53 results about "Vanishing gradient problem" patented technology

The invention belongs to the field of deep learning technology and remote sensing image segmentation, and particularly relates to a remote sensing image segmentation method based on Transsubnet edge information enhancement and multi-dimensional feature perception, and the method comprises the steps: S1, preparing a data set; s2, constructing remote sensing picture text description; s3, constructing and training a remote sensing image segmentation model; and S4, storing and testing the model. The invention designs a multi-modal feature extraction method based on parallelism of a sampling branch and a text feature extraction branch under edge feature compensation. The residual error mixing axial attention module is used for forming a transformer structure; and a text-picture multi-dimensional feature fusion enhancement module and a decoder part are embedded. According to the method, the ground feature identification capability can be improved through a text and picture multi-modal feature enhancement strategy, the segmentation boundary and small target object feature information is enhanced, more fine-grained features are reserved, the cross-regional long-distance dependency relationship is better captured, the common gradient disappearance problem in a deep network is relieved, and the method is suitable for large-scale popularization and application. And meanwhile, the small sample data set segmentation effect is improved.

The invention belongs to the field of audio processing, and particularly discloses an audio depression detection method and system based on an improved convolutional neural network autoencoder, and the method comprises the steps: carrying out the potential feature extraction and compression of input original audio data, extracting the data into a low-dimensional and dense potential feature vector, and carrying out the recognition of the potential feature vector; and the information loss in the reconstruction process is reduced. According to the model part, a residual block structure is introduced into an encoder part, input low-layer information is reserved through jump connection, the learning ability of the network is enhanced, and the gradient disappearance problem is avoided. A transpose convolution operation is introduced into a decoder part, the spatial resolution is increased through up-sampling, and a high-quality signal reconstruction task is taken as a constraint, so that the extracted potential features are ensured to contain all key information for accurate classification. Finally, the potential features are directly used for depression classification, and experimental results show that the accuracy and robustness of depression detection can be remarkably improved, and the method has high practical application value.

The invention discloses a data set distillation method and system based on semantic relation decoupling, and belongs to the technical field of data distillation. According to the method, the feature distribution of the distillation data is aligned with the category-level statistics of the original data stored in the memory bank to realize efficient and fine-grained category-level optimization of the distillation data, so that the representativeness and semantic consistency of the distillation data are improved; in addition, for the problem that distribution in the last layer of feature distribution of the pre-training model is too concentrated, the problem of downstream task gradient disappearance caused by insufficient sample diversity in the feature space is effectively relieved by maximizing diagonal elements of a covariance matrix and minimizing non-diagonal elements at the same time. Therefore, the generalization ability and the learning effect of the model on downstream tasks are improved.

The invention provides a network traffic matrix prediction method based on multi-scale convolution and attention collaborative enhancement LSTM (Long Short Term Memory), and solves the problems that an existing traffic matrix prediction model is insufficient in multi-scale feature capture, weak in space-time correlation modeling and poor in generalization. A dynamic multi-scale convolution module (adopting a 3 * 3 / 5 * 5 / 7 * 7 two-dimensional convolution kernel) is used to extract local fine-grained association and global coarse-grained trend of a two-dimensional traffic matrix, deep fusion of LSTM time sequence features and two-dimensional multi-scale spatial features is realized in combination with a cross attention mechanism, and time-space cross-dimension association is enhanced. And meanwhile, the problem of gradient disappearance of the deep network is relieved by relying on residual connection, and topology independence is realized based on a flow-by-flow method so as to adapt to traffic matrixes of different dimensions. The method specifically comprises the following steps: performing stream-by-stream division and normalization preprocessing on an original two-dimensional traffic matrix, and after processing by modules such as dynamic multi-scale convolution, cross attention enhancement LSTM, residual block and gating screening and the like, outputting a single OD stream predicted value and reconstructing a complete traffic matrix; and integrating objective functions of the encoder and the decoder, taking the MSE as a loss function, and adopting an Adam optimizer to iteratively update model parameters until the model converges. Experiments show that the MAE and the MSE of the method on real data sets such as Ailene, GEANT and the like are better than those of mainstream models such as SVR, LSTM, MTGNN and the like, the reconstruction precision of a high-dimensional traffic matrix is particularly remarkable, and a reliable basis can be provided for network congestion control and resource optimization.

The invention discloses a two-stage brain glioma segmentation system based on an improved U-Net model, and belongs to the field of medical image segmentation. According to the method, the problems of low precision, low efficiency and poor generalization ability of the existing segmentation technology are solved. In a tumor tissue detection stage, the lightweight neural network is utilized to realize preliminary identification of a brain glioma region, and the calculation range of subsequent fine segmentation is effectively reduced, so that the model can perform reasoning at relatively low resource and time cost, and the segmentation efficiency is improved. By adopting the multi-scale feature residual module, the problem of gradient disappearance caused by network deepening can be relieved, information can be propagated in a deeper level in the network, and extra parameter and calculation complexity overhead cannot be brought. Meanwhile, jump connection is changed into a multi-scale jump connection structure, the problem that a decoder network can only receive same-scale feature information is solved, and the tumor segmentation precision is remarkably improved. The method can be applied to medical image segmentation.

The invention provides a fan blade crack detection method, which comprises the following steps of: for a fan blade image acquired by an unmanned aerial vehicle, performing deblurring processing on the image through a pre-processing algorithm designed by the invention, and overcoming the problem of motion blurring when the unmanned aerial vehicle acquires the fan blade image; and carrying out crack detection on the acquired image by adopting a crack detection neural network model improved based on YOLOv8. According to the crack detection model, self texture interference of the blade can be avoided, the crack detection precision is improved, crack characteristics of different scales are captured through a Stem module of a multi-branch structure, a CAM module is added into a C2f-1 module, crack attention is integrated, and crack related characteristics are enhanced; a residual connection mode of a residual block is designed in a C2f-2 module to avoid the problem of deep network gradient disappearance, gradient flow and effective transmission of crack characteristics are guaranteed, response of crack-related channels is enhanced through channel re-calibration, and irrelevant channels such as blade textures are inhibited.

The invention discloses a mine dust concentration image recognition method based on prior features and a multi-kernel residual attention network, and the method comprises the steps: collecting mine dust images under different concentration conditions, extracting multi-dimensional image features, and constructing a prior feature library; linear and nonlinear feature subsets significantly related to the mine dust concentration are screened out; respectively mapping the linear and nonlinear feature subsets to a high-dimensional space by adopting a multi-kernel learning framework and carrying out weighted fusion to obtain fused high-order features; a multi-kernel residual attention network is constructed and trained, the gradient disappearance problem is relieved through a residual block structure, the weight of each feature channel is learned through a channel attention mechanism, and dynamic enhancement of key mine dust image features is carried out; and inputting the fused high-order features into a multi-kernel residual attention network for regression calculation, and outputting a mine dust concentration predicted value. According to the method, high-precision, non-contact and high-interpretability image recognition of mine dust concentration can be realized, and the method is suitable for real-time mine safety monitoring in a complex environment.

This invention discloses a gene regulation inference method guided by topological data analysis and gene network embedding. It combines TDA and GNN to enhance the inference capability of gene regulation networks. By capturing the topological structure of the gene regulation network graph through TDA features, the model's ability to model gene expression is enhanced. The TDA features and GAT embedding representations are effectively integrated through gating fusion. This fusion mechanism enables the model to adaptively adjust node embeddings based on global topological characteristics, which not only improves the accuracy of gene interaction representation but may also enhance the accuracy of regulatory relationship prediction. The traditional GAT architecture is extended through a four-layer graph attention mechanism. Each layer uses residual connections to alleviate the gradient vanishing problem and improve training stability. In addition, independent multilayer perceptron branches are designed for transcription factors and target gene embeddings. This deep architecture can achieve more expressive feature transformations and capture subtle patterns in gene regulation networks.

The invention provides a lightweight image super-resolution reconstruction method, device and equipment and a storage medium, standard convolution is solved into two steps of deep convolution and point convolution, a depth separable convolution layer performs independent extraction on each channel feature of a low-resolution image to realize channel decoupling, each channel only uses an independent convolution kernel, the convolution kernel is independent, and the reconstruction efficiency is improved. The original calculation amount of H * W * X * Y * M * N is reduced to H * W * X * Y * M, the parameter amount is reduced to X * Y * M from X * Y * M * N, and the compression ratio of about 1 / N is achieved. And then the information of each channel is fused through a 1 * 1 point convolution layer to generate fusion features, and the parameter quantity is further reduced while the feature expression capability is maintained. And meanwhile, residual connection is introduced to connect a plurality of feature stacking modules in series, cross-layer information transmission is realized, and bottom feature information is reserved, so that the problem of gradient disappearance during deep network training is prevented, and redundant calculation is avoided through feature multiplexing.

This invention discloses a plug-and-play model inversion attack method based on a generative model in a collaborative reasoning environment. This method uses a pre-trained StyleGAN2 as the target-independent image prior, leveraging its mapping and synthesis networks in the generator structure to map latent vectors into intermediate representations to generate images. During the attack optimization process, to address the gradient vanishing problem easily caused by traditional cross-entropy loss, a Poincaré loss function is introduced, utilizing the gradient preservation properties of non-Euclidean space to ensure the stability of the optimization process. Simultaneously, through standard image transformation and random image transformation techniques, the difference between the target data distribution and the image prior is effectively reduced, enhancing the robustness of the generated features. This method effectively overcomes limitations such as high computational resource consumption, insufficient flexibility, and sensitivity to changes in dataset distribution. In high-resolution scenarios such as face recognition, it demonstrates outstanding attack efficiency and generated image quality.

The invention relates to the technical field of network security, and discloses an XSS attack detection method and system based on multi-model fusion, and the system comprises a traffic data cleaning module, a feature extraction module, a traffic detection module and an instruction generation module. According to the method, hidden features in the front-back direction are captured by adopting a BiTCN to better obtain long-time dependency of a sequence, then, output of the BiTCN is further processed by utilizing a BiGRU layer, and the BiGRU enhances the memory ability of a model by combining a forward GRU and a reverse GRU, so that the model can learn dynamic changes of data from two directions, and therefore, the reliability of the model is improved. The method improves the perception capability of the model for the dynamic change of the time sequence, reinforces the semantic representation of the key segment of the XSS attack through the multi-head self-attention dynamic distribution weight, can reinforce the interaction between the internal features of the model through weight distribution and an attention layer, enables the model to learn more complex and abstract feature representation, and improves the accuracy of the model. And the problem of deep network gradient disappearance is solved by combining residual connection.

The invention relates to the technical field, in particular to a high-temperature furnace intelligent control system and method based on machine learning, and the method comprises the steps: a data collection module obtains and processes sensor data and outputs synchronous data streams; the twinborn modeling module is used for driving digital twinborn simulation by utilizing a synchronous data stream to obtain physical field data in the furnace; the causal analysis module identifies a key causal chain from the physical field data to generate a causal graph; the long-range prediction module adopts a neural network architecture integrated with a gating memory unit, and a causal diagram is combined to guide an attention mechanism to output a long-period prediction value; the strategy optimization module brings the predicted value into a reward function, and generates a control instruction through reinforcement learning and advanced simulation verification; the robust control module executes the instruction and monitors and predicts uncertainty to realize safe switching; the meta-learning module dynamically optimizes system parameters based on cross-batch performance data. According to the method, the gradient disappearance problem of the recurrent neural network in long-term dependence modeling is solved, and the control precision of the slow dynamic process of the high-temperature furnace is improved.

The application discloses a mine dust concentration image recognition method based on prior features and a multi-core residual attention network, which comprises the following steps: collecting mine dust images under different concentration conditions and extracting multi-dimensional image features to construct a prior feature library; screening out linear and nonlinear feature subsets significantly related to mine dust concentration; using a multi-core learning framework to respectively map the linear and nonlinear feature subsets to a high-dimensional space and perform weighted fusion to obtain fused high-order features; constructing a multi-core residual attention network and training it, relieving the gradient vanishing problem through a residual block structure, learning the weights of each feature channel through a channel attention mechanism, and dynamically strengthening key mine dust image features; inputting the fused high-order features into the multi-core residual attention network for regression calculation and outputting mine dust concentration prediction values. The method can realize high-precision, non-contact, highly interpretable image recognition of mine dust concentration and is suitable for real-time mine safety monitoring in complex environments.

The invention discloses a metal stamping defect real-time detection system based on machine learning, relates to the technical field of defect detection, and aims to solve the problems that a deep network is easily influenced by gradient disappearance, defects of different sizes need to be effectively detected according to different scales, dynamic adjustment cannot be performed according to different detection requirements, and the detection efficiency is low. The problem of dependence on a fixed data set is solved; a deep residual network is selected as a feature extraction backbone network, a common gradient disappearance problem in a deep network is effectively solved through a residual block structure, Anchor Scales are dynamically adjusted according to rust spot detection precision and a recall rate, a multi-task learning framework is combined to train defect classification and defect cause prediction, and branches of defect attribute identification are increased.

The invention provides a surface soil type identification method based on MCP-ShuffleNet V2, and the method comprises the following steps: S1, obtaining a sample soil image, and carrying out the preprocessing of the sample soil image, and then forming a training data set; s2, constructing an MCP-ShuffleNet V2 prediction model, and inputting the training data set into the MCP-ShuffleNet V2 prediction model to carry out training; and S3, obtaining a to-be-detected soil image, preprocessing the to-be-detected soil image, and inputting the preprocessed to-be-detected soil image into the trained MCP-ShuffleNet V2 prediction model to obtain a soil classification result of the to-be-detected soil image, so that interference of non-soil elements in the soil image can be effectively inhibited, a neural network is guided to focus on a key soil area, a continuous and derivable MELU activation function is provided, and the accuracy of the soil classification result is improved. And the gradient disappearance problem is effectively avoided, so that the final soil type identification precision is effectively ensured.

The application discloses a medical image segmentation method fusing multi-scale residual attention, and steps are as follows: after pre-processing of collected data sets, training sets, verification sets and test sets are divided; data enhancement is performed; in the training process, data is randomly vertically flipped, randomly horizontally flipped and randomly rotated with a probability of 0.5; a multi-scale residual attention network is constructed; in the encoder part, the network reduces the channel number of the intermediate part of double convolution, and moderately reduces network parameters; in the decoder, a multi-scale residual attention module is designed to capture the detailed features of medical images under different scales, improve the utilization rate of image feature maps, relieve the gradient vanishing problem and degradation problem caused by the network being too deep, and further improve the stability of the network; and the model is tested in the test set to obtain the segmentation result of the model. The application solves the problem that the target region of a medical image is fuzzy with a background boundary, can effectively reduce the parameter amount of the network while improving the image segmentation precision.

The application provides an electromagnetic signal recognition method based on a one-dimensional complex residual network, converts a communication signal received by a receiver into signal data in an int16 format, performs Fourier transform, draws a frequency spectrum diagram of the signal, estimates a carrier frequency of the signal, resamples the signal, performs intermediate frequency filtering, and then performs power normalization, trains a data set after power normalization processing through a one-dimensional complex residual network, and calculates the accuracy of the model on a test set, thereby solving the problem that the feature of a signal is lost due to training of a neural network by using real number representation of a communication signal in the prior art, and the residual network can avoid the gradient vanishing problem, so that a deeper network can be designed to extract the original feature of the communication signal, and a higher accuracy of modulation recognition is realized.

The application discloses a short-term and nowcasting precipitation prediction method based on GRU and LSTM space-time fusion, which comprises the following steps: step 1, reading local radar data and performing data preprocessing operation; step 2, dividing the processed data into a training set, a test set and a verification set; step 3, designing an ST-LSGRU basic module, stacking 4 layers to build a radar echo extrapolation model; step 4, inputting the training set, initializing model parameters, training a network model and saving model parameters; step 5, testing model performance by using the verification set, optimizing model hyperparameters, training a network model and saving model parameters; step 6, loading saved model parameters, inputting the test set, obtaining radar extrapolation results and obtaining precipitation in a future period of time through a Z-R relationship. The method can solve the gradient vanishing problem caused by long-time sequence prediction, and effectively improve the definition and stability of the prediction results of the model for strong precipitation weather.

The invention discloses a transmission line icing prediction method, system and device based on TCN-Attention and a medium, and relates to the technical field of transmission line icing and snowing.The method comprises the steps that time sequence data used for representing line environment conditions are collected; preprocessing the time sequence data; constructing a prediction network model composed of a time sequence modeling module and a feature attention mechanism module; inputting the preprocessed data into a prediction network model, and training the prediction model in different meteorological scenes through a multi-stage learning process; and utilizing the trained prediction model to predict the line icing torsion angle, the icing shape and the phase state parameters in the target time period. The time sequence modeling module can capture long-term time sequence dependence, the problem of gradient disappearance easily occurring in a traditional recurrent neural network is solved, meanwhile, the attention mechanism has the characteristic of dynamic weighting, and a higher weight value can be given to key parameters so as to highlight importance.

The application discloses a cell migration behavior prediction method based on physical information width learning, which utilizes cell scratch experiment data to construct a base function matrix, combines a Fisher-KPP reaction diffusion equation to construct a nonlinear least square problem containing physical mechanism constraints, and adopts an enhanced nonlinear least square disturbance algorithm to iteratively solve output weights, so as to realize prediction of cell density space-time evolution. The application converts traditional deep network iterative training into a nonlinear least square problem, greatly reduces a search space through physical information initialization and linearization approximation, and realizes order-of-magnitude improvement of training speed. The application constructs a width learning architecture containing feature nodes and enhanced nodes, effectively avoids the gradient vanishing problem in deep learning, combines analytical derivative calculation, eliminates cumulative error of automatic differentiation on high-order derivatives, and can more accurately capture space-time evolution characteristics in cell migration.

The present application discloses a shiitake mushroom growth time sequence image prediction method and system, a model construction method and system, and belongs to the technical field of time sequence image prediction. The present application solves the technical problems of insufficient sensitivity of detail features, serious gradient disappearance and weak long-term dependence modeling capability in the prior art. A shiitake mushroom fruiting body growth time sequence image dataset is constructed and preprocessed. A VMRNN-DMSA model is constructed, which is an improved model based on the VMRNN-D architecture. A skip connection mechanism integrating an MFM module is introduced into the backbone network, a SAM module is added, and an AKConv refining module is added. Based on the preprocessed dataset, a weighted loss function is used to train the VMRNN-DMSA model. After training is completed, the final shiitake mushroom growth time sequence image prediction model is obtained, and the model is used for shiitake mushroom growth prediction. The present application is used for realizing high-precision and high-quality shiitake mushroom growth time sequence image prediction.

The application discloses a remote sensing image landslide detection method and system, and belongs to the technical field of landslide detection. The method adopts a data enhancement method to expand a landslide data set, and solves the problem of a small number of landslide samples in a training set. DarkNet53 is used to replace a commonly used backbone feature extraction network of FasterR-CNN, so that the problem of gradient disappearance caused by a too deep network is avoided, and the detection precision is improved.

A hyperspectral soil nutrient inversion method based on WBL-ResNet, belonging to the field of intelligent interpretation of remote sensing images, consists of an MSFE module and a WBL-ResNet model. The MSFE module effectively captures different levels and details in hyperspectral images, enhancing the model's ability to express hyperspectral data and improving the efficiency of training and the accuracy of inversion. The design of WBL-ResNet residual connections reduces the gradient vanishing problem in deep network training, helping to maintain the learning stability of the network as its depth increases. A weighted balanced loss function is adopted, and through the design of Gaussian kernel function and variance weighting, higher weights are assigned to sparsely labeled samples, improving the inversion ability for sparse labels. This invention can accurately, quickly, and stably complete soil nutrient inversion in hyperspectral Earth observation scenarios, realizing contactless monitoring of farmland soil nutrient status using remote sensing methods.

The application discloses a kind of fresh grape variety recommendation method, system, equipment and storage medium based on knowledge graph belonging to intelligent recommendation technical field.It includes the following steps: collecting the attribute and variety preference data of fresh grape customer consumption behavior;According to attribute and variety, the knowledge graph of fresh grape field is constructed;According to the knowledge graph, the fresh grape customer variety recommendation model is established;The fresh grape customer variety recommendation model is trained, and the variety recommendation list of fresh grape customer is generated.The application solves the problem of over-smoothing and gradient vanishing that easily occurs in deep propagation of graph neural network, can prevent important information loss, and significantly improves the accuracy of fresh grape recommendation result.

The invention discloses a precipitation prediction method based on a GRU and an LSTM neural network. The precipitation prediction method comprises the following steps: selecting meteorological station data continuously monitored for a long time; removing abnormal values of the meteorological station data, and complementing missing values; selecting first 85% of historical data for model parameter training, and selecting last 15% of recent data for model generalization ability test; normalizing the three precipitation index dimensions of the training set and the test set respectively; combining an algorithm model by using GRU and LSTM components; inputting the preprocessed training set data into the network, using a mean square error as a loss function, adopting an Adam algorithm to optimize the loss function, and iteratively updating network parameters until the model converges; data to be predicted are input into the trained model after being preprocessed in the first step, and a precipitation prediction result is output. Sequence data information is processed through multiple influence factor data and the superposed GRU and LSTM combination, the gradient disappearance problem is avoided, the prediction effect is good, precision is high, and applicability is good.

The DSRINet network provided by the invention takes modular design as a core, and features of spectral data are efficiently extracted and classified in a multi-module combination mode. The infrastructure adopts an Inception module, and a multi-scale convolution kernel is utilized to capture multi-level features in spectral data, so that the expression ability of a model to complex data is enhanced, and local and global information is concerned at the same time. In the convolutional layer design, the DSRINet adopts depth separable convolution and is divided into two independent steps of deep convolution and point-by-point convolution. In the deep convolution, convolution is independently carried out on each channel along an input channel, and local features in a sequence segment can be effectively learned, such as key spectral features of a peak value, a trough and the like in a waveform; the output of the deep convolution is combined through the point-by-point convolution to generate final feature representation, so that the global correlation between different sequence segments is modeled, and the overall structure and mode of the spectral data are captured. Through the combination mode, the global mode of the spectral data can be learned while the local features are reserved by the depth separable convolution. In order to deal with the gradient disappearance problem in the deep network, the DSRINet introduces convolution-based residual connection, and compared with classical residual connection, the design not only relieves gradient disappearance, but also further enhances the nonlinear capability of the network. Besides, the invention provides an attention mechanism (FSWA) fusing feature selection and weight distribution, and the mechanism remarkably optimizes the feature extraction process by screening important features and distributing different weights for the features.

The invention discloses a time-of-use electricity price prediction method and device based on a residual Mama model. The method comprises the steps of data acquisition preprocessing and normalization processing; smoothing the normalized data; in a feature mapping stage, introducing the two trend components embedded into the spatial dimension into a residual Mama model; a residual linear layer is used for carrying out fusion learning on the double-branch features, so that the gradient disappearance problem in the training iteration process is relieved; the residual Mama model forms a richer feature representation by splicing the outputs of the two branches, the sum + of the two branches and the original mapping features; the fusion features are mapped to a space of a target prediction length and a variable dimension through a linear layer lin3; the output prediction result is subjected to inverse transformation of RevIN, and final time sequence prediction is obtained; and squaring the difference value of the time sequence prediction result and the true value, and summing to obtain the overall loss in the prediction process of the residual Mama model. The electricity price prediction method is more stable and efficient and has wide adaptability.

The invention discloses a mercuric chloride-based multi-modal medical large model and a medical image tumor segmentation modeling method. Comprising the following steps: integrating various medical image data and expert diagnosis reports, constructing a multi-modal data set, providing rich information for model training, comprehensively improving the recognition and analysis ability of tumor features, and further assisting doctors to make more accurate diagnosis and treatment decisions; through an edge computing technology, two business systems of mobile medical treatment and local medical treatment are constructed, and computing power is pushed from a cloud end or a data center to a network edge; aiming at the characteristics of a mercuric chloride server, carrying out all-parameter fine tuning on the multi-modal medical large model, carrying out fine tuning through a high-resolution image and supervised learning annotation data, and optimizing the model to adapt to a specific task; and a residual module and attention fusion are introduced, so that the complex feature extraction capability of the network is enhanced, and the gradient disappearance problem is effectively solved. The medical AI technology is effectively promoted to be widely applied to clinical scenes, and the efficiency and accuracy of medical diagnosis are practically improved.

The application discloses a kind of road disease detection method based on unmanned plane, electronic equipment and program product. The method is realized based on the road disease detection model of training completion, the C3k2-MDDSC module introduced in the backbone network of the model, the feature reuse capability is enhanced by gradient shunt and multi-scale fusion, the gradient vanishing problem is alleviated, and the model robustness is improved with the aid of jump connection and grouping convolution. The ACFP module is introduced at the end of the backbone network, then the dynamic fusion of local and global features is realized using multi-scale hollow convolution and channel-space attention mechanism, and the complex scene modeling capability is improved. The SGF module is integrated in the neck network, which can improve the spatial perception of the model for small targets. In addition, the ES-FPN proposed based on the SGF module not only can strengthen the utilization of shallow spatial information, but also can optimize the complementarity of cross-level features. During training, a fast and high-quality intersection over union loss is proposed as a regression loss, which improves the rotation box alignment accuracy and model convergence speed by introducing angle penalty.

The application discloses a photovoltaic array fault detection method based on an improved artificial neural network, and the photovoltaic array fault diagnosis model constructed by the photovoltaic array fault detection method comprises an input layer, a gated convolution layer, a flattening layer, an output layer and one or a plurality of fully connected layers connected in sequence; by adding the gated convolution layer and the plurality of fully connected layers, the gradient disappearance problem existing in the current photovoltaic array fault diagnosis is solved; the adaptive Softmax activation function is used as the activation function of the output layer, so that the accuracy and stability of the model are improved; in addition, the photovoltaic array is constructed through simulation simulation, a large number of working state parameters of the photovoltaic array can be acquired according to the module parameters of the photovoltaic array under the condition that the data is less; the application can process complex tasks with large data sets and provide fast prediction after training, and has good generalization ability and extremely high feasibility and application prospect.