An adaptive dynamic weighted fusion system and method for multi-source security data in oilfields under harsh environments

Abstract

This invention discloses an adaptive dynamic weighted fusion system and method for multi-source security data in oilfields under harsh environments, belonging to the field of industrial IoT and multi-sensor data fusion technology. Addressing the problems of low accuracy, disconnect between environmental and sensor reliability, scarcity of real data, and poor environmental recognition adaptability of traditional fixed-weight fusion methods in harsh environments such as oilfield dust storms, rain, snow, and low light, this invention employs dual-source data supply, MLP environmental recognition, dynamic weight prediction, dynamic weighted fusion, and combined loss optimization techniques to construct a six-layer closed-loop architecture. This achieves quantitative mapping between environmental factors and sensor reliability, as well as adaptive weighted fusion under harsh environments. This invention can significantly improve data fusion accuracy under harsh environments and is suitable for scenarios such as oilfield perimeter security, equipment monitoring, and intrusion detection, exhibiting strong environmental adaptability and high reliability.

Description

Technical Field

[0001] This invention belongs to the fields of industrial Internet of Things, multi-sensor data fusion, and deep learning technology, and specifically relates to a dynamic weighted fusion system and method for multi-source security data in oil fields that is adaptive to harsh environments. Background Technology

[0002] Oilfield security systems are core infrastructure for ensuring the safety of oil and gas production. They rely on data collected from multiple sources, including cameras, infrared sensors, radar sensors, and vibration sensors, for comprehensive analysis and decision-making. However, the outdoor environment of oilfields is complex and variable; severe weather conditions such as sandstorms, rain, snow, low light, and strong light can significantly affect the performance of different sensors. 1. Low accuracy of fixed-weight fusion: Existing multi-sensor data fusion methods mostly adopt fixed-weight allocation strategies, which cannot dynamically adjust the contribution of each sensor according to environmental changes. As a result, in harsh environments, severely affected sensors (such as cameras in sandstorm weather) will reduce the overall fusion accuracy and even produce wrong decisions.

[0003] 2. Disconnect between environment and sensor reliability: The lack of a quantitative mapping mechanism between environmental factors and sensor reliability makes it impossible to assess the effective working capability of each sensor based on real-time environmental conditions, resulting in a lack of scientific basis for fusion decision-making.

[0004] 3. Scarcity of real data: Security data collection in the harsh environment of oil fields is difficult and costly, and the number of labeled samples is small. Models that rely solely on real data for training have poor generalization ability and are difficult to adapt to the complex environment of different oil fields.

[0005] 4. Poor adaptability of the environment recognition module: Some technologies use convolutional neural networks (CNN) for environment recognition, but environmental factors are mostly low-dimensional numerical data (light intensity, dust concentration, rain and snow level). The CNN architecture is prone to dimensionality mismatch problems, resulting in low accuracy of environment classification. Summary of the Invention

[0006] This invention aims to address the industry pain point of the sharp drop in accuracy of traditional fixed-weight fusion methods in complex environments. It proposes an adaptive dynamic weighted fusion system and method for multi-source security data in oilfields under harsh environments. This system enables adaptive fusion of security data from multiple sensors in oilfields under harsh environments such as sandstorms, rain, snow, and low light. It constructs a complete technical system that integrates data generation, environmental identification, weight prediction, dynamic fusion, and visualization interaction, significantly improving the reliability and accuracy of oilfield security systems in harsh environments.

[0007] To achieve the above objectives, the present invention provides the following technical solution: This invention provides an adaptive, harsh environment-friendly, multi-source security data dynamic weighted fusion system for oilfields, the system comprising: The dual-source data processing module is used to generate synthetic data, load real data, build a dual-source data supply mechanism, and output a standardized dataset with environmental labels and sensor reliability labels. The environmental identification module is used to extract environmental factors from the standardized dataset, perform feature extraction through multi-level linear transformation and nonlinear activation, and output the probability distribution of oilfield environmental categories. The dynamic weight prediction module is used to map environmental factors to fusion weights of each sensor based on the MLP model, normalize the fusion weights, and output normalized fusion weights. The dynamic weighted fusion module is used to weight and fuse multi-sensor features according to the normalized fusion weights, perform adaptive weight adjustment in harsh environments, and output a unified feature vector after fusion. The training and inference module is used to complete model training and parameter optimization based on the combined loss function, and to perform real-time inference and performance evaluation. The integrated visual interaction module is used to realize full-process visual monitoring and log recording of parameter configuration, data processing, model training, inference execution and result display.

[0008] Furthermore, the synthetic data generation described in this invention specifically includes: generating a synthetic multi-source dataset with environmental labels and sensor reliability labels by simulating sensor data characteristics under different harsh environments and their corresponding reliability change patterns; The generation of real data specifically includes: loading real data in CSV format at least, and automatically parsing the loaded data. Environmental factors, multi-sensor features, environmental labels, and sensor reliability labels are used to complete data dimension verification and format conversion, and output a standardized dataset that adapts to the model input.

[0009] Furthermore, this invention employs an MLP architecture to construct the environment recognition model; the environment recognition model outputs the environment category probability distribution through a forward propagation formula, which is:

[0010] in, The input is the environmental factor matrix (B is the batch size). Here are the weight matrices for each layer. For bias terms, This represents the probability distribution of the output environment categories.

[0011] Furthermore, the dynamic weighted fusion module of the present invention realizes the dynamic fusion of multi-sensor data based on the weighted average principle. It multiplies the feature vectors of each sensor with their corresponding prediction weights and then sums them to output a unified feature vector after fusion.

[0012] Furthermore, the adaptive weight adjustment performed under harsh environments in this invention specifically involves: When the current environment is identified as a harsh environment, the weights are adjusted as follows:

[0013] Adjusted and renormalized weights:

[0014] in, For the first i The weights of each sensor.

[0015] Furthermore, the training and inference module of this invention also uses the AdamW optimizer to update model parameters and introduces weight decay to prevent overfitting.

[0016] Furthermore, the combined loss function of the training inference module described in this invention is:

[0017] in: To identify cross-entropy loss for the environment, The weighted L2 regularization loss, This is due to reliability matching loss.

[0018] Furthermore, the system described in this invention constructs a six-layer closed-loop architecture, which consists of a parameter configuration layer, a dual-source data processing layer, an environment identification layer, a dynamic weight prediction layer, a dynamic weighted fusion layer, and a visualization interaction layer, from bottom to top. Data is exchanged and results are fed back between each layer, forming a complete technical closed loop.

[0019] Based on the same inventive concept, this invention also provides a dynamic weighted fusion method for multi-source security data in oilfields that adapts to harsh environments, the method comprising the following steps: S1: System initialization and parameter configuration, completing hardware environment adaptation, core parameter settings, and model and module initialization; S2: Dual-source data preparation and loading, select synthetic data generation or real data loading, output standardized dataset and divide it into training set and test set; S3: Model training and parameter optimization. Based on the training set, the environment recognition model and dynamic weight prediction model are iteratively trained using a combination loss function and optimizer. The trained model and logs are saved. S4: Visualize training results, plot and display the training loss curve and the environment recognition accuracy curve, and output the core training metrics; S5: Single-time real-time inference, based on a single sample, performs environment recognition, dynamic weight prediction, dynamic weighted fusion and adaptive weight adjustment in harsh environments, and outputs inference results; S6: Batch inference and performance evaluation, performs inference on batch samples, calculates the accuracy of environment recognition and the accuracy improvement rate under harsh environments, and generates an evaluation report; S7: Results display and log recording. Visualize the inference results and record and save logs related to the entire process of operation, training and inference.

[0020] The beneficial effects of this invention are as follows: 1. Dynamic mapping mechanism between environmental factors and sensor reliability: This invention establishes for the first time a quantitative mapping relationship between 3D environmental factors (light intensity, dust concentration, rain and snow level) and the reliability of 4 types of oilfield security sensors (camera, infrared, radar, vibration). Through an end-to-end deep learning model, dynamic prediction of fusion weights is realized, which completely solves the problem of sharp drop in accuracy of traditional fixed weight fusion under harsh environments.

[0021] 2. Three-in-one combined loss function design: The innovative combination of cross-entropy loss for environment recognition, L2 regularization loss for weights, and reliability matching loss enables the model to learn both environment classification ability and weight prediction ability at the same time. This ensures the accuracy of environment recognition and makes the predicted weights highly consistent with the actual reliability of the sensors, thus greatly improving the model training stability and weight prediction accuracy.

[0022] 3. Adaptive accuracy improvement mechanism for harsh environments: An adaptive optimization strategy based on weight amplification / reduction for harsh environments was designed. After identifying a harsh environment, the weight of reliable sensors is automatically amplified, the weight of unreliable sensors is reduced and renormalized, which improves the overall fusion accuracy in harsh environments by more than 40% and significantly enhances the environmental adaptability of the system.

[0023] 4. Dual-source data processing mechanism for harsh oilfield scenarios: A dual-source data supply system of synthetic data generation + real data loading has been constructed, which can simulate multi-source security data and reliability labels in batches under five typical oilfield environments. This effectively solves the industry pain points of scarce real data and high labeling costs in harsh oilfield environments and improves the generalization ability of the model.

[0024] 5. MLP Environment Recognition Architecture Adapted to Low-Dimensional Environmental Factors: The environment recognition module was reconstructed to address the characteristic of environmental factors being 3-dimensional numerical data. The MLP architecture was adopted instead of the traditional CNN architecture, which completely solved the RuntimeError problem caused by dimensionality mismatch. At the same time, the accuracy of environment recognition was improved to over 95%, providing a reliable environmental basis for dynamic weight prediction.

[0025] 6. This invention specifically relates to multi-source security data processing in harsh oilfield environments, environmental factor and sensor reliability mapping, dynamic weight prediction and weighted fusion technology, and is specifically applied to scenarios such as oilfield perimeter security, equipment monitoring, and intrusion detection. This invention achieves adaptive fusion of multi-sensor security data in harsh environments such as sandstorms, rain, snow, and low light, solving the industry pain point of traditional fixed-weight fusion methods experiencing a sharp drop in accuracy in complex environments. It constructs a complete technical system encompassing data generation, environmental identification, weight prediction, dynamic fusion, and visual interaction, significantly improving the reliability and accuracy of oilfield security systems in harsh environments. Attached Figure Description

[0026] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0027] Figure 1 This is the overall framework of the adaptive harsh environment oilfield multi-source security data dynamic weighted fusion system described in this invention; Figure 2 This is a flowchart of the adaptive harsh environment multi-source security data dynamic weighted fusion method for oilfields described in this invention. Detailed Implementation

[0028] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. The following examples will help those skilled in the art to further understand the present invention, but do not limit the present invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

[0029] Example 1, Combination Figure 1 This embodiment addresses the core shortcomings of existing oilfield multi-source security data fusion technologies by proposing an adaptive dynamic weighted fusion system for oilfield multi-source security data in harsh environments, aiming to achieve the following objectives: 1. Construct a quantitative mapping mechanism between environmental factors and sensor reliability, and dynamically adjust the fusion weights of each sensor according to the real-time environmental status to solve the problem of sharp drop in accuracy of fixed-weight fusion under harsh environments.

[0030] 2. Design a combined loss function that combines environment recognition loss, weight regularization loss, and reliability matching loss to improve the stability of model training and the accuracy of weight prediction.

[0031] 3. Establish a dual-source data processing mechanism of synthetic data generation + real data loading to simulate multi-source security data under different harsh environments in batches, making up for the shortage of real data in oil fields.

[0032] 4. The environment recognition module was reconstructed, and a multilayer perceptron (MLP) was used to adapt to low-dimensional environmental factor inputs, solving the problem of dimension mismatch and improving the accuracy of environment classification.

[0033] 5. Implement an adaptive accuracy enhancement mechanism for harsh environments. In harsh environments, the weight of reliable sensors is automatically strengthened and unreliable sensors are suppressed, thereby improving the overall fusion accuracy by more than 40%.

[0034] 6. Build an integrated visual operation interface that supports data generation, model training, real-time inference and result display, and adapts to the usage needs of oilfield field operation and maintenance personnel.

[0035] The oilfield multi-source security data dynamic weighted fusion system adapted to harsh environments described in this embodiment includes six core modules: dual-source data processing module, environment identification module, dynamic weight prediction module, dynamic weighted fusion module, training and inference module, and integrated visualization and interaction module. The dual-source data processing module is used to generate synthetic data, load real data, construct a dual-source data supply mechanism, and output a standardized dataset with environmental labels and sensor reliability labels. The environmental identification module is used to extract environmental factors from the standardized dataset, perform feature extraction through multi-level linear transformation and nonlinear activation, and output the probability distribution of oilfield environmental categories. The dynamic weight prediction module is used to map environmental factors into fusion weights of each sensor based on the MLP model, normalize the fusion weights, and output normalized fusion weights. The dynamic weighted fusion module is used to weight and fuse multi-sensor features according to the normalized fusion weights, perform adaptive weight adjustment in harsh environments, and output a unified feature vector after fusion. The training and inference module is used to complete model training and parameter optimization based on the combined loss function, and to perform real-time inference and performance evaluation. The integrated visual interaction module is used to realize full-process visual monitoring and log recording of parameter configuration, data processing, model training, inference execution and result display.

[0036] Furthermore, in this embodiment, the design concept, core improvements, and core formulas of the above modules are explained as follows: (1) Dual-source data processing module Design concept: Construct a dual data source supply mechanism of synthetic data and real data. Synthetic data generates multi-source datasets with complete labels by simulating the characteristics and reliability changes of sensor data under different harsh environments. Real data supports loading and parsing of standardized CSV format, realizing seamless switching between the two data sources and providing sufficient and diverse data support for model training.

[0037] Core improvements: It solves the problem of scarcity and difficulty in labeling real security data in the harsh environment of oil fields. The synthesized data can simulate sensor data and reliability labels in batches under five typical environments. Real data achieves automatic format verification and dimension conversion, improving data processing efficiency by more than 50%.

[0038] Key Design: 1. Synthetic Data Generation: The system presets the environmental factor ranges for five typical oilfield environments and automatically generates corresponding data on illumination, dust concentration, and rain / snow levels based on the environmental type. Simultaneously, based on the environment-reliability mapping relationship, it generates reliability labels and characteristic data for each sensor to simulate the degradation law of sensor performance under harsh environments.

[0039] 2. Real Data Loading: Supports loading real data in CSV format, automatically parses environmental factors, multi-sensor features, environmental labels and sensor reliability labels, completes data dimension verification and format conversion, and outputs a standardized dataset adapted to the model input.

[0040] (2) Environmental Recognition Module Design concept: In view of the low-dimensional numerical data of environmental factors, an MLP architecture is adopted to build an environmental recognition model to replace the traditional CNN architecture and solve the problem of dimensionality mismatch. Through multi-level linear transformation and non-linear activation, deep features of environmental factors are extracted to achieve accurate classification of five typical oilfield environments.

[0041] Core improvements: The environment recognition module has been reconstructed, and MLP has been adopted to adapt to 3D environmental factor input, which completely solves the dimensionality mismatch problem of the original CNN architecture; the accuracy of environment recognition has been improved to over 95%, providing a reliable environmental basis for subsequent dynamic weight prediction.

[0042] Core formula: The forward propagation process of the environment recognition model is as follows:

[0043] in, The input is the environmental factor matrix (B is the batch size). Here are the weight matrices for each layer. For bias terms, This represents the probability distribution of the output environment categories.

[0044] (3) Dynamic weight prediction module Design concept: Construct a dynamic weight prediction model based on MLP to map 3D environmental factors to the fusion weights of each sensor; normalize the weights through the Softmax layer to ensure that the sum of all weights is 1, so that the weight allocation conforms to the probability distribution characteristics.

[0045] Key improvements: For the first time, an end-to-end mapping relationship between environmental factors and sensor fusion weights has been established, which can automatically adjust the weights according to the real-time environment without manual intervention; the error between the weight prediction results and the actual reliability of the sensors is less than 5%, achieving a scientific and reasonable weight allocation.

[0046] Core formula: The forward propagation process for dynamic weight prediction is as follows:

[0047] in, The Softmax function guarantees that the sum of the weights in each row is 1, which is the output normalized weight matrix (S is the number of sensors).

[0048] (4) Dynamic weighted fusion module Design concept: Dynamic fusion of multi-sensor data is achieved based on the weighted average principle. The feature vectors of each sensor are multiplied by their corresponding prediction weights and then summed to output a unified feature vector after fusion. At the same time, an adaptive weight adjustment mechanism for harsh environments is designed to further improve the fusion accuracy under harsh conditions.

[0049] Key improvements: Dynamic weighted fusion of multi-sensor data has been achieved, replacing the traditional fixed-weight fusion; an adaptive weight adjustment mechanism for harsh environments has been introduced, which improves the overall fusion accuracy by more than 40% in harsh environments.

[0050] Core formula: 1. Basic weighted fusion:

[0051] in, Let D be the feature matrix of the i-th sensor (where D is the feature dimension). Let i be the weight of the i-th sensor. This is the feature matrix after fusion.

[0052] 2. Adaptive weight adjustment for harsh environments: When the current environment is identified as a harsh environment (dust storm, rain, snow, low light), the weights are adjusted as follows:

[0053] Adjusted and renormalized weights:

[0054] (5) Training reasoning module Design concept: Construct a combined loss function to simultaneously optimize the accuracy of environment recognition and the accuracy of weight prediction; use the AdamW optimizer to update model parameters and introduce weight decay to prevent overfitting; design a real-time inference process to achieve end-to-end execution of environment recognition, weight prediction, data fusion and accuracy improvement.

[0055] Core improvements: A combined loss function was designed, which includes environment recognition loss, weight regularization loss, and reliability matching loss, enabling the model to learn both environment classification and weight prediction capabilities simultaneously; an accuracy enhancement mechanism for harsh environments is automatically triggered during inference, achieving adaptive optimization.

[0056] Core formula: The combined loss function is:

[0057] in: Identify cross-entropy loss for the environment: , One-hot encoding for real-world environment labels. This represents the predicted probability of the environment category.

[0058] Weighted L2 regularization loss: This is used to prevent weights from becoming extreme.

[0059] For reliability matching loss: , For the first i The first sample j The true reliability of each sensor makes the prediction weights approximate the true reliability.

[0060] (6) Integrated visual interaction module Design concept: Based on a graphical user interface (GUI), an oilfield multi-source security data fusion system is built, integrating five core functions: data generation, data loading, model training, single inference, and batch inference; training curves, inference results, and sensor weight changes are displayed in real time, realizing full-process visualized operation and monitoring.

[0061] Key improvements: Low-code operation of the entire process of multi-source security data fusion in oilfields has been achieved, and the system can be deployed and used without the need for professional algorithm knowledge; real-time display of results and logging are supported to meet the needs of on-site operation and maintenance and auditing in oilfields.

[0062] In summary, the technical solution of this embodiment is based on a closed-loop process of dual-source data processing → environment identification → dynamic weight prediction → multi-source data fusion → training inference → visualization interaction, covering 6 core functional modules. All modules are customized for the harsh environmental characteristics of oil fields. The core formulas and judgment rules are derived from the core logic of environmental factor quantification, sensor reliability mapping, dynamic weight calculation and fusion loss optimization, realizing adaptive high-precision fusion of multi-source security data in complex environments.

[0063] Furthermore, in practical applications, the system architecture of this embodiment is centered on dynamic weight prediction and adaptive fusion in harsh environments, supported by dual-source data, MLP environment identification, and combined loss optimization, and uses full-process visualization as the operational carrier. It constructs a 6-layer closed-loop architecture, enabling data exchange and result feedback between each layer. The specific architecture layers are (from bottom to top) as follows: Figure 1 As shown: 1. Parameter Configuration Layer: Enables flexible configuration of core parameters such as environment type definition, number of sensors and feature dimensions, model training hyperparameters (learning rate, batch size, training epochs), and harsh environment identification.

[0064] 2. Dual-source data processing layer: Enables automatic generation of synthetic data in harsh environments and standardized loading of real security data from oil fields, outputting multi-source datasets with environmental labels and sensor reliability labels.

[0065] 3. Environmental Identification Layer: Based on the MLP architecture, the three-dimensional environmental factors (light intensity, dust concentration, rain and snow level) are classified, and the probability distribution of five typical oilfield environments (normal, dust, rain and snow, weak light, strong light) is output.

[0066] 4. Dynamic Weight Prediction Layer: Predicts the fusion weights of each sensor based on real-time environmental factors and outputs a normalized weight vector (summing to 1).

[0067] 5. Dynamic weighted fusion layer: Based on the predicted dynamic weights, the features of multiple sensors are weighted and averaged to fuse, and the unified feature vector after fusion is output.

[0068] 6. Visual Interaction Layer: Integrates the entire process of data generation / loading, model training, single / batch inference, and result display into a visual interface, enabling one-click operation and real-time result feedback; Example 2, Combination Figure 2 This embodiment describes a dynamic weighted fusion method for multi-source security data in oilfields that adapts to harsh environments. The specific implementation steps are S1-S7, which are executed sequentially and involve data exchange to form a complete technical closed loop. S1: System initialization and parameter configuration, completing hardware environment adaptation, core parameter settings, and model and module initialization; S2: Dual-source data preparation and loading, select synthetic data generation or real data loading, output standardized dataset and divide it into training set and test set; S3: Model training and parameter optimization. Based on the training set, the environment recognition model and dynamic weight prediction model are iteratively trained using a combination loss function and optimizer. The trained model and logs are saved. S4: Visualize training results, plot and display the training loss curve and the environment recognition accuracy curve, and output the core training metrics; S5: Single-time real-time inference, based on a single sample, performs environment recognition, dynamic weight prediction, dynamic weighted fusion and adaptive weight adjustment in harsh environments, and outputs inference results; S6: Batch inference and performance evaluation, performs inference on batch samples, calculates the accuracy of environment recognition and the accuracy improvement rate under harsh environments, and generates an evaluation report; S7: Results display and log recording. Visualize the inference results and record and save logs related to the entire process of operation, training and inference.

[0069] Furthermore, each of the above steps will be explained in detail; S1: System Initialization and Parameter Configuration Start the visualization system and complete hardware environment adaptation (automatic CPU / GPU identification) and initialization of each module; Configure core parameters: environment type, number of sensors and feature dimensions, model training hyperparameters (learning rate, batch size, training epochs), harsh environment labeling, etc. Initialize the environment recognition model, dynamic weight prediction model, and dynamic fusion module to prepare for subsequent processes.

[0070] S2: Preparation and Loading of Dual-Source Data Select the data source in the visualization interface: synthetic data or real data; If you choose synthetic data: set the number of samples, and the system will automatically generate multi-source security data for 5 typical environments, including environmental factors, sensor features, environmental labels and reliability labels; If you choose real data: Select a real data file in CSV format, and the system will automatically complete data parsing, format validation, and dimension conversion to output a standardized dataset; The processed dataset is divided into training and testing sets to provide data support for model training and performance verification.

[0071] S3: Model Training and Parameter Optimization Load the training set data and convert it into a model-compatible tensor format; The AdamW optimizer was used to iteratively train the environment recognition model and the dynamic weight prediction model with the combined loss function as the optimization objective. Real-time monitoring of training loss and environmental recognition accuracy, and recording of training process data; After training is complete, save the model weight file and training log.

[0072] S4: Visualization of Training Results The training loss curve and the environment recognition accuracy curve are plotted in the visualization interface to intuitively show the model training process; Output the final training results, including core metrics such as total loss value and environmental recognition accuracy.

[0073] S5: Single Real-Time Inference Randomly select a sample from the dataset and extract its environmental factors and multi-sensor features; Input the trained model and execute environment recognition, dynamic weight prediction, data fusion and adaptive weight adjustment for harsh environments in sequence; The output inference results include the identified environment type, the prediction weights of each sensor, the fused features, and the accuracy improvement rate.

[0074] S6: Batch Inference and Performance Evaluation 100 samples were randomly selected for batch inference, and the overall environmental recognition accuracy and the harsh environment recognition accuracy were statistically analyzed. Calculate the accuracy improvement under harsh environments and verify the adaptive optimization effect of the system; Output a batch inference results report, including various performance metrics and statistical data.

[0075] S7: Results Display and Log Recording The results of single-inference and batch-inference are displayed centrally in a visual interface; It automatically records the entire process of operation logs, training data, and inference results, and supports local saving and querying of log files.

[0076] In summary, the core innovations of this invention are reflected in five aspects. All innovations address the core defects of existing technologies and the characteristics of harsh oilfield environments, demonstrating significant technical originality and practical application value: 1. Dynamic mapping mechanism between environmental factors and sensor reliability: For the first time, a quantitative mapping relationship between 3D environmental factors (light intensity, dust concentration, rain and snow level) and the reliability of 4 types of oilfield security sensors (camera, infrared, radar, vibration) was established. Dynamic prediction of fusion weights was realized through an end-to-end deep learning model, which completely solved the problem of sharp drop in accuracy of traditional fixed weight fusion under harsh environments.

[0077] 2. Three-in-one combined loss function design: The innovative combination of cross-entropy loss for environment recognition, L2 regularization loss for weights, and reliability matching loss enables the model to learn both environment classification ability and weight prediction ability at the same time. This ensures the accuracy of environment recognition and makes the predicted weights highly consistent with the actual reliability of the sensors, thus greatly improving the model training stability and weight prediction accuracy.

[0078] 3. Adaptive accuracy improvement mechanism for harsh environments: An adaptive optimization strategy based on weight amplification / reduction for harsh environments was designed. After identifying a harsh environment, the weight of reliable sensors is automatically amplified, the weight of unreliable sensors is reduced and renormalized, which improves the overall fusion accuracy in harsh environments by more than 40% and significantly enhances the environmental adaptability of the system.

[0079] 4. Dual-source data processing mechanism for harsh oilfield scenarios: A dual-source data supply system of synthetic data generation + real data loading has been constructed, which can simulate multi-source security data and reliability labels in batches under five typical oilfield environments. This effectively solves the industry pain points of scarce real data and high labeling costs in harsh oilfield environments and improves the generalization ability of the model.

[0080] 5. MLP Environment Recognition Architecture Adapted to Low-Dimensional Environmental Factors: The environment recognition module was reconstructed to address the characteristic of environmental factors being 3-dimensional numerical data. The MLP architecture was adopted instead of the traditional CNN architecture, which completely solved the RuntimeError problem caused by dimensionality mismatch. At the same time, the accuracy of environment recognition was improved to over 95%, providing a reliable environmental basis for dynamic weight prediction.

[0081] Example 3: To verify the effectiveness and technical advantages of the technical solution of the present invention, two sets of comparative experiments were designed for the multi-source security data fusion scenario in the harsh environment of oil fields, to verify the performance advantages of the dynamic weighted fusion method and the harsh environment adaptive mechanism of the present invention respectively.

[0082] (1) Experimental environment and dataset settings Experimental hardware environment: CPU is Intel Core i7-12700K, GPU is NVIDIA RTX 3090Ti (24G VRAM), memory is 32G, software environment is Python 3.9, PyTorch 2.0, CUDA 11.8.

[0083] Experimental dataset: A synthetic dataset containing 2000 samples is used, covering 5 environmental categories: normal, dust, rain and snow, low light, and strong light, with 400 samples for each category. Each sample contains 3-dimensional environmental factors, 16-dimensional features from 4 sensors, environmental labels, and sensor reliability labels.

[0084] Evaluation metrics: Environmental identification accuracy, harsh environment identification accuracy, mean square error of weighted prediction (MSE), and harsh environment accuracy improvement rate were selected as the core evaluation metrics.

[0085] (2) Comparative experiment of dynamic weighted fusion and fixed weighted fusion The dynamic weighted fusion method of this invention is compared with the traditional fixed-weight fusion method (each sensor has a weight of 0.25). Tests were conducted under the same dataset and hardware environment. The core performance comparison is shown in Table 1. Table 1

[0086] Experimental results show that the dynamic weighted fusion method of the present invention slightly improves the fusion accuracy under normal conditions, and the fusion accuracy improves by 40% under harsh conditions, fully achieving the design goal of improving accuracy in harsh environments.

[0087] (3) Experiment to verify the effectiveness of the combined loss function The combined loss function of this invention is compared with the method using only cross-entropy loss to verify the improvement effect of the combined loss function on model performance. The experimental results are shown in Table 2: Table 2

[0088] Experimental results show that the combined loss function not only improves the accuracy of environment recognition by 6.5 percentage points, but also reduces the weight prediction MSE by 74.4%, while significantly shortening the convergence rounds of model training, thus proving the effectiveness of the combined loss function.

[0089] The above description of the technical solution provided by the present invention through several specific embodiments is intended to highlight the advantages and benefits of the technical solution provided by the present invention. However, the above-described specific embodiments are not intended to limit the present invention. Any reasonable modifications and improvements to the present invention, reasonable combinations of implementation methods and equivalent substitutions based on the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An oilfield multi-source security data dynamic weighted fusion system that adapts to harsh environments, characterized in that, The system includes: The dual-source data processing module is used to generate synthetic data, load real data, build a dual-source data supply mechanism, and output a standardized dataset with environmental labels and sensor reliability labels. The environmental identification module is used to extract environmental factors from the standardized dataset, perform feature extraction through multi-level linear transformation and nonlinear activation, and output the probability distribution of oilfield environmental categories. The dynamic weight prediction module is used to map environmental factors to fusion weights of each sensor based on the MLP model, normalize the fusion weights, and output normalized fusion weights. The dynamic weighted fusion module is used to weight and fuse multi-sensor features according to the normalized fusion weights, perform adaptive weight adjustment in harsh environments, and output a unified feature vector after fusion. The training and inference module is used to complete model training and parameter optimization based on the combined loss function, and to perform real-time inference and performance evaluation. The integrated visual interaction module is used to realize full-process visual monitoring and log recording of parameter configuration, data processing, model training, inference execution and result display.

2. The adaptive harsh-environment oilfield multi-source security data dynamic weighted fusion system according to claim 1, characterized in that, The specific data generation includes: generating a synthetic multi-source dataset with environmental and sensor reliability labels by simulating the characteristics of sensor data under different harsh environments and their corresponding reliability change patterns; The generation of real data specifically includes: loading real data in CSV format at least, and automatically parsing the loaded data. Environmental factors, multi-sensor features, environmental labels, and sensor reliability labels are used to complete data dimension verification and format conversion, and output a standardized dataset that adapts to the model input.

3. The adaptive harsh-environment oilfield multi-source security data dynamic weighted fusion system according to claim 1, characterized in that, The environment recognition model is constructed using an MLP architecture; the environment recognition model outputs the probability distribution of environment categories through a forward propagation formula, which is: wherein, is the input environmental factor matrix (B is the batch size), is the weight matrix of each layer, is the bias term, is the output environmental category probability distribution.

4. The adaptive harsh-environment oilfield multi-source security data dynamic weighted fusion system according to claim 1, characterized in that, The forward propagation process of the dynamic weight prediction module is as follows: wherein, The Softmax function guarantees that the sum of the weights in each row is 1 for the output normalized weight matrix (S is the number of sensors).

5. The adaptive harsh-environment oilfield multi-source security data dynamic weighted fusion system according to claim 1, characterized in that, The dynamic weighted fusion module realizes the dynamic fusion of multi-sensor data based on the weighted average principle. It multiplies the feature vectors of each sensor with their corresponding prediction weights and sums them to output a unified feature vector after fusion.

6. The adaptive harsh-environment oilfield multi-source security data dynamic weighted fusion system according to claim 1, characterized in that, The specific steps for performing adaptive weight adjustment under adverse conditions are as follows: When the current environment is identified as a harsh environment, the weights are adjusted as follows: Adjusted and renormalized weights: wherein, is the weight of the i th sensor.

7. The adaptive harsh-environment oilfield multi-source security data dynamic weighted fusion system according to claim 1, characterized in that, The training and inference module also uses the AdamW optimizer to update model parameters and introduces weight decay to prevent overfitting.

8. The adaptive harsh-environment oilfield multi-source security data dynamic weighted fusion system according to claim 1, characterized in that, The combined loss function of the training inference module is: in: To identify cross-entropy loss for the environment, The weighted L2 regularization loss, This is due to reliability matching loss.

9. The adaptive harsh environment multi-source security data dynamic weighted fusion system for oilfields according to claim 1, characterized in that, The system constructs a six-layer closed-loop architecture, consisting of a parameter configuration layer, a dual-source data processing layer, an environment identification layer, a dynamic weight prediction layer, a dynamic weighted fusion layer, and a visualization interaction layer, from bottom to top. Data is exchanged and results are fed back between each layer, forming a complete technical closed loop.

10. The fusion method implemented by the adaptive harsh environment multi-source security data dynamic weighted fusion system for oilfields as described in any one of claims 1-9, characterized in that, The methods include: S1: System initialization and parameter configuration, completing hardware environment adaptation, core parameter settings, and model and module initialization; S2: Dual-source data preparation and loading, select synthetic data generation or real data loading, output standardized dataset and divide it into training set and test set; S3: Model training and parameter optimization. Based on the training set, the environment recognition model and dynamic weight prediction model are iteratively trained using a combination loss function and optimizer. The trained model and logs are saved. S4: Visualize training results, plot and display the training loss curve and the environment recognition accuracy curve, and output the core training metrics; S5: Single-time real-time inference, based on a single sample, performs environment recognition, dynamic weight prediction, dynamic weighted fusion and adaptive weight adjustment in harsh environments, and outputs inference results; S6: Batch inference and performance evaluation, performs inference on batch samples, calculates the accuracy of environment recognition and the accuracy improvement rate under harsh environments, and generates an evaluation report; S7: Results display and log recording. Visualize the inference results and record and save logs related to the entire process of operation, training and inference.

Images