Offshore equipment data processing method and system based on convolutional neural network model
The convolutional neural network model processes offshore equipment data, which solves the problems of multimodal data fusion and timing feature capture, achieves higher fault diagnosis accuracy and equipment status judgment, and provides reliable preventive maintenance.
Patent Information
- Application Number
- CN202510306174.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-14
- Publication Date
- 2025-07-01
AI Technical Summary
The existing technology is difficult to effectively integrate multimodal data of offshore equipment, and cannot capture different scales and timing characteristics, resulting in insufficient accuracy of fault diagnosis and difficulty in achieving accurate maintenance and early warning.
The method based on the convolutional neural network model is adopted to process the electrical, equipment status and meteorological data of offshore equipment through the dynamic receptive field convolution layer, the cross-modal feature interactive attention mechanism layer and the cyclic convolutional gating unit layer, so as to realize the fusion of multimodal features and the capture of timing information.
It significantly improves the accuracy of fault diagnosis, provides a more reliable preventive maintenance basis for equipment, and reduces equipment downtime and repair costs.
Smart Images
Figure CN120234564A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of data processing, and particularly to a method and system for processing data of marine equipment based on a convolutional neural network model. Background Art
[0002] The statements in this part merely provide background technical information related to the present invention and do not necessarily constitute prior art.
[0003] With the continuous development of offshore operations, for example, the continuous construction of onshore power facilities at offshore ports, as a solution to reduce the use of auxiliary diesel generators by ships when docking at ports, onshore power technology has been widely applied and developed. Among them, the onshore power pile is a key device of the onshore power system. However, due to the long-term exposure of onshore power piles to a complex working environment and being affected by various factors such as electrical stress, temperature changes, and humidity, various faults are likely to occur. Therefore, it is necessary to diagnose the faults of marine equipment such as onshore power piles. Traditional fault diagnosis of marine equipment relies on manual inspections and regular maintenance, and this method has obvious limitations. Manual inspections are difficult to detect potential faults in real time and accurately, and it is often difficult to detect some intermittent faults or early fault signs. Although regular maintenance can prevent the occurrence of faults to a certain extent, due to the fixed maintenance cycle, faults may occur between two maintenance periods and not be discovered in time.
[0004] With the continuous development of neural network model technology, data processing based on neural network models has become a research hotspot. However, marine equipment such as onshore power piles is often affected by various factors. When obtaining data, multi-modal data needs to be obtained. Existing technologies are difficult to effectively fuse the complex information between different modalities when processing multi-modal data of marine equipment such as onshore power piles, and cannot highlight the key modal feature combinations, resulting in insufficient information utilization and limited accuracy of fault diagnosis. Moreover, existing technologies are difficult to simultaneously capture the feature information of different scales in marine equipment data, and there is insufficient consideration of local details and overall trends, affecting the comprehensive and accurate judgment of the equipment operation status. At the same time, the temporal characteristics of marine equipment data such as onshore power pile data are difficult to be fully utilized in existing neural networks, and the changing trend of the equipment operation status cannot be effectively captured, resulting in low fault diagnosis accuracy and difficulty in achieving precise maintenance and early warning of the equipment. Summary of the Invention
[0005] To solve the above problems, the present invention proposes a method and system for processing data of marine equipment based on a convolutional neural network model, aiming to solve the deficiencies existing in the prior art mentioned in the background art.
[0006] To achieve the above object, the present invention adopts the following technical solutions:
[0007] In a first aspect, the present invention provides a method for processing data of marine equipment based on a convolutional neural network model, including:
[0008] Collect electrical data, equipment status data, and meteorological data of marine equipment, and preprocess the obtained electrical data, equipment status data, and environmental data to obtain corresponding feature vectors;
[0009] Input the feature vectors into a convolutional neural network model, which sequentially includes a dynamic receptive field convolutional layer, a pooling layer, a cross-modal feature interaction attention mechanism layer, a recurrent convolutional gated unit layer, and a fully connected layer; the feature vectors are input into the dynamic receptive field convolutional layer for spectral analysis to obtain data in each frequency band, and dynamic receptive field convolutional operations are performed on the data in different frequency bands at multiple scales to output sub-feature maps. After fusing all the sub-feature maps, a multi-modal feature map is obtained and output to the pooling layer for pooling. The pooling layer outputs the pooled multi-modal feature map and inputs it into the cross-modal feature interaction attention mechanism layer. In the cross-modal feature interaction attention mechanism layer, first, modal feature vectors are obtained, then an interaction matrix and an attention weight matrix are calculated, and a fused feature map is obtained according to the attention weight matrix; the recurrent convolutional gated unit layer obtains the fused feature map, updates the hidden state, and outputs it to the fully connected layer for fully connected calculation to output a fault result.
[0010] In a further technical solution, the electrical data includes voltage data and current data; the equipment status data includes equipment temperature data and switch status data; the meteorological data includes environmental temperature data and humidity data.
[0011] In a further technical solution, the specific method for inputting the feature vectors into the dynamic receptive field convolutional layer for spectral analysis is as follows: First, perform a two-dimensional discrete Fourier transform on the feature vectors to obtain a spectrogram, divide the spectrogram into data in multiple frequency bands, calculate its energy distribution, and determine the receptive field adjustment parameters corresponding to the data in each frequency band according to the energy distribution; then perform dynamic receptive field convolutional operations on the data in each frequency band respectively, and each frequency band data outputs a sub-feature map. The size of the sub-feature map is: where k is the convolution kernel size, s is the stride, H, W, and C are the height, width, and number of channels of the input feature vectors respectively, α i and β i are receptive field adjustment parameters; finally, all the sub-feature maps are fused to obtain a multi-modal feature map.
[0012] In a further technical solution, the specific method for calculating the interaction matrix is: Flatten the pooled multi-modal feature map to obtain a modal feature vector f i =flatten(F i ), where F iIs the multi-modal feature map after pooling; the elements in the interaction matrix Among them, W ij Is a learnable weight matrix, used to measure the feature interaction strength between modality i and modality j.
[0013] Further technical solution, the specific method for calculating the attention weight matrix is: normalize the interaction matrix to obtain the attention weight matrix A, and the elements in the attention weight matrix
[0014] The specific method for obtaining the fused feature map according to the attention weight matrix is: fuse the modality feature vectors through the attention weight matrix to obtain the modality feature fusion vector Then reshape the modality feature fusion vector to obtain the fused feature map.
[0015] Further technical solution, the specific method for updating the hidden state is: first perform multi-scale convolution operation on the obtained fused feature map, then calculate the gating signal, where the gating signal includes an update gate and a forget gate, then calculate the candidate state, and finally update the hidden state.
[0016] Further technical solution, the multi-scale convolution operation is:[[]] Among them, x t Is the fused feature map obtained at time t, Conv k Represents the convolution operation with a convolution kernel of k, K is the number of different convolution kernel sizes set, Is the fused feature map after convolution using the convolution kernel, Is to The multi-scale feature map spliced in sequence;
[0017] The calculation of the gating signal is:[[]] Among them, W z And W r Are both learnable weight matrices, σ is the sigmoid function, z t Is the update gate, r t Is the forget gate, h t-1 Is the hidden state at the previous moment;
[0018] The calculation of the candidate state is:[[]] Among them, W g Is a learnable weight matrix, ⊙ is the element-wise multiplication;
[0019] The update of the hidden state is: h t =(1 - z t )⊙h t-1 + z t ⊙gt 。
[0020] In a second aspect, the present invention provides a marine equipment data processing system based on a convolutional neural network model, including:
[0021] A data acquisition module, configured to: acquire electrical data, equipment status data, and meteorological data of marine equipment, and preprocess the acquired electrical data, equipment status data, and environmental data to obtain corresponding feature vectors;
[0022] A data processing module, configured to: input the feature vectors into a convolutional neural network model, which sequentially includes a dynamic receptive field convolutional layer, a pooling layer, a cross-modal feature interaction attention mechanism layer, a recurrent convolutional gated unit layer, and a fully connected layer; the feature vectors are input into the dynamic receptive field convolutional layer for spectral analysis to obtain data in each frequency band, and dynamic receptive field convolutional operations are performed on the data in different frequency bands at multiple scales to output sub-feature maps. After fusing all the sub-feature maps, a multi-modal feature map is obtained and output to the pooling layer for pooling. The pooling layer outputs the pooled multi-modal feature map and inputs it into the cross-modal feature interaction attention mechanism layer, where an interaction matrix and an attention weight matrix are calculated, and a fused feature map is obtained according to the attention weight matrix; the recurrent convolutional gated unit layer obtains the fused feature map, updates the hidden state, and outputs it to the fully connected layer for fully connected calculation to output a fault result.
[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0024] 1. The convolutional neural network model designed in the present invention can more accurately capture the features and temporal changes in marine equipment data such as onshore power pile data through the dynamic receptive field convolutional layer and the recurrent convolutional gated unit layer. Combined with the effective fusion of multi-modal data and the highlighting of key features by the cross-modal feature interaction attention mechanism layer, it can effectively fuse the complex information between different modalities, highlight the key modal feature combinations, make full use of the information, significantly improve the accuracy of fault diagnosis, provide a more reliable basis for preventive maintenance of equipment, and reduce equipment downtime and maintenance costs.
[0025] 2. The convolutional neural network model designed in the present invention can capture feature information at different scales from complex marine equipment data (onshore power pile data), making the feature information more comprehensive and discriminative. At the same time, it can take into account local details and overall trends, and can comprehensively and accurately judge the operating state of the equipment.
[0026] 3. The designed convolutional neural network model in the present invention can more effectively process multi-modal, multi-scale, and time-series data. On the one hand, it makes full use of the time-series characteristics of offshore equipment data such as onshore power pile data in the convolutional neural network model, effectively capturing the changing trends of the equipment operating state, so that the convolutional neural network model has stronger adaptability and generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] The accompanying drawings forming a part of this specification are used to provide a further understanding of the present invention. The schematic embodiments and descriptions thereof of the present invention are used to explain the present invention and do not constitute an improper limitation to the present invention.
[0028] Figure 1 is a flowchart of a method for processing offshore equipment data based on a convolutional neural network model; DETAILED DESCRIPTION OF THE EMBODIMENTS
[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0030] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs. In the case of no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.
[0031] Embodiment 1
[0032] The present embodiment provides a method for processing offshore equipment data based on a convolutional neural network model, including the following steps:
[0033] S1: Collect electrical data, equipment status data, and meteorological data of offshore equipment, and preprocess the obtained electrical data, equipment status data, and environmental data to obtain corresponding feature vectors;
[0034] S2: Input the feature vector into a convolutional neural network model, which successively includes a dynamic receptive field convolutional layer, a pooling layer, a cross-modal feature interaction attention mechanism layer, a recurrent convolutional gated unit layer, and a fully connected layer. The feature vector is input into the dynamic receptive field convolutional layer for spectrum analysis to obtain data in each frequency band, and dynamic receptive field convolutional operations are performed on the data in different frequency bands at multiple scales to output sub-feature maps. After fusing all the sub-feature maps, a multi-modal feature map is obtained and output to the pooling layer for pooling. The pooling layer outputs the pooled multi-modal feature map and inputs it into the cross-modal feature interaction attention mechanism layer. In the cross-modal feature interaction attention mechanism layer, first obtain the modal feature vectors, then calculate the interaction matrix and the attention weight matrix, and obtain the fused feature map according to the attention weight matrix. The recurrent convolutional gated unit layer obtains the fused feature map, updates the hidden state and outputs it to the fully connected layer for fully connected calculation to output the fault result.
[0035] Among them, in step S1, the offshore equipment in this embodiment is a shore power pile. The electrical data collected from the offshore equipment (shore power pile) includes but is not limited to voltage data, current data, and power data; the equipment status data includes but is not limited to the temperature data of the equipment (shore power pile) and the switch status data; the meteorological data includes but is not limited to the ambient temperature data and humidity data. Among them, the voltage value output by the shore power pile is one of the most basic monitoring data, which can reflect the power supply status of the shore power pile. The current data can reflect the load condition of the shore power pile. The power data is the product of voltage and current, which directly reflects the electric energy output of the shore power pile. The temperature data of the equipment (shore power pile). During the operation of the electrical components inside the shore power pile, heat will be generated. Excessive temperature may affect the performance and service life of the equipment. The switch status data records the status of various switches of the shore power pile (such as circuit breakers, contactors, etc.), including the opening and closing time of the switch, the number of operations, etc. These data help to analyze the operation frequency of the equipment, service life, and the operation of the switch during a fault. The ambient temperature data affects the heat dissipation efficiency of the shore power pile, affecting the output power and voltage stability of the entire shore power pile. The humidity data affects the insulation performance of the shore power pile.
[0036] And in step S1, the preprocessing includes data cleaning and feature extraction. Traditional data preprocessing methods often use fixed thresholds or simple statistical methods when cleaning outliers and extracting features, lacking sufficient consideration of the dynamic characteristics and potential laws of the data. In this embodiment, by introducing an adaptive weighting mechanism, the thresholds for data cleaning and the weights for feature extraction can be dynamically adjusted according to the local and global feature distributions of the data, so as to more accurately remove noise and outliers and strengthen the key features closely related to the equipment operation state: The data cleaning is as follows: For a data sequence L = [l1, l2…, l n of a certain modality, first calculate its local mean μ iand the local standard deviation σ i , the local mean and standard deviation are calculated by means of a sliding window with a window size of w. Let l j be the current data point, and calculate its deviation degree from the local mean Introduce an adaptive weight function W(d j ), and its form is σ w represents the standard deviation of the weight function, which can be adaptively adjusted according to the overall fluctuation of the data. This weight function makes the data points with smaller deviations have higher weights, while the weights of data points with larger deviations drop rapidly. The cleaned data set X clean : X clean = [l j × W(d j ) | j = 1, 2, 3…, n], in this way, the influence of outliers (i.e., data points where W(d j ) is close to 0) is greatly weakened, while normal data points are retained and weighted and adjusted according to their deviation degrees from the local mean, making the data smoother and more reliable.
[0037] For the cleaned data X clean , wavelet transform is used for feature extraction. Let ψ a，b (t) be the wavelet function, where a is the scale parameter and b is the translation parameter. Perform continuous wavelet transform on X clean to obtain where is the conjugate function of ψ a，b (t). By analyzing the coefficients after wavelet transform select the coefficients within a specific scale a k and frequency band b k related to the fault characteristics of the onshore power pile to construct the feature vector where K and L are the numbers of selected scales and frequency bands respectively. To further enhance the discriminant ability of the features, introduce a feature enhancement weight matrix W F , and its elements are obtained through the analysis and learning of a large amount of historical onshore power pile data, and are used to highlight the feature components that play a key role in fault diagnosis and prediction. Finally, the enhanced feature vector F enhanced = W F · F.
[0038] In step S2, in the dynamic receptive field convolutional layer, when facing complex and variable shore power pile data, the traditional dynamic receptive field method may not be able to accurately and flexibly adjust the receptive field according to the diversity of data features. In the dynamic receptive field convolutional layer proposed in this embodiment, by introducing an adaptive mechanism based on the spectral analysis of data features, it is possible to dynamically adjust the receptive field simultaneously at multiple scales, so as to more comprehensively capture various feature information in the data. Whether it is high-frequency mutation signals or low-frequency trend changes, effective feature extraction can be obtained. Specifically: First, perform a two-dimensional discrete Fourier transform (2D-DT) on the feature vector X (with dimensions H×W×C, where H is the height, W is the width, and C is the number of channels) to obtain the spectrogram F = DFT2(X). Divide the spectrogram into N band data F = {F1, F2, …, F N}, calculate the energy distribution E i for each band data, and determine the receptive field adjustment parameters corresponding to each band data according to the energy distribution, which are α i and β i , for example, through a learnable mapping function g: [α i , β i = g(E i ); then perform dynamic receptive field convolution operations on each band data respectively. Each band data outputs a sub-feature map, and the size of the sub-feature map is: where k is the convolution kernel size, s is the stride, H, W, and C are the height, width, and number of channels of the input feature vector respectively, α i and β i are the receptive field adjustment parameters; finally, fuse all the sub-feature maps to obtain a multi-modal feature map where i is the i-th band data, Y i is the output sub-feature map after the i-th band data passes through the dynamic receptive field convolution, which contains the feature information extracted from this band data through a specific receptive field.
[0039] In step S2, in the cross-modal feature interaction attention mechanism layer, in multi-modal data processing, existing attention mechanisms often only focus on intra-modal or simple inter-modal fusion, while ignoring the complex interaction relationships between modalities. This improved cross-modal feature interaction attention mechanism can deeply explore the potential associations and mutual influences between different modal features by constructing a cross-modal feature interaction graph, so as to more accurately highlight the cross-modal feature combinations that play a key role in fault diagnosis and prediction during the fusion process, further improving the diagnostic accuracy and reliability of the model. Specifically: Flatten the pooled multi-modal feature map to obtain the modal feature vector f i = flatten(F i ), where Fi is the multi-modal feature map after pooling; the elements in the interaction matrix where W ij is a learnable weight matrix used to measure the feature interaction strength between modality i and modality j. The interaction matrix is normalized to obtain the attention weight matrix A, and the elements in the attention weight matrix
[0040] The specific method for obtaining the fused feature map according to the attention weight matrix is: the modality feature vectors are fused through the attention weight matrix to obtain the modality feature fusion vector Then the modality feature fusion vector is reshaped to obtain the fused feature map.
[0041] In step S2, in the recurrent convolutional gated unit layer, when the traditional recurrent convolutional network processes time series data, the way of feature extraction and fusion of time series information is relatively simple, and it is difficult to fully capture the complex time-dependent relationships and variable feature patterns in the onshore power pile data. In this embodiment, the recurrent convolutional gated unit layer is essentially a time-series-based recurrent convolutional gated unit layer, which introduces a hierarchical time-series feature extraction mechanism, combines adaptive convolutional filtering and dynamic gated recurrent units, and can perform feature extraction and information screening on data at different time scales, so as to more accurately process the long-term trends and short-term fluctuations in the onshore power pile operation data, and improve the sensitivity and prediction accuracy to changes in the device operation state. The specific method for updating the hidden state is: first, perform multi-scale convolutional operations on the obtained fused feature map, then calculate the gating signals, where the gating signals include the update gate and the forget gate, then calculate the candidate state, and finally update the hidden state. Specifically: performing multi-scale convolutional operations is:[[]] where, x t is the fused feature map obtained at time t, Conv k represents the convolutional operation with a convolution kernel of k, K is the number of different convolution kernel sizes set, is the fused feature map after convolution using the convolution kernel, is to The multi-scale feature map concatenated in sequence;
[0042] Calculating the gating signals is:[[]] where, W z and W r are both learnable weight matrices, σ is the sigmoid function, z t is the update gate, r t is the forget gate, h t-1 is the hidden state at the previous moment;
[0043] Calculating the candidate state is:[[]] where, Wg is a learnable weight matrix, and ⊙ is element-wise multiplication;
[0044] Update the hidden state to: h t =(1 - z t )⊙h t-1 + z t ⊙g t .
[0045] Example Two
[0046] This example provides a marine equipment data processing system based on a convolutional neural network model, which specifically includes the following modules:
[0047] Data acquisition module, configured to: collect electrical data, equipment status data, and meteorological data of marine equipment, and preprocess the obtained electrical data, equipment status data, and environmental data to obtain corresponding feature vectors;
[0048] Data processing module, configured to: input the feature vectors into a convolutional neural network model, which successively includes a dynamic receptive field convolutional layer, a pooling layer, a cross-modal feature interaction attention mechanism layer, a recurrent convolutional gated unit layer, and a fully connected layer; the feature vectors are input into the dynamic receptive field convolutional layer for spectral analysis to obtain data of each frequency band, and dynamic receptive field convolutional operations are performed on the data of different frequency bands at multiple scales to output sub-feature maps. After fusing all the sub-feature maps, a multi-modal feature map is obtained and output to the pooling layer for pooling. The pooling layer outputs the pooled multi-modal feature map and inputs it into the cross-modal feature interaction attention mechanism layer. In the cross-modal feature interaction attention mechanism layer, first, modal feature vectors are obtained, then an interaction matrix and an attention weight matrix are calculated, and a fused feature map is obtained according to the attention weight matrix; the recurrent convolutional gated unit layer obtains the fused feature map, updates the hidden state, and outputs it to the fully connected layer for fully connected calculation to output a fault result.
[0049] It should be noted here that each module in this example corresponds one-to-one with the method in Example One, and the specific implementation process is the same, so it will not be repeated here.
[0050] Example Three
[0051] This example provides a computer-readable storage medium with a program stored thereon. When the program is executed by a processor, it implements the steps in the marine equipment data processing method based on a convolutional neural network model as described in Example One above.
[0052] Example Four
[0053] An electronic device, comprising a memory, a processor, and a program stored on the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for processing maritime device data based on the convolutional neural network model described in the first embodiment above are implemented.
[0054] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.
[0055] Although the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, it is not a limitation on the protection scope of the present invention. Those skilled in the art should understand that based on the technical solutions of the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present invention.
Claims
1. A method for processing offshore equipment data based on a convolutional neural network model, characterized in that: include: Collect electrical data, equipment status data and meteorological data of offshore equipment, and pre-process the acquired electrical data, equipment status data and environmental data to obtain corresponding feature vectors; The feature vector is input into a convolutional neural network model, which includes a dynamic receptive field convolution layer, a pooling layer, a cross-modal feature interaction attention mechanism layer, a recurrent convolution gating unit layer, and a fully connected layer in sequence; the feature vector is input into the dynamic receptive field convolution layer for spectrum analysis to obtain data of each frequency band, and a dynamic receptive field convolution operation is performed on the data of different frequency bands at multiple scales, and a sub-feature map is output. All sub-feature maps are fused to obtain a multi-modal feature map and output to the pooling layer for pooling. The pooling layer outputs the pooled multi-modal feature map and inputs it into the cross-modal feature interaction attention mechanism layer. In the cross-modal feature interaction attention mechanism layer, the modal feature vector is first obtained, and then the interaction matrix and the attention weight matrix are calculated, and the fused feature map is obtained according to the attention weight matrix; the recurrent convolution gating unit layer obtains the fused feature map, updates the hidden state and outputs it to the fully connected layer for full connection calculation, and outputs the fault result.
2. The offshore equipment data processing method based on the convolutional neural network model according to claim 1, characterized in that: The electrical data includes voltage data and current data; the device status data includes device temperature data and switch status data; and the meteorological data includes ambient temperature data and humidity data.
3. The offshore equipment data processing method based on the convolutional neural network model according to claim 1, characterized in that: The feature vector is input into the dynamic receptive field convolution layer for spectrum analysis. The specific method is as follows: first, a two-dimensional discrete Fourier transform is performed on the feature vector to obtain a spectrum map, the spectrum map is divided into multiple frequency band data, its energy distribution is calculated, and the receptive field adjustment parameters corresponding to each frequency band data are determined according to the energy distribution; then, a dynamic receptive field convolution operation is performed on each frequency band data, and each frequency band data outputs a sub-feature map, and the size of the sub-feature map is: Where k is the convolution kernel size, s is the step size, H, W and C are the height, width and number of channels of the input feature vector respectively, α i and β i Adjust parameters for the receptive field; finally, fuse all sub-feature maps to obtain a multimodal feature map.
4. The offshore equipment data processing method based on the convolutional neural network model according to claim 1, characterized in that: The specific method of calculating the interaction matrix is: flattening the pooled multimodal feature map to obtain the modal feature vector f i = flatten(F i ), where F i is the multimodal feature map after pooling; the element M in the interaction matrix ij =f i T ·W ij ·f j , where W ij is a learnable weight matrix used to measure the feature interaction strength between modality i and modality j.
5. The offshore equipment data processing method based on the convolutional neural network model according to claim 1, characterized in that: The specific method for calculating the attention weight matrix is: normalizing the interaction matrix to obtain the attention weight matrix A, and the elements in the attention weight matrix The specific method of obtaining the fusion feature map according to the attention weight matrix is: fusing the modal feature vectors through the attention weight matrix to obtain the modal feature fusion vector WijT·fj, and then reshape the modal feature fusion vector to obtain the fused feature map.
6. The offshore equipment data processing method based on the convolutional neural network model according to claim 1, characterized in that: The specific method for updating the hidden state is: first, a multi-scale convolution operation is performed on the acquired fusion feature map, and then a gating signal is calculated, wherein the gating signal includes an update gate and a forget gate, and then a candidate state is calculated, and finally the hidden state is updated.
7. The offshore equipment data processing method based on the convolutional neural network model according to claim 6, characterized in that: The multi-scale convolution operation is as follows: Among them, x t is the fusion feature map obtained at time t, Conv k represents a convolution operation with a convolution kernel of k, where K is the number of different convolution kernel sizes set. To use the fused feature map after convolution with the convolution kernel, For the general Multi-scale feature maps spliced in sequence; The calculation gating signal is: Among them, W z and W r are all learnable weight matrices, σ is the sigmoid function, z t is the update gate, r t is the forget gate, h t-1 is the hidden state at the previous moment; The candidate states for calculation are: Among them, W g is the learnable weight matrix, ⊙ is the element-wise multiplication; The updated hidden state is: h t =(1-z t )⊙h t-1 +z t ⊙g t .
8. Offshore equipment data processing system based on convolutional neural network model, characterized in that: include: The data acquisition module is configured to: collect electrical data, equipment status data and meteorological data of offshore equipment, and pre-process the acquired electrical data, equipment status data and environmental data to obtain corresponding feature vectors; The data processing module is configured to: input the feature vector into a convolutional neural network model, which includes a dynamic receptive field convolution layer, a pooling layer, a cross-modal feature interaction attention mechanism layer, a recurrent convolution gating unit layer, and a fully connected layer in sequence; input the feature vector into the dynamic receptive field convolution layer for spectrum analysis to obtain data of each frequency band, and perform dynamic receptive field convolution operations on different frequency band data at multiple scales, output a sub-feature map, fuse all sub-feature maps to obtain a multi-modal feature map and output it to the pooling layer for pooling, the pooling layer outputs the pooled multi-modal feature map and inputs it to the cross-modal feature interaction attention mechanism layer, in the cross-modal feature interaction attention mechanism layer, first obtain the modal feature vector, then calculate the interaction matrix and the attention weight matrix, and obtain the fused feature map according to the attention weight matrix; the recurrent convolution gating unit layer obtains the fused feature map, updates the hidden state and outputs it to the fully connected layer for full connection calculation, and outputs the fault result.
9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps in the offshore equipment data processing method based on a convolutional neural network model as described in any one of claims 1 to 7 are implemented.
10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the offshore equipment data processing method based on the convolutional neural network model as described in any one of claims 1 to 7 are implemented.