost matrix based on global features; solving the optimal transmission problem based on the Sinkhorn algorithm to obtain a transmission scheme, the transmission scheme satisfies the first-order moment matching constraint and the second-order moment correlation constraint; calculating the multi-scale optimal transmission distance and obtaining the transmission distance through a scale-adaptive fusion network; and performing a weighted combination of the transmission distance and the mutual information through a parameterized sigmoid mapping function to obtain the edge weight; A transfer matrix is ​​constructed based on the edge weights, the importance of nodes in the directed graph structure is obtained through iterative calculation, and the product of the node importance and the edge weight is normalized to obtain a local stability score; Constructing a multi-scale feature sequence of the time series correlation feature, calculating the sample entropy of the multi-scale feature sequence, constructing an entropy complexity plane based on the sample entropy to obtain a complexity index, and calculating a global stability score based on the complexity index; A feature state vector including the local stability score, the global stability score and their time difference is constructed, the feature state vector is input into an adaptive weight network with a Softmax activation function to obtain a fusion weight, and the local stability score and the global stability score are combined by the fusion weight to obtain a stability score of the feature.

7. The method according to claim 5, characterized in that The time series correlation feature is calibrated based on the stability score. The calibration generates attention weights through a Squeeze-and-Excitation modulation layer in a dynamic calibration unit, and the modulated features are calibrated using a residual nonlinear transformation. The optimal historical feature sequence is selected based on dynamic programming, and the enhanced view of the time series correlation feature and the enhanced view of the historical feature are used to construct a comparative learning sample pair. The calibration features obtained through multi-objective optimization include: A Squeeze-and-Excitation modulation layer is constructed in the dynamic calibration unit, global average pooling is performed on the temporal correlation features, and attention weights are generated through a feedforward network, the attention weights are adjusted based on the stability score, and the temporal correlation features are modulated with the attention weights to obtain modulation features; The modulation feature is subjected to a residual nonlinear transformation to obtain an initial calibration feature, wherein the residual nonlinear transformation includes passing the modulation feature through a first transformation layer with a ReLU activation function and a second transformation layer with a tanh activation function, multiplying the transformation result with a sigmoid output of a feature stability score and then adding the result to the modulation feature; Construct a self-similarity matrix of the time series correlation feature, the self-similarity matrix is ​​calculated by combining the cosine similarity and Gaussian kernel similarity of the feature vector, calculate the state score between the current moment and the historical moment, construct a state transfer equation based on the state score to calculate the optimal state value, the calculation range of the state transfer equation is N moments before the current moment, backtrack from the optimal state value to obtain the feature selection path and extract the corresponding historical features, where the formula for the state score is: ； Where t is the current moment, k is the historical moment, |tk| is the time interval between two moments, and f t is the feature vector at the current time t, f k is the feature vector of historical moment k, sim cos (f k， f t ) is the feature vector f k and f t The cosine similarity between ||f k -f t || 2 is the feature vector f k and f t The square of the Euclidean distance, σ is the bandwidth parameter of the Gaussian kernel, λ is the time decay factor, the larger the value, the faster the time decay; Generate enhanced views of the initial calibration feature and the historical feature respectively through random mask and Gaussian noise, construct different enhanced views of the same feature as a positive sample pair, construct the enhanced views of the initial calibration feature and the historical feature as a negative sample pair, and calculate the contrast loss through the InfoNCE loss function; The initial calibration feature is fused with the weighted average of the historical features to obtain the calibration feature, a multi-objective loss function including mean square error, contrast loss and time series smoothing regularization term is constructed, and the parameters of the dynamic calibration unit are optimized using a cosine annealing learning rate strategy.

8. The method according to claim 2, characterized in that: Establishing an associative mapping relationship between the data in the environment state matrix and the image quality degradation type includes: The dynamic coupling relationship between environmental parameters is calculated to obtain the dynamic coupling matrix C(t), which is calculated as follows: ； Among them, C ij (t) is the coupling strength between the i-th environmental parameter and the j-th environmental parameter at time t; E i (t) is the value of the i-th environmental parameter at time t; E j (t) is the jth environmental parameter value at time t; ψ(E i (t), E j (t)) is the parameter similarity function; τ(t) is the time-varying attenuation coefficient, which controls the influence of the rate of change; ΔE ij (t) is the difference in parameter change rate, which is the difference in the speed of change of the two parameters; S ij (t) is the spatial correlation coefficient, which reflects the correlation degree of the parameters in spatial distribution; Based on the environmental parameters and the dynamic coupling matrix, the image quality index Q(t) is calculated, and its calculation formula is: ； Among them, t is the time stamp of the current moment; k is the historical moment index, from 1 to N, where N is the size of the historical time window considered; ω k (t) is the adaptive weight coefficient; E(t) is the environmental parameter vector at the current moment; C(t) is the dynamic coupling matrix at the current moment; f k is the mapping function between environmental parameters and quality indicators; λ is the time decay factor; |tk| is the time interval between the current moment and the historical moment; The image quality index is time-series smoothed and calibrated, and the statistical characteristics of the quality index at the latest N moments are calculated; the current quality index is weighted smoothed based on the statistical characteristics; the smoothing result is calibrated according to historical observation data to obtain a blur index and a contrast index.

Images