A cross-system risk data fusion method and system based on federated learning

Abstract

This invention belongs to the field of data processing technology, and in particular relates to a cross-system risk data fusion method and system based on federated learning. This invention solves the defect of one-size-fits-all aggregation in traditional federated learning by dynamically adjusting weights based on real-time verification indicators and data volume, so that high-value data can automatically obtain higher contributions. Under the premise of not sharing the original data at all, it achieves a faster response speed than centralized training through encrypted aggregation and dynamic weight allocation.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to a cross-system risk data fusion method and system based on joint learning. Background Technology

[0002] Risk monitoring data for port infrastructure is scattered across multiple independent systems, such as equipment operation and maintenance systems, meteorological monitoring systems, ship AIS systems, and customs logistics systems. These systems are typically managed by different entities, such as port operators, equipment suppliers, and regulatory authorities.

[0003] With the rapid development of information technologies such as artificial intelligence, big data, and the Internet of Things, intelligent decision-making algorithms have been widely applied in various fields. Introducing intelligent decision-making algorithms into the field of emergency response to operational risks in port infrastructure, through the fusion analysis of multi-source data, the optimization and construction of intelligent models, and real-time dynamic risk assessment, can provide scientific and accurate decision support for port emergency management. This not only helps improve the speed and efficiency of port emergency response, enabling timely and effective countermeasures and reducing accident losses, but also enhances the overall operational safety and reliability of the port, ensuring the smooth flow of the logistics supply chain, and has significant practical implications for promoting the stable development of the regional economy. Existing technologies include solutions for port risk data fusion using federated learning. Traditional distributed training in federated learning typically employs fixed-weight aggregation and static iteration strategies, such as pre-setting a fixed training cycle. However, in the scenario of port infrastructure risk fusion, the following problems exist: the contribution of data from different participants to risk scoring varies significantly; the contribution of crane stress sensor data to equipment failure risk is far higher than that of conventional wind speed data from weather stations; the port environment and operational status cause data distribution to change over time, making fixed aggregation strategies unsuitable; some nodes may upload noisy parameters or low-quality updates, affecting the stability of the global model. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a cross-system risk data fusion method and system based on joint learning, thereby resolving the issues existing in the prior art.

[0005] This invention provides a cross-system risk data fusion method based on joint learning, comprising the following steps:

[0006] S1: Acquire operational data from various monitoring systems, including equipment monitoring systems, environmental sensing systems, and operation management systems;

[0007] S2: Perform data preprocessing on the operational data of each monitoring system;

[0008] S3: Construct a joint learning model for cross-system risk data fusion; the joint learning model includes a local model, a global model, a secure aggregation protocol, and a communication mechanism;

[0009] S4: Train the joint learning model;

[0010] S4.1: Global model initialization; S4.2: Determine dynamic aggregation weights based on the initialized global model; S4.3: The federated coordinator performs weighted aggregation of the parameters of the participants based on the dynamic aggregation weights to obtain the global model weights, thereby achieving training;

[0011] S5: Input the preprocessed operational data of each monitoring system into the trained joint learning model to obtain the risk data fusion result.

[0012] Preferably, step S4.2 specifically involves: each participant using its local training set to train a local model of the initialized global model, optimizing the local risk classification loss; each participant allocating 5%-10% of its local training set as a validation set, calculating the validation metrics of the local model; and calculating the validation metrics based on the number of samples N in this round of local training. i (t) Calculate the basic weights;

[0013] Among them, the base weights of the i-th participant in t rounds of training The calculation formula is:

[0014] ;

[0015] In this step, the validation metrics are AUC value, F1 score, or risk classification accuracy.

[0016] The federal coordinator collects verification metrics from each participating party. Basic weights Historical credibility Calculate dynamic aggregate weights ;

[0017] The calculation formula is: ;

[0018] In the formula, α, β, γ are hyperparameters that are dynamically adjusted through the meta-learning process of the federal coordinator, and Norm() is the normalization function.

[0019] Preferably, step S4.3 specifically involves: each participating party using the TLS protocol to update the parameters of the local model. Encrypted upload; the federal coordinator performs weighted aggregation of the parameter update amounts; the specific formula is:

[0020]

[0021] ;

[0022] In the formula, Let $\mathbf{t}$ be the global model weight update amount during the $t$-th training round. Let be the global model weights for the t-th training round. The global model weights for training in round t-1 Let N be the weights of the i-th local model trained in the t-th round, and N be the total number of samples. This represents the learning rate of the global model.

[0023] Preferably, in step S4.1, the global model's federal coordinator pre-trains a lightweight artificial neural network model based on historical risk data and distributes it to all participants. Each participant registers metadata with the federal coordinator upon first connecting with the federal coordinator. The metadata includes: data type, data volume, data timeliness, historical risk association tags, and initial credibility score. The data type includes equipment monitoring data, environmental monitoring data, and business operation data; the data volume is the number of historical samples; and the data timeliness is the most recent data update time.

[0024] Preferably, in step S2, the data preprocessing includes data cleaning, spatiotemporal alignment, feature extraction, and feature standardization.

[0025] Preferably, the data cleaning specifically includes outlier removal and missing value repair; the outlier removal is achieved using statistical methods, specifically by calculating the mean (μ) and standard deviation (σ) within a sliding window for the time-series data, and removing outliers that exceed the range of [μ-3σ, μ+3σ]; the missing value repair specifically involves: for numerical data, linear interpolation is used to repair missing values; for non-numerical data, if key fields are missing, default values ​​are used to fill them.

[0026] Preferably, the spatiotemporal alignment includes time alignment and spatial alignment; the time alignment specifically involves sampling the operating data of different systems into a fixed time window, then downsampling the high-frequency data, upsampling the low-frequency data, and converting all timestamps into a unified format; the spatial alignment specifically involves converting data from different coordinate systems into the same reference.

[0027] Preferably, the feature extraction specifically includes: feature extraction from equipment monitoring data, feature extraction from environmental monitoring data, and feature extraction from business operation data; the feature extraction from equipment monitoring data specifically includes: for each type of data, extracting the mean, standard deviation, maximum value, minimum value, and peak factor of that type of data; and performing a fast Fourier transform on each type of data to extract the main frequency component and frequency band energy ratio of that type of data; and extracting the sliding window trend and difference features of that type of data; the feature extraction from environmental monitoring data specifically includes: extracting the maximum wind speed, maximum temperature, minimum visibility, extreme precipitation, and 24-hour cumulative precipitation from the environmental monitoring data; the business operation data is a mixed categorical / numerical feature, wherein, for numerical business data, extracting the cargo stacking height, continuous equipment operation time, and ship berthing density from the business operation data; for categorical business data, extracting text log keywords from the business operation data, and converting the text log keywords into binary data or category labels through natural language processing.

[0028] Preferably, in step S3, the input of the global model is the feature vector extracted by each of the local models, and the output is a unified risk score; the global model is a lightweight artificial neural network model, which includes: an input layer, the dimension of which is the total number of features of the feature vector; a hidden layer, which includes 1-2 fully connected layers, each with 128 neurons and a ReLU activation function, used to learn the interaction relationship of the standardized feature set; and an output layer, which includes one neuron and a Sigmoid activation function, outputting a risk score P∈[0,1]. For the local model, each participant customizes the local model according to the local system type. For device monitoring data, the local model is a long short-term memory network model, and the input of the long short-term memory network is a 5-minute window. The local model is a convolutional neural network (CNN) model for environmental monitoring data. The CNN model's input consists of environmental features within a 5-minute window. The CNN model's structure includes one CNN layer and one fully connected layer. The output of the CNN model is a 16-dimensional feature vector representing the environmental risk contribution. For business operation data, the local model is a fully connected neural network (WNN) model. The WNN model's input consists of business features. The WNN model's structure includes one embedded layer and two fully connected layers. The output of the WNN model is a 16-dimensional feature vector representing operational management risk.

[0029] According to another aspect of the present invention, a cross-system risk data fusion system based on joint learning is provided, the system employing the aforementioned cross-system risk data fusion method based on joint learning, the system comprising:

[0030] The data acquisition module is used to acquire the operational data of various monitoring systems, including equipment monitoring systems, environmental sensing systems, and operation management systems.

[0031] The data preprocessing module is used to perform data preprocessing operations on the operating data of the various monitoring systems.

[0032] A joint learning model building module is used to construct a joint learning model for cross-system risk data fusion; the joint learning model includes a local model, a global model, a secure aggregation protocol, and a communication mechanism.

[0033] A joint learning model training module is used to train the joint learning model;

[0034] The data fusion module is used to input the preprocessed operational data of each monitoring system into the trained joint learning model to obtain the risk data fusion result.

[0035] The embodiments of the present invention have the following technical effects:

[0036] This invention addresses the shortcomings of traditional federated learning's one-size-fits-all aggregation by dynamically adjusting weights based on real-time verification metrics and data volume. This allows high-value data to automatically receive higher contributions, and achieves a faster response speed than centralized training through encrypted aggregation and dynamic weight allocation, without sharing the original data at all. Attached Figure Description

[0037] 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.

[0038] Figure 1 This is a flowchart of a cross-system risk data fusion method based on joint learning provided in an embodiment of the present invention;

[0039] Figure 2 This is a flowchart for training a joint learning model, provided as an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0041] Example 1, such as Figure 1 As shown, a cross-system risk data fusion method based on joint learning includes the following steps:

[0042] S1: Acquire operational data from various monitoring systems, including equipment monitoring systems, environmental sensing systems, and operation management systems;

[0043] In the complex environment of seaports, infrastructure operation risks are influenced by multiple coupled factors, and related data are scattered across several independently managed systems. These systems may be controlled by different entities such as port operators, equipment suppliers, regulatory authorities, and third-party service providers, and their data types, management permissions, and security requirements vary significantly.

[0044] The equipment monitoring system comprises physical sensors and supporting systems used to directly monitor the operational status of key port infrastructure equipment, with data reflecting the real-time health status of the port infrastructure. The port equipment monitoring system includes a crane monitoring subsystem, a conveyor belt and storage equipment monitoring subsystem, a power and energy supply monitoring subsystem, and a port machinery control monitoring subsystem. The crane monitoring subsystem includes a crane main beam stress sensor to monitor the stress on the crane main beam steel structure; a hoisting mechanism vibration sensor to detect abnormalities in the hoisting mechanism wire rope / reducer; a motor temperature sensor to reflect electrical load; and a brake pressure sensor to determine the braking performance of the brake. The conveyor belt and storage equipment monitoring subsystem includes a belt misalignment sensor to detect material transport deviation; a roller bearing vibration sensor to monitor wear of transmission components; and a stacker crane position encoder to determine the accuracy of cargo stacking. The power and energy supply monitoring subsystem includes a transformer oil temperature sensor to monitor the insulation status of the transformer; a distribution cabinet current / voltage sensor to determine the overload risk of the distribution cabinet; and an uninterruptible power supply (UPS) status monitoring module to detect potential power outages.

[0045] The environmental perception system is a system that monitors the external natural and operational environment of the port. Its data indirectly affects the port infrastructure risks through the effects of extreme weather causing equipment overload, reduced visibility leading to collisions, etc.

[0046] The environmental perception system includes a meteorological monitoring station, a hydrological monitoring subsystem, and a visibility and lighting monitoring subsystem. The meteorological monitoring station includes an anemometer (to measure instantaneous / average wind speed in m / s), a wind vane (to monitor prevailing wind direction), a rain gauge (to monitor cumulative rainfall in mm), a temperature and humidity sensor (to monitor ambient temperature / humidity in ℃ / %), and a barometer (to monitor atmospheric pressure changes). The hydrological monitoring subsystem includes a tide meter (to monitor real-time water level in m), a wave sensor (to measure wave height / period in m / s), and a current meter (to monitor water flow velocity in m / s). The visibility and lighting monitoring subsystem includes a visibility meter (to monitor visibility distance in fog / haze weather in m) and a dock lighting brightness sensor (to determine visibility for nighttime operations).

[0047] The operation management system includes an Automatic Identification System (AIS), a terminal operation scheduling system, and a warehouse management system. The AIS is used to obtain the real-time location (latitude and longitude), speed (knots), draft (m), estimated time of arrival (ETA), and cargo type (container / bulk cargo). The terminal operation scheduling system is used to obtain loading and unloading task plans, cargo stacking height records, and truck dispatch logs. The warehouse management system is used to obtain cargo inventory lists, temperature records of temperature-controlled warehouses, and rack load-bearing monitoring data.

[0048] S2: Perform data preprocessing on the operational data of each monitoring system;

[0049] The data preprocessing includes data cleaning, spatiotemporal alignment, feature extraction, and feature standardization.

[0050] The data cleaning process involves removing invalid data and supplementing missing data to ensure the temporal continuity of the running data. Specifically, it includes outlier removal and missing value repair. Outlier removal employs statistical methods, specifically calculating the mean (μ) and standard deviation (σ) within a sliding window for the time-series data and removing outliers exceeding the range [μ-3σ, μ+3σ]. Missing value repair involves using linear interpolation for numerical data and filling missing values ​​with default values ​​for non-numerical data if key fields are missing.

[0051] The spatiotemporal alignment is used to solve the problems of time asynchrony and spatial reference system inconsistency in the operating data of different systems, ensuring that the characteristics of the operating data of different systems are comparable in the same spatiotemporal dimension. The spatiotemporal alignment includes time alignment and spatial alignment. The time alignment specifically involves sampling the operating data of different systems into a fixed time window, then downsampling the high-frequency data and upsampling the low-frequency data, and converting all timestamps into a unified format. The spatial alignment specifically involves converting the data of different coordinate systems into the same reference.

[0052] The feature extraction is used to extract numerical features that directly reflect or are strongly correlated with risks from the cleaned and aligned operational data; specifically: feature extraction of equipment monitoring data, feature extraction of environmental monitoring data, and business operation data.

[0053] The specific features extraction of the equipment monitoring data includes: for each type of data, extracting the mean, standard deviation, maximum value, minimum value, and peak factor; the mean and standard deviation are used to describe the central tendency and dispersion of the data, and the maximum and minimum values ​​are used to capture the extreme states of the data; and performing a Fast Fourier Transform on each type of data to extract the dominant frequency component and frequency band energy proportion; the dominant frequency component is used to describe the frequency of energy concentration, and the frequency band energy proportion is used to reflect the abnormal contribution of different frequency bands; and for each type of data, extracting the sliding window trend and differential features, the sliding window trend slope is used to determine whether the equipment is in an accelerated aging state, and the differential features are used to monitor abrupt changes.

[0054] The specific feature extraction of the environmental monitoring data involves extracting the maximum wind speed, highest temperature, lowest visibility, extreme precipitation, and 24-hour cumulative precipitation from the environmental monitoring data.

[0055] The business operation data has mixed categorical and numerical features. For numerical business data, the stacking height of goods, continuous operation time of equipment, and berthing density of ships are extracted. For categorical business data, text log keywords are extracted, including overload warning, loose screws, maintenance status, and operation priority. Natural language processing (NLP) is used to convert the text log keywords into binary data or category labels.

[0056] S3: Construct a joint learning model for cross-system risk data fusion; the joint learning model includes a local model, a global model, a secure aggregation protocol, and a communication mechanism;

[0057] The joint learning model adopts a hierarchical architecture and includes the following core components:

[0058] A local model is deployed on each participating node, such as a port equipment operation and maintenance system, a weather station, and a dispatch center. The local model is customized based on local data features and is responsible for local data training and parameter updates. The parameters include the weights and gradients of the local model.

[0059] The global model is maintained by the federal coordinator. In this embodiment, the federal coordinator is the port data management center, which aggregates the parameters of each local model to generate a unified risk assessment function.

[0060] The secure aggregation protocol is used to ensure that the model parameters uploaded by each participating party are encrypted during transmission, can only be aggregated by the federal coordinator, and cannot be decrypted individually, thus ensuring privacy and security.

[0061] The communication mechanism involves the local model and the federal coordinator exchanging parameters through an encrypted channel, typically employing an iterative process of local training, parameter uploading, global aggregation, and model distribution. In this embodiment, the encrypted channel uses the TLS protocol.

[0062] Furthermore, the input to the global model is the feature vector extracted by each of the local models, and the output is a unified risk score; in this embodiment, the global model is a lightweight artificial neural network model, which includes:

[0063] An input layer, the dimension of which is the total number of features of the feature vector;

[0064] The hidden layer comprises 1-2 fully connected layers, each with 128 neurons and a ReLU activation function, used to learn the interaction relationships of the standardized feature set.

[0065] The output layer consists of one neuron, with the activation function being the Sigmoid activation function, and outputs a risk score P∈[0,1].

[0066] For the local model, each participant customizes the local model according to the local system type. The core objective is to extract the risk-sensitive features of the data and map them into an intermediate representation compatible with the global model. Specifically, for device monitoring data, the local model is a Long Short-Term Memory (LSTM) network model. The input of the LSTM network is a sequence of device features within a 5-minute window. The structure of the LSTM network includes one LSTM layer and one fully connected layer to extract the temporal dependencies of the features of the device monitoring data. The output of the LSTM network is a feature vector representing the current risk state of the device. For environmental monitoring data, the local model is a convolutional neural network (CNN) model. The input of the CNN model is environmental features within a 5-minute window. The structure of the CNN model includes one CNN layer and one fully connected layer. The output of the CNN model is a 16-dimensional feature vector representing the contribution of environmental risk. For business operation data, the local model is a fully connected neural network (MLP) model. The input of the MLP model is business features. The structure of the MLP model includes one embedded layer and two fully connected layers. The output of the MLP model is a 16-dimensional feature vector representing operational management risk.

[0067] In this model, the feature vectors output by each local model are concatenated into a unified feature vector in the global model, and then the multi-source interaction relationship is learned through the hidden layer of the global model.

[0068] S4: Train the joint learning model;

[0069] Traditional distributed training in joint learning typically employs fixed-weight aggregation and static iteration strategies, such as pre-setting a fixed number of training rounds. However, in the scenario of risk fusion for port infrastructure, the following problems exist: the contribution of data from different participants to risk scoring varies significantly, with crane stress sensor data contributing far more to equipment failure risk than conventional wind speed data from weather stations; the port environment and operational status cause data distribution to change over time, making fixed aggregation strategies unsuitable; and some nodes may upload noisy parameters or low-quality updates, affecting the stability of the global model.

[0070] To address this, this embodiment proposes a federated training process with dynamic weight adaptive aggregation, which improves the convergence speed of the joint learning model and the accuracy of risk scoring by dynamically adjusting the contribution weights of the participants.

[0071] Specifically, such as Figure 2 As shown, S4 specifically includes:

[0072] S4.1: Global model initialization;

[0073] The global model's federal coordinator pre-trains a lightweight artificial neural network model based on historical risk data and distributes it to all participants. Each participant registers metadata with the federal coordinator upon their first connection with the federal coordinator. This metadata includes: data type, data volume, data timeliness, historical risk association tags, and initial credibility score. The data type includes equipment monitoring data, environmental monitoring data, and business operation data; the data volume is the number of historical samples; and the data timeliness is the most recent data update time.

[0074] S4.2: Determine the dynamic aggregation weights based on the initialized global model;

[0075] Each participant trains a local model of the initialized global model using its local training set, optimizing the local risk classification loss; each participant allocates 5%-10% of its local training set as a validation set, calculates the validation metric of the local model, and calculates the validation metric based on the number of samples N in this round of local training. i (t) Calculate the basic weights;

[0076] Among them, the base weights of the i-th participant in t rounds of training The calculation formula is:

[0077] ;

[0078] In this step, the validation metrics are AUC value, F1 score, or risk classification accuracy.

[0079] The federal coordinator collects verification metrics from each participating party. Basic weights Historical credibility Calculate dynamic aggregate weights ;

[0080] The calculation formula is: ;

[0081] In the formula, α, β, γ are hyperparameters that are dynamically adjusted through the meta-learning process of the federal coordinator, and Norm() is the normalization function.

[0082] S4.3: The federal coordinator performs weighted aggregation of the parameters of the participating parties according to the dynamic aggregation weights to obtain the global model weights, thereby achieving training;

[0083] In this process, each participating party uses the TLS protocol to update the parameters of the local model. Encrypted upload; the federal coordinator performs weighted aggregation of the parameter update amounts; the specific formula is:

[0084]

[0085] ;

[0086] In the formula, Let $\mathbf{t}$ be the global model weight update amount during the $t$-th training round. Let be the global model weights for the t-th training round. The global model weights for training in round t-1 Let N be the weights of the i-th local model trained in the t-th round, and N be the total number of samples. This represents the learning rate of the global model.

[0087] This embodiment addresses the shortcomings of traditional federated learning's one-size-fits-all aggregation by dynamically adjusting weights based on real-time verification metrics and data volume. This allows high-value data to automatically receive higher contributions, and achieves a faster response speed than centralized training through encrypted aggregation and dynamic weight allocation, without sharing the original data at all.

[0088] S5: Input the preprocessed operational data of each monitoring system into the trained joint learning model to obtain the risk data fusion result.

[0089] The feature data of each participant is input into the trained global model, which outputs a comprehensive risk score.

[0090] In this embodiment, when the comprehensive risk score is ≥0.7, the risk data fusion result is high risk; when the comprehensive risk score is between 0.4 and 0.7, the risk data fusion result is medium risk; and when the comprehensive risk score is ≤0.4, the risk data fusion result is low risk.

[0091] Example 2: This invention also provides a cross-system risk data fusion system based on joint learning. The system employs a cross-system risk data fusion method based on joint learning as described in Example 1. The system includes:

[0092] The data acquisition module is used to acquire the operational data of various monitoring systems, including equipment monitoring systems, environmental sensing systems, and operation management systems.

[0093] The data preprocessing module is used to perform data preprocessing operations on the operating data of the various monitoring systems.

[0094] A joint learning model building module is used to construct a joint learning model for cross-system risk data fusion; the joint learning model includes a local model, a global model, a secure aggregation protocol, and a communication mechanism.

[0095] A joint learning model training module is used to train the joint learning model;

[0096] The data fusion module is used to input the preprocessed operational data of each monitoring system into the trained joint learning model to obtain the risk data fusion result.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A cross-system risk data fusion method based on joint learning, characterized in that, The method includes the following steps: S1: Acquire operational data from various monitoring systems, including equipment monitoring systems, environmental sensing systems, and operation management systems; S2: Perform data preprocessing on the operational data of each monitoring system; S3: Construct a joint learning model for cross-system risk data fusion; the joint learning model includes a local model, a global model, a secure aggregation protocol, and a communication mechanism; S4: Train the joint learning model; S4.1: Global model initialization; S4.2: Determine dynamic aggregation weights based on the initialized global model; S4.2 specifically involves: each participant using its local training set to train a local model of the initialized global model, optimizing the local risk classification loss; each participant allocating 5%-10% of its local training set as a validation set, calculating the validation metrics of the local model; and calculating the validation metrics based on the number of samples N in this round of local training. i (t) Calculate the basic weights; Among them, the base weights of the i-th participant in t rounds of training The calculation formula is: ； In this step, the validation metrics are AUC value, F1 score, or risk classification accuracy. The federal coordinator collects verification metrics from all participating parties. Basic weights Historical credibility Calculate dynamic aggregate weights ; The calculation formula is: ; In the formula, α, β, γ are hyperparameters that are dynamically adjusted through the meta-learning process of the federal coordinator, and Norm() is a normalization function. S4.3: The federal coordinator performs weighted aggregation of the parameters of the participating parties according to the dynamic aggregation weights to obtain the global model weights, thereby achieving training; S5: Input the preprocessed operational data of each monitoring system into the trained joint learning model to obtain the risk data fusion result.

2. The cross-system risk data fusion method based on joint learning according to claim 1, characterized in that, Specifically, S4.3 involves each participating party using the TLS protocol to update the parameters of the local model. Encrypted upload; the federal coordinator performs weighted aggregation of the parameter update amounts; the specific formula is: ； In the formula, Let $\mathbf{t}$ be the global model weight update amount during the $t$-th training round. Let be the global model weights for the t-th training round. These are the global model weights trained in the (t-1)th round. Let N be the weights of the i-th local model trained in the t-th round, and N be the total number of samples. This represents the learning rate of the global model.

3. The cross-system risk data fusion method based on joint learning according to claim 2, characterized in that, In S4.1, the federal coordinator of the global model pre-trains a lightweight artificial neural network model based on historical risk data and distributes it to all participating parties. Each participating party registers metadata with the federal coordinator upon its first connection with the federal coordinator. The metadata includes: data type, data volume, data timeliness, historical risk association tags, and initial credibility score. The data type includes equipment monitoring data, environmental monitoring data, and business operation data. The data volume is the number of historical samples, and the data timeliness is the most recent data update time.

4. The cross-system risk data fusion method based on joint learning according to claim 1, characterized in that, In S2, the data preprocessing includes data cleaning, spatiotemporal alignment, feature extraction, and feature standardization.

5. The cross-system risk data fusion method based on joint learning according to claim 4, characterized in that, The data cleaning specifically includes outlier removal and missing value repair. Outlier removal is achieved using statistical methods, specifically by calculating the mean μ and standard deviation σ within a sliding window for the time-series data, and removing outliers that exceed the range [μ-3σ, μ+3σ]. Missing value repair is achieved by using linear interpolation for numerical data and filling missing values ​​with default values ​​for non-numerical data if key fields are missing.

6. The cross-system risk data fusion method based on joint learning according to claim 4, characterized in that, The spatiotemporal alignment includes time alignment and spatial alignment; the time alignment specifically involves sampling the operating data of different systems into a fixed time window, then downsampling the high-frequency data and upsampling the low-frequency data, and converting all timestamps into a unified format; the spatial alignment specifically involves converting data from different coordinate systems into the same reference.

7. The cross-system risk data fusion method based on joint learning according to claim 4, characterized in that, The feature extraction specifically includes: feature extraction of equipment monitoring data, feature extraction of environmental monitoring data, and feature extraction of business operation data; the feature extraction of equipment monitoring data specifically includes: for each type of data, extracting the mean, standard deviation, maximum value, minimum value, and peak factor of that type of data; and performing a fast Fourier transform on each type of data to extract the main frequency component and frequency band energy ratio of that type of data; and extracting the sliding window trend slope and difference features of that type of data. The environmental monitoring data feature extraction specifically involves: extracting the maximum wind speed, highest temperature, lowest visibility, extreme precipitation, and 24-hour cumulative precipitation from the environmental monitoring data; the business operation data is a hybrid categorical / numerical feature, wherein, for numerical business data, the cargo stacking height, continuous equipment operation time, and ship berthing density are extracted from the business operation data; for categorical business data, text log keywords are extracted from the business operation data, and the text log keywords are converted into binary data or category labels through natural language processing.

8. The cross-system risk data fusion method based on joint learning according to claim 1, characterized in that, In S3, the input of the global model is the feature vector extracted by each of the local models, and the output is a unified risk score. The global model is a lightweight artificial neural network model, which includes: an input layer, the dimension of which is the total number of features of the feature vector; a hidden layer, which includes 1-2 fully connected layers, each with 128 neurons and a ReLU activation function, used to learn the interaction relationship of the standardized feature set; and an output layer, which includes 1 neuron and a Sigmoid activation function, outputting a risk score P∈[0,1].

9. A cross-system risk data fusion system based on joint learning, characterized in that, The system employs a cross-system risk data fusion method based on joint learning as described in any one of claims 1-8, and the system comprises: The data acquisition module is used to acquire the operational data of various monitoring systems, including equipment monitoring systems, environmental sensing systems, and operation management systems. The data preprocessing module is used to perform data preprocessing operations on the operating data of the various monitoring systems. A joint learning model building module is used to construct a joint learning model for cross-system risk data fusion; the joint learning model includes a local model, a global model, a secure aggregation protocol, and a communication mechanism. A joint learning model training module is used to train the joint learning model; The data fusion module is used to input the preprocessed operational data of each monitoring system into the trained joint learning model to obtain the risk data fusion result.

Images