Image fuzzy classification and icing prediction method and system based on Mama and storage medium

Abstract

The invention relates to the technical field of image fuzzy classification and power grid safety monitoring, in particular to an image fuzzy classification and icing prediction method and system based on Mama and a storage medium. The image fuzzy classification model provided by the invention comprises a Stem-P network and four Mamba-E networks which are connected in sequence, the input of the Stem-P network is the input of the basic model, and the output of the last Mamba-E network is the output of the basic model; the Stem-P network comprises a convolutional layer, a pooling layer, a linear layer and a nonlinear feature network which are connected in sequence; the Mama-E network comprises a first-layer normalization network, a first multi-layer perceptron, a second-layer normalization network, a Mama network, a first dimension superposition unit, a third-layer normalization network, a second multi-layer perceptron and a second dimension superposition unit which are connected in sequence; according to the training method of the image fuzzy classification model based on the Mama, pooling and linear processing are carried out on the input image by using the Stem-P network, dominant processing of hidden features of the image is realized, feature capture of image data information by a subsequent Mama-E network is facilitated, and thus the recognition efficiency and precision of the image fuzzy degree are improved. Besides, according to the icing prediction method provided by the invention, the rainfall capacity condition is predicted in combination with an image fuzzy degree classification result and the reflectivity of the rain measurement radar, the icing degree of the monitoring area can be predicted, and thus early warning of icing is realized. And the image fuzzy degree classification only needs to input a remotely collected visual image into the model for identification, the reflectivity of the rain measuring radar is provided by a meteorological radar of the meteorological bureau on line, and the accurate icing degree prediction result can be quickly obtained by combining simple input data, so that the applicability is very wide.

Description

Technical Field

[0001] The present invention relates to the technical field of image fuzzy classification and power grid safety monitoring, in particular to a Mamba-based image fuzzy classification and ice prediction method, system and storage medium. Background Art

[0002] According to statistics during the icing period, transmission lines usually experience rainfall before icing. In recent years, due to the influence of extremely severe weather, rime icing has caused frequent accidents such as line disconnection, tower collapse, and dancing of transmission lines, which seriously threaten the safe and stable operation of the power grid. Rime has the characteristics of high density, strong adhesion and tight structure, and the degree of damage it causes to transmission lines is far greater than that of rime, snow rime and mixed rime. Therefore, if the formation of rime icing can be warned in advance, it is of great significance to optimize the operation mode of the power grid and ensure the safe operation of the power grid.

[0003] Freezing rain is the main cause of rime and icing on transmission lines. At present, domestic and foreign scholars have conducted a lot of research on the formation of freezing rain. According to current research, the formation of freezing rain is related to specific geographical environment and climatic conditions. The formation of freezing rain cannot be accurately predicted using meteorological station monitoring and sounding data. Therefore, it is difficult to quantitatively evaluate the damage of rime and icing to transmission lines based on rainfall monitoring.

[0004] However, the images taken by the visual online monitoring device on the transmission line tower in bad weather have different degrees of blur, which shows that image blur can also be used as one of the conditions for transmission line icing prediction. However, there is currently a lack of image blur classification methods suitable for different weather conditions, making it difficult to provide a reliable reference for transmission line icing prediction. Summary of the invention

[0005] In order to overcome the defect of the above-mentioned prior art that there is a lack of image fuzzy classification methods suitable for different weather conditions, the present invention proposes a training method for an image fuzzy classification model based on Mamba, and the trained image fuzzy classification model realizes high-precision rapid classification of image blur degree.

[0006] The present invention proposes a training method for a Mamba-based image fuzzy classification model, which first constructs a basic model and a labeled data set; the labeled data set is used to store a visualization image labeled with a fuzzy degree classification label; the basic model inputs a visualization image and outputs a fuzzy degree classification label of the visualization image; the basic model is trained on the labeled data set to obtain a converged basic model as an image fuzzy classification model;

[0007] The basic model includes a Stem-P network and four Mamba-E networks connected sequentially. The input of the Stem-P network is the input of the basic model, and the output of the last Mamba-E network is the output of the basic model. The Stem-P network includes sequentially connected convolutional layers, pooling layers, linear layers, and nonlinear feature networks.

[0008] The Mamba-E network includes a first layer normalized network, a first multi-layer perceptron, a second layer normalized network, a Mamba network, a first dimension superposition unit, a third layer normalized network, a second multi-layer perceptron, and a second dimension superposition unit connected in sequence; wherein the input of the first layer normalized network is the input of the Mamba-E network, and the output of the second dimension superposition unit is the output of the Mamba-E network; the input of the Mamba-E network is also connected to the input of the first dimension superposition unit, and the output of the first dimension superposition unit is also connected to the input of the second dimension superposition unit.

[0009] Preferably, the Mamba network in the Mamba-E network adopts bidirectional sequence modeling.

[0010] Preferably, when modeling bidirectional sequences, the Mamba network includes both the forward and reverse sequences in the horizontal direction and the forward and reverse sequences in the vertical direction.

[0011] Preferably, the basic model training process includes the following steps:

[0012] SA1. Build a basic model and annotate a dataset; the annotate dataset is used to store visualization images annotated with blur classification labels;

[0013] SA2, extract training samples from the labeled data set and let the basic model learn the training samples;

[0014] SA3: Extract verification samples from the annotated data set and input them into the basic model to obtain the predicted labels output by the basic model and determine whether the basic model has converged. If not, execute step SA4. If yes, fix the basic model as the image fuzzy classification model.

[0015] SA4. Calculate the model loss by combining the predicted label and the true label on the verification sample, update the parameters of the base model through back propagation of the model loss, and then return to step SA2.

[0016] Preferably, in SA4, the model loss adopts cross entropy loss or mean square error loss.

[0017] The present invention also provides an icing prediction method. First, the image fuzzy classification model is obtained by using the training method of the Mamba-based image fuzzy classification model described above. The visual image of the monitoring area is input into the image fuzzy classification model to obtain the fuzzy category of the visual image. The rainfall radar reflectivity Z and the rainfall amount R are acquired. The icing evaluation value E is calculated.

[0018]

[0019] Among them, a is the weight value reflecting the influence degree of rainfall on icing, and b is the weight value reflecting the influence degree of image blurriness on icing; Y i represents the fuzzy degree value corresponding to the image fuzzy category.

[0020] If 0 < E ≤ 2b, it is determined that there is no icing in the monitoring area; if 2b < E ≤ 5a + 2b and R ≤ 5, it is determined that the monitoring area has mild to moderate icing; if E > 5a + 2b and R > 5, it is determined that the monitoring area has severe icing.

[0021] Preferably, the image fuzzy categories include: clear, mild to moderate blur, and severe blur. Let i ∈ {0, 1, 2}, where the fuzzy degree value Y0 corresponding to "clear" is 0, the fuzzy degree value Y1 corresponding to "mild to moderate blur" is 1, and the fuzzy degree value Y2 corresponding to "severe blur" is 2.

[0022] Preferably, the calculation method of the rainfall amount R is: obtain the Z - R curve, substitute the rainfall radar reflectivity Z into the Z - R curve to obtain the corresponding rainfall amount R; or, the rainfall amount R is obtained through a weather forecasting system or a rain gauge sensor.

[0023] An icing prediction system provided by the present invention includes a memory and a processor. The memory stores a computer program and the image fuzzy classification model obtained by using the training method of the Mamba-based image fuzzy classification model described above. The processor is connected to the memory, and the processor is used to execute the computer program to implement the icing prediction method described above.

[0024] A storage medium provided by the present invention stores a computer program, and when the computer program is executed, it is used to implement the icing prediction method described above.

[0025] The advantages of the present invention are as follows:

[0026] (1) The training method of the Mamba-based image fuzzy classification model provided by the present invention uses the Stem-P network to perform pooling and linear processing on the input image, realizing the explicit processing of the hidden features of the image, which is beneficial for the subsequent Mamba-E network to capture the feature information of the image data, thereby improving the recognition efficiency and accuracy of the image blurriness degree.

[0027] (2) In the present invention, by introducing the Selective State Space Model (Selective SSM), the computational efficiency is significantly improved, and the computational resource consumption can be greatly reduced while maintaining high performance. The image data information is captured using the Mamba network, and based on the unidirectional Mamba network, an improved bidirectional sequence modeling is introduced to improve the perception ability of visual input.

[0028] (3) In the present invention, bidirectional sequence modeling can simultaneously consider the previous and next information in the sequence, and effectively capture the bidirectional dependency between pixels in the image, thereby improving the ability to obtain and understand global contextual information. In the present invention, bidirectional sequence modeling extracts the forward and reverse sequences in the vertical direction (up and down) on the basis of extracting the forward and reverse sequences in the horizontal direction (left and right). At the same time, the horizontal and vertical bidirectional sequences of the image are extracted, making the extracted image features more complete, capturing the complete information of the image more comprehensively and effectively, and making the image blur classification more accurate.

[0029] (4) The ice coverage prediction method proposed in the present invention combines the image blur level classification results and the rainfall prediction of the rain radar reflectivity to predict the ice coverage in the monitoring area, thereby realizing early warning of ice coverage. The image blur level classification only requires the input of the remotely collected visual image into the model for recognition, and the rain radar reflectivity is provided online by the meteorological radar of the Meteorological Bureau. Combined with simple input data, accurate ice coverage prediction results can be quickly obtained, and the applicability is very wide. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a flow chart of the training method of the Mamba-based image fuzzy classification model;

[0031] Figure 2 This is the structure diagram of the image fuzzy classification model;

[0032] Figure 3 To improve the structure diagram of the Mamba network;

[0033] Figure 4 Schematic diagram of image enhancement;

[0034] a0 is a clear image; b0 is a slightly to moderately blurred image; c0 is a severely blurred image;

[0035] a1 is the image after random cropping of a0, a2 is the image after random rotation of a0, a3 is the image after horizontal flipping of a0, a4 is the image after vertical stretching of a0, and a5 is the image after brightness change of a0;

[0036] b1 is the image after b0 is randomly cropped, b2 is the image after b0 is randomly rotated, b3 is the image after b0 is horizontally flipped, b4 is the image after b0 is vertically stretched, and b5 is the image after b0 is changed in brightness;

[0037] c1 is the image after random cropping of c0, c2 is the image after random rotation of c0, c3 is the image after horizontal flipping of c0, c4 is the image after vertical stretching of c0, and c5 is the image after brightness of c0 is changed;

[0038] Figure 5 Comparison of loss values ​​of the models in the embodiments;

[0039] Figure 6 The accuracy comparison of the models in the embodiments;

[0040] Figure 7 2 is a comparison of the recall rates of the models in the embodiments. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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.

[0042] The present invention proposes a training method for a Mamba-based image fuzzy classification model. First, a basic model is constructed. The input of the basic model is a visualized image, and the output is a fuzzy degree classification label of the visualized image. Then, the basic model is trained on a labeled data set, and the basic model trained to convergence is used as an image fuzzy classification model. The fuzzy degree of the visualized image is classified by the image fuzzy classification model.

[0043] Specifically, the visualization image can be collected through a visualization online monitoring device installed on a transmission line tower; the image blur level classification can be specifically divided into: clear, light to medium blur and heavy blur.

[0044] Reference Figure 1 In this implementation, the training of the image fuzzy classification model includes the following steps:

[0045] SA1. Build a basic model and annotate a dataset. The annotated dataset is used to store visualization images annotated with blur classification labels. The blur classification labels in the annotated dataset can be annotated manually or using existing high-precision tools.

[0046] SA2, extract training samples from the labeled data set and let the basic model learn the training samples;

[0047] SA3: Extract verification samples from the annotated data set and input them into the basic model to obtain the predicted labels output by the basic model and determine whether the basic model has converged. If not, execute step SA4. If yes, fix the basic model as the image fuzzy classification model.

[0048] SA4. Calculate the model loss by combining the predicted label and the true label on the verification sample, update the parameters of the base model through back propagation of the model loss, and then return to step SA2.

[0049] In the specific implementation, in SA4, the model loss uses mean square error loss or cross entropy loss.

[0050] Reference Figure 2 The basic model in this implementation manner includes: a Stem-P network and four Mamba-E networks connected in sequence; the input of the Stem-P network is the input of the basic model, the output of the Stem-P network is connected to the input of the first Mamba-E network, the output of the first Mamba-E network is connected to the input of the second Mamba-E network, the output of the second Mamba-E network is connected to the input of the third Mamba-E network, the output of the third Mamba-E network is connected to the input of the fourth Mamba-E network, and the output of the fourth Mamba-E network is the output of the basic model.

[0051] The initial input image size of the base model is H×W×C, and each subsequent Mamba-E network successively reduces the image size to H / 4×W / 4×C, H / 8×W / 8×2C, H / 16×W / 16×4C, and the last Mamba-E network outputs an image size of H / 32×W / 32×8C.

[0052] During the image processing process, the basic model reduces the image size and increases the channel dimension through progressive downsampling to achieve efficient image feature extraction, thereby generating the final blur degree classification label.

[0053] The Stem-P network consists of sequentially connected convolutional layers, pooling layers, linear layers, and nonlinear feature networks. The Stem-P network divides the input image into patches by downsampling the input feature map using an improved "three layers one network" operation, thereby generating a feature map with a spatial dimension of H / 4×W / 4. The Mamba-E network is used to create hierarchical representations with resolutions of H / 4×W / 4, H / 8×W / 8, H / 16×W / 16, and H / 32×W / 32, and performs downsampling operations through patch merging.

[0054] Reference Figure 3, the Mamba-E network includes a first layer normalized network, a first multi-layer perceptron, a second layer normalized network, a Mamba network, a first dimension superposition unit, a third layer normalized network, a second multi-layer perceptron, and a second dimension superposition unit connected in sequence; wherein the input of the first layer normalized network is the input of the Mamba-E network, and the output of the second dimension superposition unit is the output of the Mamba-E network; the input of the Mamba-E network is also connected to the input of the first dimension superposition unit, and the output of the first dimension superposition unit is also connected to the input of the second dimension superposition unit;

[0055] That is, the input image features of the Mamba-E network are processed by the first normalization network, the first multi-layer perceptron, the second normalization network and the Mamba network in sequence, the output data of the Mamba network is dimensionally superimposed with the input image data of the Mamba-E network and then processed by the third normalization network and the second multi-layer perceptron again, and then the output data output by the first dimension superposition unit and the output data of the second multi-layer perceptron are dimensionally superimposed by the second dimension superposition unit as the output data of the Mamba-E network; which is equivalent to the output data of the Mamba-E network being the dimensionally superimposed result of the input data of the Mamba-E network, the output data of the Mamba network and the output data of the second multi-layer perceptron.

[0056] In this implementation, the Mamba network in the Mamba-E network adopts a bidirectional sequence, wherein feature x is convolved in a horizontally leftward and vertically upward direction during forward processing, and feature x is convolved in a vertically downward and horizontally rightward direction during backward processing.

[0057] Specifically, the network structure of the Mamba network includes: a normalization layer, an activation function, a forward convolution layer, a forward SSM, a backward convolution layer, a backward SSM, a third dimension stacking unit, and a fourth dimension stacking unit. The forward convolution layer and the forward SSM are connected to form a forward branch, and the backward convolution layer and the backward SSM are connected to form a backward branch.

[0058] The input of the Mamba network is converted into features z and x after being processed by the normalization layer. On the forward branch, feature x is convolved horizontally to the left and vertically upward by the forward convolution layer, and the forward SSM processes the convolution data output by the forward convolution layer as the output of the forward branch. On the backward branch, feature x is convolved vertically downward and horizontally to the right by the backward convolution layer, and the backward SSM processes the convolution data output by the backward convolution layer as the output of the backward branch.

[0059] The output of the backward branch is convolved with the data of feature z after being activated by the activation function and then input into the third-dimension superposition unit. The output of the forward branch is convolved with the data of feature z after being activated by the activation function and then input into the third-dimension superposition unit. The third-dimension superposition unit performs superposition processing on the two input data. The output of the third-dimension superposition unit and the output of the normalization layer are processed by the fourth-dimension superposition unit and then used as the output of the Mamba network.

[0060] The icing prediction method proposed in this embodiment includes the following steps S1 - S3.

[0061] S1. First, obtain the visual image of the monitoring area, classify the visual image using the image fuzzy classification model, and obtain the fuzziness value Y i ; obtain the rain radar reflectivity Z and the rainfall R;

[0062] Specifically, the rain radar reflectivity Z is the measured value of the radar; the rainfall R can be measured, or the rainfall R corresponding to the rain radar reflectivity Z can be obtained based on the Z - R curve (the fitting relationship between the rain radar reflectivity and the rainfall).

[0063] Specifically, the image fuzzy categories can be set to: clear, light - moderate blur, and severe blur. Let i ∈ {0, 1, 2}, where the fuzziness value Y0 corresponding to "clear" is 0, the fuzziness value Y1 corresponding to "light - moderate blur" is 1, and the fuzziness value Y2 corresponding to "severe blur" is 2.

[0064] S2. Substitute the rain radar reflectivity Z and the fuzziness value Y of the image i into the following formula to calculate the icing evaluation value E;

[0065]

[0066] where a is the weight value reflecting the influence degree of rainfall on icing, and b is the weight value reflecting the influence degree of image fuzziness on icing; a and b can be iterated according to the particle swarm optimization algorithm; Y i represents the fuzziness value corresponding to the image fuzzy category.

[0067] In specific implementation, first obtain the data set {rainfall, fuzziness; icing level}, and then use the particle swarm optimization algorithm to iterate the a and b weight values on the data set.

[0068] S3. Substitute the icing evaluation value E into the following judgment condition to obtain the icing degree prediction result:

[0069]

[0070] That is, if 0 < E ≤ 2b, it is judged that there is no icing in the monitoring area;

[0071] If \(2b < E\leqslant5a + 2b\) and \(R\leqslant5\), it is determined that the monitored area has light to moderate icing;

[0072] If \(E>5a + 2b\) and \(R>5\), it is determined that the monitored area has severe icing;

[0073] If \(1 < E\leqslant5a + 2b\) and \(R>5\), or \(E>5a + 2b\) and \(R\leqslant5\), a program error is returned and the next judgment is executed.

[0074] Embodiment 1

[0075] The above image blur classification model and icing prediction method are verified below in conjunction with specific embodiments.

[0076] In this embodiment, more than 70,000 sets of RGB three-channel images collected by visual online monitoring devices are collected from the transmission panoramic intelligent control platform in a certain place. The data collection time is from November 2023 to February 2024, covering five rounds of icing periods (a total of 53 days, among which, there are 8 days with an icing thickness of more than 5 mm, and 45 days with an icing thickness of more than 0 mm and less than 5 mm). The visual online monitoring device adopts an all-weather acquisition method and collects once every 30 minutes, ensuring comprehensive acquisition of image data during the day and night.

[0077] In this embodiment, images are randomly selected from the visual images collected by the visual online monitoring device and subjected to a three-person voting mechanism for cleaning and auditing in three rounds to obtain the original images labeled with blur degree classification labels. The resolution of the original images is compressed into image samples of \(640\times640\); then the cropped image samples are subjected to image enhancement, including random cropping, random rotation, horizontal flipping, vertical stretching, changing brightness, etc., to increase the number of samples, as specifically Figure 4 shown. Random cropping is to randomly select one of the center, upper left, lower left, upper right, and lower right of the image for cropping, and the cropping ratio is set to 0.8; random rotation is to randomly rotate the image at an angle of -30 degrees to 30 degrees; changing brightness is to randomly change the brightness of the image from 0.5 to 1.5.

[0078] In this embodiment, all image samples are divided into a training set and a validation set according to a ratio of 8:2; in this embodiment, the image blur classification model provided by the present invention and various comparison models are trained on the training set in combination with the above steps SA1 - SA4, and the performance of the converged model is tested on the validation set.

[0079] The comparison models include: YOLOv8x network model using ResNet-18 network, YOLOv8x network model using ResNet-101 network and Mamba-1 network model. The Mamba-1 network model is composed of Stem network and four traditional Mamba networks connected in sequence. The Stem network preprocesses the input image samples, and the Mamba network downsamples the input image features in turn to extract image features and output blur classification labels.

[0080] In this implementation, during the training process of the YOLOv8x network model YOLOv8x (ResNet-18) using the ResNet-18 network, the YOLOv8x network model YOLOv8x (ResNet-101) using the ResNet-101 network, the Mamba-1 network model, and the model of the present invention on the training set, the loss value, precision, and recall rate of the model are as follows: Figure 5 , Figure 6 , Figure 7 As shown, it can be seen that in the training set, the performance indicators of the model of the present invention are higher than those of the comparison model.

[0081] In this embodiment, when the four models are verified on the verification set, the performance of each fuzzy category is shown in Table 1 below.

[0082] Table 1 Comparison of validation results of four models

[0083]

[0084] It can be seen from Table 1 above that compared with the YOLOv8x (ResNet-18) model, the model of the present invention has obvious improvements in accuracy, recall rate and F1 value; compared with the YOLOv8x (ResNet-101) model, great progress has been made in accuracy, recall rate and F1 value; it can be seen that the model of the present invention has achieved different degrees of performance improvement in various performance indicators of the model compared with the YOLOv8x framework.

[0085] From Table 1, it can be seen that the model of the present invention has made progress in inference time compared with the three comparison models, especially compared with the YOLOv8x (ResNet-101) model, the inference time has been greatly shortened, as shown in Table 2 for details.

[0086] Table 2 Comparison of inference time of three models

[0087]

[0088]

[0089] As can be seen from Table 2, in terms of reasoning time, the model of the present invention saves 0.08 seconds on clear images, 0.139 seconds on lightly and moderately blurred images, and 0.071 seconds on heavily blurred images compared to the YOLOv8x (ResNet-101) model, and the calculation speed is increased by up to 51.29%; the model of the present invention saves 0.013 seconds on clear images, 0.012 seconds on lightly and moderately blurred images, and 0.002 seconds on heavily blurred images compared to the Mamba-1 model, and the calculation speed is increased by up to 8.33%.

[0090] It can be seen that the model of the present invention has either achieved significant improvement in model performance indicators or greatly improved computing speed compared to different YOLOV8 frameworks; compared to the Mamba-1 model, the model of the present invention has not only improved model performance, but also further improved the model's computational reasoning efficiency.

[0091] In this embodiment, the transmission line image collected by the online monitoring device installed on the transmission line tower is fuzzy classified in combination with the model of the present invention, and the rainfall is monitored by the rain sensor installed on site. Then, the ice coverage evaluation value E is calculated in combination with the rain radar reflectivity Z, the image blur value and the rainfall R:

[0092]

[0093] In this embodiment, the weights a and b are both 0.5.

[0094] In this implementation, the predicted value of ice cover level is obtained according to the following judgment conditions, and the measured ice cover thickness, image blur degree, measured rainfall and ice cover level predicted by the model of some monitoring results are statistically analyzed, as shown in Table 3.

[0095]

[0096] Table 3 Evaluation results of image blur, rainfall and ice level

[0097]

[0098]

[0099] In the line names in Table 3, 220kV, 800kV and 500kV represent voltage levels, SQ, SN, JT, AQ, LS, GJ and SA represent line abbreviations, # represents the line tower number, “±” represents a DC transmission line, and no “±” represents an AC transmission line.

[0100] It can be seen from Table 3 above that in this embodiment, ice thickness of 14 mm and below is identified as light to moderate ice, and ice thickness of 25 mm and above is identified as heavy ice. The pattern is obvious, and the prediction results are consistent with the actual values, which proves the reliability of the ice prediction method provided by the present invention.

[0101] Of course, it is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, but also includes the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any reference numerals in the claims should not be regarded as limiting the claims involved.

[0102] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

[0103] The techniques, shapes, and structural parts not described in detail in the present invention are all well-known techniques.

Claims

1. A training method for a Mamba-based image fuzzy classification model, characterized in that: First, construct a basic model and an annotation dataset; the annotation dataset is used to store visual images labeled with blur degree classification labels; the input of the basic model is a visual image, and the output is the blur degree classification label of the visual image; Train the basic model on the annotation dataset to obtain the converged basic model as the image blur classification model; The basic model includes, connected in sequence: a Stem-P network and four Mamba-E networks; the input of the Stem-P network is the input of the basic model, and the output of the last Mamba-E network is the output of the basic model; the Stem-P network includes, connected in sequence, a convolutional layer, a pooling layer, a linear layer, and a non-linear feature network; The Mamba-E network includes, connected in sequence, a first normalization network, a first multi-layer perceptron, a second normalization network, a Mamba network, a first dimension stacking unit, a third normalization network, a second multi-layer perceptron, and a second dimension stacking unit; wherein, the input of the first normalization network is the input of the Mamba-E network, and the output of the second dimension stacking unit is the output of the Mamba-E network; the input of the Mamba-E network is also connected to the input of the first dimension stacking unit, and the output of the first dimension stacking unit is also connected to the input of the second dimension stacking unit.

2. The training method of the Mamba-based image fuzzy classification model according to claim 1, characterized in that: The Mamba network in the Mamba-E network adopts bidirectional sequence modeling.

3. The training method of the Mamba-based image fuzzy classification model as claimed in claim 2, characterized in that: When the Mamba network performs bidirectional sequence modeling, it simultaneously includes forward and backward sequences in the horizontal direction and forward and backward sequences in the vertical direction.

4. The training method of the Mamba-based image fuzzy classification model according to claim 1, characterized in that: The training process of the basic model includes the following steps: SA1. Construct a basic model and an annotation dataset; the annotation dataset is used to store visual images labeled with blur degree classification labels; SA2. Extract training samples from the annotation dataset and let the basic model learn from the training samples; SA3. Extract validation samples from the annotation dataset and input them into the basic model, obtain the predicted labels output by the basic model, and determine whether the basic model has converged; if not, execute step SA4; if so, fix the basic model as the image blur classification model; SA4. Calculate the model loss on the validation samples by combining the predicted labels and the true labels, update the parameters of the basic model through the backpropagation of the model loss, and then return to step SA2.

5. The training method of the Mamba-based image fuzzy classification model as claimed in claim 4, characterized in that: In SA4, the model loss adopts cross-entropy loss or mean squared error loss.

6. An ice prediction method using the training method of the Mamba-based image fuzzy classification model according to any one of claims 1 to 5, characterized in that: First, obtain an image blur classification model by using the training method of the Mamba-based image blur classification model described in any one of claims 1-5; input the visual image of the monitoring area into the image blur classification model to obtain the blur category of the visual image; and obtain the rain gauge radar reflectivity Z and the rainfall R; calculate the icing evaluation value E; Among them, a is the weight value reflecting the influence of rainfall on ice cover, b is the weight value reflecting the influence of image blur on ice cover; Y i Indicates the blur degree value corresponding to the image blur category; If 0 < E ≤ 2b, it is determined that there is no icing in the monitoring area; if 2b < E ≤ 5a + 2b and R ≤ 5, it is determined that the monitoring area has light to moderate icing; if E > 5a + 2b and R > 5, it is determined that the monitoring area has severe icing.

7. The ice prediction method according to claim 6, characterized in that: Image blur categories include: clear, light to medium blur and heavy blur. Let i∈{0, 1, 2}, where the blur degree value Y0=0 for "clear", the blur degree value Y1=1 for "light to medium blur" and the blur degree value Y2=2 for "heavy blur".

8. The ice prediction method according to claim 6, characterized in that: The calculation method of the rainfall R is: obtain the ZR curve, substitute the rain radar reflectivity Z into the ZR curve to obtain the corresponding rainfall R; or, the rainfall R is obtained through a weather forecast system or a rainfall sensor.

9. An ice cover prediction system, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program and an image fuzzy classification model obtained by the training method of the image fuzzy classification model based on Mamba as described in any one of claims 1 to 5, and the processor is connected to the memory, and the processor is used to execute the computer program to implement the ice prediction method as described in any one of claims 6, 7 or 8.

10. A storage medium, characterized in that: A computer program is stored, and when the computer program is executed, it is used to implement the ice prediction method according to any one of claims 6, 7 or 8.

Images