An MLP-based electrical equipment operating environment intelligent early warning method and system suitable for offshore booster stations

Abstract

The application discloses an intelligent early warning method and system for an electrical equipment operation environment of a sea-based booster station based on an MLP, relates to the technical field of offshore wind power, and comprises the following steps: constructing a first data set and a second data set, combining a transfer learning and a continuous label smoothing strategy, effectively solving the problem of high dependence on historical data of a target station, and only needing to collect a small amount of target station data to complete model adaptation, thereby reducing data collection costs and model deployment periods and improving the generalization ability of the model in different sea-based booster station scenarios. The MLP model with classification and regression functions can simultaneously output equipment failure probability and environmental parameter optimization target values, and realizes early warning and adjustment integration. The Bayesian optimization algorithm is introduced to optimize the MLP model parameters, the L2 regularization and Dropout technologies are combined, the prediction accuracy and the anti-overfitting ability of the model are improved, and the PID closed-loop control of the dynamic adjustment module is used to realize accurate adjustment of the environmental parameters.

Description

Technical Field

[0001] This invention relates to the field of offshore wind power technology, and more specifically to an intelligent early warning method and system for the operating environment of electrical equipment based on MLP, applicable to offshore substations. Background Technology

[0002] Currently, offshore substations are the core facilities for collecting, boosting, and transmitting electricity in offshore wind farms. Their internal electrical equipment is exposed to harsh marine environments with high humidity, high salt spray concentration, strong vibration, and drastic temperature fluctuations for extended periods. As a result, problems such as decreased equipment insulation performance, corrosion of metal parts, and wear of mechanical structures occur frequently, and the risk of failure is significantly higher than that of land-based electrical equipment.

[0003] Traditional equipment operating environment early warning methods mostly employ fixed threshold judgment mechanisms, that is, preset safety thresholds for environmental parameters such as temperature, humidity, and gas concentration, triggering an early warning when the monitored value exceeds the threshold. However, this method has significant limitations: on the one hand, marine environmental parameters are highly dynamic and coupled; for example, increased salt spray concentration can exacerbate the damage of humidity to equipment insulation, making fixed thresholds difficult to adapt to complex and changing environmental conditions. On the other hand, traditional methods rely on monitoring data from only one or a few environmental parameters, failing to effectively integrate multi-source heterogeneous data such as equipment vibration and the concentration of insulation gas decomposition products, resulting in low early warning accuracy and high false alarm and missed alarm rates.

[0004] Existing early warning systems generally lack dynamic modeling and proactive adjustment capabilities: they cannot build correlation models between equipment status and environmental parameters based on historical operating data and failure cases, making it difficult to predict failure risks in advance. Consequently, they cannot determine whether there is a deviation trend and intervene when a certain environmental parameter deviates from the optimization target value but has not reached the preset early warning threshold, resulting in significant early warning lag and easily triggering unplanned shutdowns or even safety accidents. In addition, the geographical environment, equipment layout, and operating conditions of different offshore substations vary, and existing models need to be retrained by collecting a large amount of historical data for each site, resulting in poor model adaptability and long deployment cycles. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide an intelligent early warning method and system for the operating environment of electrical equipment based on MLP, applicable to offshore substations. Through multi-source data fusion, improved MLP modeling, and active adjustment mechanisms, it achieves accurate monitoring of environmental parameters, early warning of equipment failures, and dynamic optimization of the operating environment. Simultaneously, it improves the model's adaptability to different substation scenarios and reduces data acquisition and model training costs.

[0006] To achieve the above objectives, according to a first aspect of the present invention, the present invention adopts the following technical solution:

[0007] A method for intelligent early warning of the operating environment of electrical equipment based on MLP, applicable to offshore substations, includes:

[0008] Collect data from the target site to construct the first dataset, and filter publicly available datasets to construct the second dataset;

[0009] The first and second datasets are preprocessed and subjected to feature extraction, dimensionality reduction, and fusion.

[0010] A multilayer perceptron regression and classification model was constructed, and the initial model was trained based on the second dataset.

[0011] A transfer learning method with a continuous label smoothing strategy is adopted, and the initial model is fine-tuned using the first dataset to adapt to the target site scenario;

[0012] The edge computing node receives environmental parameters and equipment status data collected by sensors in real time, processes them for features, and inputs them into the fine-tuned model to output the equipment failure probability and the optimized target value of environmental parameters.

[0013] Dynamic adjustment and graded early warning response are implemented based on the optimization target values ​​of equipment failure probability and environmental parameters.

[0014] Optionally, the process of collecting data from target sites to construct the first dataset includes:

[0015] Data is collected through a sensor network deployed at offshore substations. The sensor types include temperature and humidity sensors, vibration sensors, gas detectors, and infrared thermal imagers. At the same time, the operating status data of environmental control equipment is also collected.

[0016] The collected raw data were preprocessed, and time series data were extracted using the sliding window technique. Z-score normalization was used to eliminate the influence of dimensions, and wavelet transform was used to remove noise interference from vibration signals and gas concentration signals. Finally, the first dataset was formed, in which the features include preprocessed environmental parameters, equipment status parameters, and regulating equipment operating parameters, and the labels are equipment fault status and environmental parameter optimization target values.

[0017] Optionally, constructing a second dataset by filtering the public dataset includes:

[0018] We selected publicly available datasets related to the operating environment of electrical equipment at offshore substations and filtered out data that matched the sensor bands and measurement dimensions of the target sites.

[0019] The filtered public data is preprocessed using the same normalization and denoising methods as the first dataset. At the same time, the data is labeled according to the equipment failure records and the optimal operating range of environmental parameters in the public dataset to form the second dataset.

[0020] Optionally, the preprocessing and feature extraction of the first and second datasets includes: extracting statistical features, frequency domain features, time series features, and linkage features of regulating devices from the preprocessed first and second datasets.

[0021] Optionally, the dimensionality reduction and fusion process includes: using principal component analysis to reduce the dimensionality of the extracted high-dimensional features, retaining principal components with a cumulative variance contribution rate exceeding n%, and eliminating redundancy and multicollinearity among features.

[0022] Optionally, constructing the multilayer perceptron regression and classification model includes:

[0023] Construct an MLP model that combines classification and regression functions. The model structure includes an input layer, a hidden layer, and an output layer.

[0024] The number of nodes in the input layer is equal to the dimension of the feature vector after dimensionality reduction.

[0025] The hidden layers are set to p layers. The number of neurons in each layer is determined by Bayesian optimization algorithm. The activation function is ReLU.

[0026] The output layer adopts a dual-output structure. The first is the device failure probability, which is output through the Sigmoid activation function and used for classification tasks. The second is the environmental parameter optimization target value, which is output through the linear activation function and used for regression tasks.

[0027] Optionally, the initial model training based on the second dataset includes:

[0028] The MLP model was trained using the second dataset as training data. The model performance was evaluated using 5-fold cross-validation. The prediction accuracy of the environmental parameter optimization target value was evaluated using root mean square error. The equipment fault classification performance was evaluated using accuracy and recall.

[0029] The Bayesian optimization algorithm is used to search for the optimal model parameters. The optimization objective is to minimize the RMSE. The parameters to be optimized include the number of hidden layer neurons, learning rate, batch size, and regularization coefficient.

[0030] During training, L2 regularization and Dropout techniques are introduced to prevent overfitting. The Adam optimizer is used to minimize the loss function until the model converges, thus obtaining the initial model for intelligent early warning of electrical equipment operating environment.

[0031] Optionally, the transfer learning method employing a continuous label smoothing strategy, which fine-tunes the initial model using a first dataset to adapt it to the target site scenario, includes:

[0032] Based on the distribution of equipment failure probability labels and environmental parameter optimization target value labels in the first dataset, label intervals are divided. The division criterion is that there are no more than m intervals with a continuous sample count of 0. The sample count in each interval is then counted.

[0033] The interval weight is calculated based on the number of samples in each interval and the weighted sum of the number of samples in its neighboring intervals, as shown in the following formula: ;

[0034] in, Let be the weight of the i-th interval. Let be the number of samples in the i-th interval. When i exceeds the upper and lower limits, =0;

[0035] Freeze the parameters of the initial model's lower-level networks and train only the higher-level networks. During training, adjust the loss function using interval weights, as shown in the following formula:

[0036] ;

[0037] in, , These are the predicted and actual values ​​of the failure probability, respectively. , These represent the predicted and actual values ​​of the environmental parameter optimization target values, respectively; α and β are weighting coefficients; and Weight is the weight of the interval where the actual label is located.

[0038] The training continues until the loss value no longer decreases, resulting in an intelligent early warning model for the operating environment of electrical equipment adapted to the target site.

[0039] Optionally, the dynamic adjustment and graded early warning response based on the optimization target value of equipment failure probability and environmental parameters includes:

[0040] When environmental parameters deviate from the optimization target value but do not reach the warning threshold, the PID algorithm generates adjustment commands to control the operating power or speed of the ventilation and air conditioning equipment. During the adjustment process, environmental parameter changes are collected in real time to form a closed loop.

[0041] Three levels of early warning thresholds are set. When the probability of a fault exceeds the corresponding threshold, the system will notify the maintenance personnel through audible and visual alarms and SMS push notifications. At the same time, the system will coordinate with inspections to focus on the abnormal areas. If the early warning is triggered by abnormal environmental parameters, the adjustment actions will be strengthened first.

[0042] Optionally, the ventilation and air conditioning equipment may be any one, some or all of the following: air conditioner, fresh air dehumidifier, exhaust fan, electric heater and salt spray filter.

[0043] According to a second aspect of the present invention, the present invention adopts the following technical solution:

[0044] An intelligent early warning system for the operating environment of electrical equipment based on MLP, applicable to offshore substations, comprising:

[0045] Data collection module: Used to collect data from target sites to build the first dataset, and to filter publicly available datasets to build the second dataset;

[0046] Feature processing module: used to preprocess the first and second datasets and perform feature extraction, dimensionality reduction and fusion processing;

[0047] Initial model building module: used to build multilayer perceptron regression and classification models, and to train the initial model based on the second dataset;

[0048] Transfer learning module: Used to fine-tune the initial model using the first dataset to adapt it to the target site scenario, employing a transfer learning method with a continuous label smoothing strategy.

[0049] Real-time prediction module: It is used to receive environmental parameters and equipment status data collected by sensors in real time through edge computing nodes, and after feature processing, input them into the fine-tuned model to output the equipment failure probability and the optimized target value of environmental parameters.

[0050] Feedback adjustment module: used for dynamic adjustment and graded early warning response based on the target value of equipment failure probability and environmental parameters.

[0051] As can be seen from the above technical solution, compared with the prior art, the present invention provides an intelligent early warning method and system for the operating environment of electrical equipment based on MLP applicable to offshore substations, which has the following beneficial effects:

[0052] This invention proposes an intelligent early warning method for the operating environment of electrical equipment based on MLP applicable to offshore substations, including: collecting target site data to construct a first dataset, and screening public datasets to construct a second dataset; preprocessing the first and second datasets and performing feature extraction, dimensionality reduction, and fusion processing; constructing a multilayer perceptron regression and classification model, and training the initial model based on the second dataset; using a transfer learning method with a continuous label smoothing strategy, fine-tuning the initial model using the first dataset to adapt to the target site scenario; receiving environmental parameters and equipment status data collected by sensors in real time through edge computing nodes, inputting them into the fine-tuned model after feature processing, and outputting the equipment failure probability and environmental parameter optimization target value; and dynamically adjusting and hierarchically responding to the early warning based on the equipment failure probability and environmental parameter optimization target value. This invention effectively solves the problem of high dependence of traditional MLP models on historical data of target sites by constructing the first and second datasets (1) and combining transfer learning and continuous label smoothing strategies. Only a small amount of target site data needs to be collected to complete the model adaptation, reducing data collection costs and model deployment cycle, while improving the generalization ability of the model in different offshore substation scenarios. (2) An MLP model with both classification and regression functions is adopted, which can simultaneously output the probability of equipment failure and the target value of environmental parameter optimization, realizing the integration of early warning and regulation, transforming the traditional passive early warning into active prevention, effectively reducing the risk of equipment failure and reducing unplanned downtime. The Bayesian optimization algorithm is introduced to optimize the parameters of the MLP model, and combined with L2 regularization and Dropout technology, the prediction accuracy and anti-overfitting ability of the model are improved. Through the PID closed-loop control of the feedback regulation module, the environmental parameters are precisely regulated, further improving the equipment operating environment. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0054] Figure 1 This invention provides a schematic flowchart of an intelligent early warning method for the operating environment of electrical equipment based on MLP, applicable to offshore substations.

[0055] Figure 2 The present invention provides a structural framework diagram of an intelligent early warning system for the operating environment of electrical equipment based on MLP, applicable to offshore substations.

[0056] Figure 3This invention provides an extended structural diagram of an MLP-based intelligent early warning system for the operating environment of electrical equipment in offshore substations.

[0057] Figure 4 This invention provides a control principle diagram for an intelligent early warning system for the operating environment of electrical equipment based on MLP, applicable to offshore substations.

[0058] Among them, 1-control terminal, 2-1-electrical equipment room 1, 2-2-electrical equipment room 2, 11-network switch, 12-monitoring workstation, 13-central control room, 21-PLC equipment, 22-display panel, 23-acquisition equipment. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] This invention discloses an intelligent early warning method for the operating environment of electrical equipment based on MLP, applicable to offshore substations, such as... Figure 1 As shown, it includes:

[0061] Collect data from the target site to construct the first dataset, and filter publicly available datasets to construct the second dataset;

[0062] The first and second datasets are preprocessed and subjected to feature extraction, dimensionality reduction, and fusion.

[0063] A multilayer perceptron regression and classification model was constructed, and the initial model was trained based on the second dataset.

[0064] A transfer learning method with a continuous label smoothing strategy is adopted, and the initial model is fine-tuned using the first dataset to adapt to the target site scenario;

[0065] The edge computing node receives environmental parameters and equipment status data collected by sensors in real time, processes them for features, and inputs them into the fine-tuned model to output the equipment failure probability and the optimized target value of environmental parameters.

[0066] Dynamic adjustment and graded early warning response are implemented based on the optimization target values ​​of equipment failure probability and environmental parameters.

[0067] In a specific implementation, an intelligent early warning system for the operating environment of electrical equipment based on MLP, applicable to offshore substations, can apply the above method. The control principle is as follows: Figure 4 As shown, it specifically includes:

[0068] The system comprises a control terminal 1 and a controlled terminal. The control terminal 1 is connected to the controlled terminal via a Category 6 network cable. The control terminal 1 includes a central control room 13, a monitoring workstation 12, and a network switch 11. The monitoring workstation 12 and the network switch 11 are both connected to the central control room 13 via signal transmission. The network switch 11 is used for information exchange, the monitoring workstation 12 is used for visual display, and the central control room 13 is used to output control signals based on the collected information.

[0069] The monitoring workstation 12 also includes a display panel.

[0070] The controlled terminals include multiple controlled terminals such as electrical equipment room 1 and electrical equipment room 2.

[0071] For example, the electrical equipment room 1 includes: a data acquisition device 23, a display panel 22, and a PLC device 21. The data acquisition device 23 and the display panel 22 are both connected to the PLC device 21 by signals. The data acquisition device 23 is used to acquire real-time information, the display panel 22 is used for visual display, and the PLC device 21 is used for signal acquisition and transmission.

[0072] The collection of target site data to construct the first dataset includes:

[0073] Data is collected from multiple controlled terminals, including Electrical Equipment Room 1 and Electrical Equipment Room 2, through a sensor network deployed at the offshore substation. The sensor types of the data collection device 23 include: temperature and humidity sensors: monitoring the ambient temperature and relative humidity around the equipment; vibration sensors: collecting vibration acceleration signals of key equipment such as transformers and switchgear; differential pressure sensors: collecting differential pressure changes in the environment around the equipment; gas detectors: measuring the concentration of SF6 insulating gas and combustible gas (hydrogen); and infrared thermal imagers: acquiring the surface temperature distribution of the equipment, while simultaneously collecting the operating status data of environmental control equipment (fresh air unit, dehumidifier, air conditioner, exhaust fan): operating power, switch status, and speed.

[0074] PLC device 21 is connected to data acquisition device 23, and the acquired data is displayed in real time via display panel 22. PLC device 21 is used for signal acquisition and transmits the data to control terminal 1 via Category 6 network cable. In control terminal 1, network switch 11 is used for information exchange. Central control room 13 is used to output control signals based on the acquired information. The specific steps are as follows:

[0075] Step 1: Design a data verification mechanism based on the principle of constant weight drying: For temperature and humidity sensor data, periodically collect condensation samples from the equipment surface and measure their dry and wet weights to calculate the actual humidity value to calibrate the sensor data; for gas detector data, verify the measurement accuracy through a standard gas concentration calibration experiment.

[0076] The collected raw data were preprocessed, and time series data were extracted using sliding window technology. Z-score normalization was used to eliminate the influence of dimensions, and wavelet transform was used to remove noise interference from vibration signals and gas concentration signals. Finally, the first dataset was formed. The features included preprocessed environmental parameters, equipment status parameters, and adjustment equipment operating parameters. The labels were the equipment fault status (0 indicates normal, 1 indicates fault) and the environmental parameter optimization target value.

[0077] Step 2: Filter publicly available datasets to construct a second dataset:

[0078] Public datasets related to the operating environment of electrical equipment at offshore substations are selected, such as the OWC-Offshore Wind Farm Environment Dataset and the IEEE PES-Offshore Substation Equipment Reliability Dataset. Data with the same sensor band and measurement dimension as the target site are selected. For example, if the temperature and humidity sensor at the target site measures -20℃ to 60℃ and 0 to 100%RH, then temperature and humidity data with the same measurement range in the public dataset are selected. If the vibration sensor sampling frequency is 500Hz, then vibration data with the same sampling frequency in the public dataset are matched.

[0079] The filtered public data is preprocessed using the same normalization and denoising methods as the first dataset. At the same time, the data is labeled with the equipment fault records and optimal operating range of environmental parameters in the public dataset, i.e., the fault status and the optimization target value, to form the second dataset.

[0080] Step 3: Preprocess the first and second datasets and extract features:

[0081] Statistical features, frequency domain features, time series features, and control device linkage features are extracted from the preprocessed first and second datasets.

[0082] The statistical characteristics are: calculating the mean, variance, maximum, and minimum values ​​of environmental parameters and equipment status parameters within the sliding window (such as the average temperature and humidity variance within a 10-minute window).

[0083] Frequency domain features: Fast Fourier Transform (FFT) is performed on the vibration signal to extract the amplitude and phase information of the characteristic frequency bands; wavelet packet decomposition is performed on the gas concentration time series to extract the energy entropy of different frequency bands; Time series features: Autocorrelation coefficients and cross-correlation coefficients (such as the cross-correlation coefficients of temperature and humidity, and the autocorrelation coefficient of vibration signals) of environmental parameters and equipment state parameters are calculated; Regulation equipment linkage features: The lag time between the start-up and shutdown of the fresh air unit and temperature changes, and the correlation between the power of the dehumidifier and the rate of humidity decrease are analyzed to construct a linkage feature vector.

[0084] Step 4: Dimensionality reduction and fusion processing includes:

[0085] Principal component analysis was used to reduce the dimensionality of the extracted high-dimensional features, retaining principal components with a cumulative variance contribution rate of over 95%, eliminating redundancy and multicollinearity among features, and forming the final feature vector for subsequent MLP model training.

[0086] Step 5: Constructing a multilayer perceptron regression and classification model includes:

[0087] Construct an MLP model that combines classification and regression functions. The model structure includes an input layer, a hidden layer, and an output layer.

[0088] The number of nodes in the input layer is equal to the dimension of the feature vector after dimensionality reduction.

[0089] Two to three hidden layers are set, and the number of neurons in each layer is determined by Bayesian optimization algorithm (search range is 64-256). The ReLU function is used as the activation function to enhance the model's ability to fit nonlinear relationships.

[0090] The output layer adopts a dual-output structure. The first is the device failure probability, which is output through the Sigmoid activation function and used for classification tasks. The second is the environmental parameter optimization target value, which is output through the linear activation function and used for regression tasks.

[0091] Step Six: Perform initial model training based on the second dataset:

[0092] The MLP model was trained using the second dataset as training data. The model performance was evaluated using 5-fold cross-validation. The prediction accuracy of the environmental parameter optimization target value was evaluated using root mean square error (RMSE). The equipment fault classification performance was evaluated using accuracy and recall.

[0093] The Bayesian optimization algorithm is used to search for the optimal model parameters. The optimization objective is to minimize the RMSE. The parameters to be optimized include the number of hidden layer neurons, the learning rate (0.001-0.01), the batch size (32-128), and the regularization coefficient (0.0001-0.001).

[0094] During training, L2 regularization and Dropout techniques (with the Dropout rate set to 0.2-0.3) are introduced to prevent overfitting. The Adam optimizer is used to minimize the loss function, which is a weighted sum of classification loss and regression loss. The classification loss uses cross-entropy loss, and the regression loss uses mean squared error loss, until the model converges, thus obtaining the initial model for intelligent early warning of electrical equipment operating environment.

[0095] Step 7: Employ a transfer learning method with a continuous label smoothing strategy, fine-tune the initial model using the first dataset to adapt it to the target site scenario:

[0096] Based on the distribution of equipment failure probability labels (0-1) and environmental parameter optimization target value labels (such as temperature 15-30℃, humidity 30-60%RH) in the first dataset, label intervals are divided. The division criterion is that there are no more than 5 intervals with a continuous sample count of 0, and the sample count in each interval is counted.

[0097] The interval weight is calculated based on the number of samples in each interval and the weighted sum of the number of samples in its neighboring intervals, as shown in the following formula: ;

[0098] in, Let be the weight of the i-th interval. Let be the number of samples in the i-th interval. When i exceeds the upper and lower limits, =0;

[0099] Freeze the initial model's bottom-level network parameters (input layer and first hidden layer), and train only the high-level network (second hidden layer and output layer). During training, adjust the loss function using interval weights, as shown in the following formula:

[0100] ;

[0101] in, , These are the predicted and actual values ​​of the failure probability, respectively. , These represent the predicted and actual values ​​of the environmental parameter optimization target values, respectively. α and β are weighting coefficients, with α=0.6 and β=0.4. Weight is the weight of the interval containing the actual label.

[0102] The training continues until the loss value no longer decreases, resulting in an intelligent early warning model for the operating environment of electrical equipment adapted to the target site.

[0103] Step 8: Dynamically adjust and implement tiered early warning response based on optimized target values ​​for equipment failure probability and environmental parameters.

[0104] (1) The environmental parameters and equipment status data collected by the sensors are received in real time through the edge computing nodes. Specifically, the data is collected through the sensor network deployed at the offshore booster station. The sensor types include: temperature and humidity sensors, vibration sensors, differential pressure sensors, gas detectors, and infrared thermal imagers. At the same time, the operating status data of the environmental control equipment is collected. After feature processing, the data is input into the intelligent early warning model. The model outputs the probability of equipment failure and the target value of environmental parameters optimization within the next hour.

[0105] (2) When environmental parameters deviate from the optimized target value but do not reach the warning threshold, an adjustment command is generated through the PID algorithm to control the operating power or speed of the ventilation and air conditioning equipment. For example, when the humidity is 5%-10% higher than the optimized target value, the operating power of the dehumidifier is increased; when the SF6 concentration is close to the threshold, the speed of the exhaust fan is increased. During the adjustment process, changes in environmental parameters are collected in real time to form an overall monitoring-adjustment-feedback closed loop. Specifically, the ventilation and air conditioning equipment includes an air conditioner, a fresh air dehumidifier, an exhaust fan, an electric heater, and a salt spray filter. The specific steps are as follows:

[0106] 1) Environmental parameter deviation judgment: Real-time comparison between the current environmental parameters collected by the sensors and the optimized target value output by the model to determine whether there is a deviation: When an environmental parameter deviates from the optimized target value but does not reach the preset warning threshold (e.g., the humidity optimized target value is 50%, the current humidity is 56%, which does not reach the humidity threshold of 60% corresponding to the yellow warning), the dynamic adjustment process is started; if the parameter is within the optimized target value range, the current environmental adjustment equipment operation status is maintained.

[0107] 2) Generation and execution of PID algorithm adjustment instructions: Environmental parameter deviation data (such as humidity deviation of 6% and SF6 concentration deviation of 80μL / L) are transmitted to the PID controller of the edge computing node. The PID controller automatically tunes the proportional (P), integral (I), and derivative (D) parameters according to the deviation magnitude, deviation change rate, and system response characteristics (such as ventilation duct air volume and equipment adjustment lag time).

[0108] Based on the tuned PID parameters, targeted adjustment commands are generated, such as:

[0109] When the humidity is 5%-10% higher than the optimized target value (e.g., the optimized target value is 50%, and the current humidity is 53%-55%), a command to "increase the operating power of the fresh air dehumidifier" will be generated. The power increase will be calculated based on the humidity deviation (e.g., for every 1% deviation, the power will increase by 5%-8%), and will not exceed 80% of the rated power. If the humidity is more than 10% higher than the optimized target value (e.g., 55%-60%), the exhaust fan speed will be increased at the same time (for every 5% deviation, the speed will increase by 10%).

[0110] When the SF6 concentration approaches the warning threshold (e.g., the warning threshold is 1000 μL / L, and the current concentration is 900-950 μL / L), a "increase exhaust fan speed" command is generated, and the speed is gradually increased from the current operating value to 70%-90% of the rated speed, while the standby exhaust fan is turned on.

[0111] When the ambient temperature is lower than the optimized target value (e.g., the optimized target value is 25℃ and the current temperature is 20℃), a command to "start the electric heater and adjust the power" is generated. The initial power is set to 50% of the rated power and is gradually adjusted according to the temperature rise rate. If the temperature is higher than the optimized target value (e.g., 28℃-32℃), the cooling power of the air conditioner is reduced or the cooling fan speed is increased.

[0112] When the pressure difference in the compartment deviates from the optimized target value (e.g., the equipment compartment is required to have a positive pressure of 50Pa relative to the outside, but it is currently 30Pa), adjust the speed ratio of the supply fan and the exhaust fan to increase the supply air volume or decrease the exhaust air volume until the pressure difference reaches the target value.

[0113] The adjustment command is transmitted to the controller of each environmental control device through the industrial bus. The device controller executes the command and provides real-time feedback on the execution status (such as the dehumidifier power has been increased to 60% of the rated power and the exhaust fan speed has been adjusted to 800 r / min).

[0114] 3) Monitoring and feedback of regulation effects

[0115] During the execution of the adjustment command, the sensor continuously collects environmental parameter change data at the original collection frequency (such as collecting humidity data once every 30 seconds and SF6 concentration data once every 15 seconds) and transmits it back in real time.

[0116] Compare the environmental parameters before and after adjustment and analyze the adjustment effect: if the parameters gradually approach the optimized target value (e.g., humidity drops from 56% to 52%), then maintain the current adjustment command; if the parameters do not change significantly or the deviation increases (e.g., humidity remains at 56%), then the deviation data is transmitted to the PID controller again, the parameters are readjusted and a new adjustment command is generated until the environmental parameters stabilize within the optimized target value range, forming a complete "monitoring-adjustment-feedback" closed loop.

[0117] (3) Set three-level early warning thresholds: yellow: 30%-50% probability of failure; orange: 50%-80% probability of failure; red: 80% or more probability of failure. When the probability of failure exceeds the corresponding threshold, the system will notify the operation and maintenance personnel through sound and light alarms and SMS push notifications. At the same time, the system will link the inspection to focus on the abnormal area. If the early warning is triggered by abnormal environmental parameters, the system will prioritize strengthening the adjustment action, such as increasing the power of the fresh air unit to 80%-100% of the rated power to reduce the risk of failure.

[0118] (4) The monitoring workstation 12 is used to visualize the results of the graded early warning response.

[0119] In a specific implementation, step nine may be further included: performing dynamic model updates.

[0120] New data from target sites are collected regularly, including environmental parameters, equipment status data, adjustment actions and fault records. This data is added to the first dataset and preprocessed. An incremental training method is used to fine-tune the intelligent early warning model using the updated first dataset, optimizing model parameters and early warning thresholds. At the same time, the public dataset is re-selected every 6 months to update the second dataset and retrain the initial model, further improving the model's generalization ability and prediction accuracy.

[0121] The dynamic updating of the model specifically includes:

[0122] I. New Data Collection and Preprocessing at the Target Site, Update of the First Dataset

[0123] 1.1 New Data Collection Scope and Indicator Definitions

[0124] Regularly (ideally monthly) collect new data from the target offshore substation, ensuring the collected metrics remain consistent with the initial dataset. This includes:

[0125] Environmental parameters: Temperature T new (t) (unit: °C), relative humidity H new (t) (unit: %RH), SF6 concentration C SF6,new (t) (unit: ppm), Combustible gas concentration C comb,new (t) (unit: %LEL), sampling frequency is 1 time / minute;

[0126] Equipment status data: Transformer vibration acceleration A new (t) (unit: m / s²), switchgear surface temperature T cab,new (t) (unit: °C), sampling frequency is 1 time / 10 minutes;

[0127] Adjustment action record: Fresh air unit power P fan,new (t) (unit: kW), dehumidifier operating speed G deh,new (t) (Discrete values: 1-5), Exhaust fan speed R exh,new (t) (unit: r / min), recording frequency synchronized with equipment action triggering;

[0128] Fault Log: Equipment Fault Type F new (Discrete values: 0 = normal, 1 = insulation fault, 2 = mechanical fault, 3 = corrosion fault), fault occurrence time t f .

[0129] 1.2 New Data Preprocessing Formula Steps

[0130] Step 1: Data Cleaning and Outlier Removal

[0131] The 3σ criterion is used to remove outliers. For continuous data (such as temperature, humidity, and concentration), the mean μ and standard deviation σ of the historical data (last 3 months) for this indicator are calculated. If the new data x new satisfy If the value is not found, it is considered an outlier and replaced using linear interpolation, as shown in the following formula:

[0132] ;

[0133] Where, x new,corr For the revised new data, x new+1 These are the normal data adjacent to the new data, t new t new+1 These represent the data collection times.

[0134] Step 2: Data normalization processing

[0135] The Z-score normalization method is used to convert the cleaned new data into dimensionless data, as shown in the following formula:

[0136] ;

[0137] Where, x new,norm For the new data after normalization, μ hist σ hist These are the mean and standard deviation of the initial dataset for this indicator, respectively, to ensure that the distribution of the new data is consistent with that of the historical data.

[0138] Step 3: Feature Reconstruction and Merging with the First Dataset

[0139] Using the initial feature engineering method, statistical features, frequency domain features, and time series features are extracted from the preprocessed new data:

[0140] Statistical characteristics: Calculate the mean μ of new data within a 10-minute sliding window. new,win ,variance The formula is as follows:

[0141] ;

[0142] Where N is the number of data points in the sliding window (N=10, corresponding to 10 minutes of sampling data);

[0143] Frequency domain characteristics: For new vibration data A new,norm (t) Perform Fast Fourier Transform (FFT) to extract the amplitude A corresponding to the feature frequency f0 (e.g., 50Hz, 100Hz). new,fft (f0), the formula is as follows:

[0144] ;

[0145] Where k is the frequency point index, j is the imaginary unit, and the amplitude of the characteristic frequency f0 corresponding to k is taken as the frequency domain feature;

[0146] Time-series characteristics: Calculate the cross-correlation coefficient ρ between new environmental parameters and adjustment actions. new Taking temperature and fresh air unit power as an example, the formula is as follows:

[0147] ;

[0148] Where M is the data length, τ is the lag time (taken as 5 minutes), and μ T,new μ P,new These are the average values ​​of the new temperature data and the new air blower power data, respectively.

[0149] The reconstructed new feature vectors are added to the initial first dataset to form the updated first dataset. , where X new,feat For the new feature vector, Y new,lab New data labels (fault probability labels y) cls,new Environmental optimization target value label y reg,new ).

[0150] II. Incremental training of the model based on the updated first dataset

[0151] 2.1 Incremental Training Parameter Initialization

[0152] Retain the original intelligent early warning model's network structure (input layer - hidden layer - output layer), and freeze the parameters θ of the bottom layer network (input layer and first hidden layer). bottom Only the trainable parameters of the high-level network (the second hidden layer and the output layer) are initialized. The initial value is set to 0.9 times the corresponding layer parameter of the original model, as shown in the following formula:

[0153] ;

[0154] in, These are the parameters of the high-level network in the original model.

[0155] 2.2 Definition of Incremental Training Loss Function

[0156] Following the continuous label smoothing strategy, based on the updated first dataset The label distribution is redefined, label intervals are redefined, and weights are calculated. The loss function still uses a weighted sum of classification loss and regression loss, as shown in the following formula:

[0157] ;

[0158] Where: α=0.6 and β=0.4 are weighting coefficients; , These are the predicted fault probability and the predicted environmental optimization target value output by the incrementally trained model, respectively. , These are the true labels of the fault probability and the true labels of the environmental optimization target value in the first dataset after the update. The weights for the updated label intervals are calculated in the same way as in the initial transfer learning:

[0159] ;

[0160] in, This represents the number of samples in the i-th label interval of the updated first dataset. When i exceeds the range... =0.

[0161] 2.3 Incremental Training Optimization Process

[0162] Minimize incremental training loss using the Adam optimizer incr The learning rate is set to 0.5 times the initial training learning rate (i.e., lr). incr =0.5×lr init Iteratively update the high-level network parameters θ top The formula is as follows:

[0163] Calculate the gradient:

[0164] Update momentum and second-order momentum:

[0165] ;

[0166] ;

[0167] Where β1=0.9 and β2=0.999 are momentum coefficients, m t v t These are the first-order momentum and the second-order momentum, respectively, and t is the number of iterations.

[0168] Parameter update:

[0169] ;

[0170] in, To prevent the minimum value where the denominator is 0.

[0171] Iterative training until the loss function is reached. incr No decrease for 5 consecutive epochs (decline <10) -4 Training is stopped, and the incrementally updated intelligent early warning model is obtained.incr .

[0172] 2.4 Early Warning Threshold Optimization

[0173] Based on the updated first dataset The Youden exponent maximization method was used to optimize the three-level early warning threshold (yellow Thr.). y Orange Th o Red Th r ):

[0174] Calculate different candidate thresholds Th cand The corresponding sensitivity S ens With specificity S pec :

[0175] ;

[0176] Among them, TP is a true positive (actual fault and model prediction fault), FN is a false negative (actual fault but model prediction normal), TN is a true negative (actual normal and model prediction normal), and FP is a false positive (actual normal but model prediction fault).

[0177] Calculate the Youden index The candidate threshold corresponding to the maximum value of J is selected as the optimized warning threshold:

[0178] .

[0179] III. Public Dataset Updates and Initial Model Retraining

[0180] 3.1 Filtering and updating the public dataset, updating the second dataset

[0181] Every 6 months, data is re-filtered from public databases (such as the OWC-Offshore Wind Farm Environmental Dataset and the IEEE PES Offshore Substation Dataset), using the same filtering criteria as the initial second dataset.

[0182] Sensor band matching: Temperature sensor measurement range -20℃-60℃, humidity sensor 0-100%RH, vibration sensor sampling frequency 500Hz;

[0183] Data time span: Prioritize publicly available data from the past year to ensure data timeliness;

[0184] Label completeness: Includes equipment fault records and labels indicating the optimal operating range of environmental parameters.

[0185] The newly selected public data is preprocessed using the same methods as the first dataset (cleaning, normalization, feature extraction), and then added to the initial second dataset to form the updated second dataset. ,in, For the feature vector of the newly released data, For the corresponding tag.

[0186] 3.2 Retraining the initial model based on the updated second dataset

[0187] Step 1: Model parameter initialization

[0188] The initial MLP model is reconstructed with the same number of input layer nodes and feature vector dimension. The number of hidden layers and neurons are re-searched using Bayesian optimization, with the optimization objective being to minimize the root mean square error (RMSE). The RMSE calculation formula is as follows:

[0189] ;

[0190] in, This represents the number of samples in the updated second dataset. To optimize the target value prediction for the environment in which the model is retrained, The target value is optimized for the true label in the updated second dataset.

[0191] Step 2: Retrain the loss function and optimize

[0192] The loss function and optimization strategy used in the initial training are as follows:

[0193] ;

[0194] in, Cross-entropy loss is used for fault classification tasks, and mean squared error loss is used for environmental target value regression tasks. α=0.6 and β=0.4 are weighting coefficients.

[0195] Using the Adam optimizer (lr retrain =0.005), L2 regularization (regularization coefficient λ=0.001) and Dropout (dropout rate=0.25) are introduced to prevent overfitting, and training is iterated until the loss is reached. retrain Convergence (decline <10 for 5 consecutive epochs) -4 ), to obtain the initial model after retraining .

[0196] Step 3: Verify the model's generalization ability

[0197] Five-fold cross-validation was used to evaluate the generalization ability of the initial model after retraining, and the fault prediction accuracy A on the validation set was calculated. cc RMSE of predicted environmental target values:

[0198] ;

[0199] ;

[0200] Where, N val To determine the number of samples in the validation set, A is required to... cc ≥90%, RMSE val ≤2 (RMSE ≤ 2℃, RMSE ≤ 5%RH). If not met, readjust the model parameters (such as the number of hidden layer neurons and the learning rate) and retrain until the requirements are met.

[0201] In summary, the present invention provides an intelligent early warning system for the operating environment of electrical equipment based on MLP, applicable to offshore substations, such as... Figure 2 As shown, it includes:

[0202] Data collection module: Used to collect data from target sites to build the first dataset, and to filter publicly available datasets to build the second dataset;

[0203] Feature processing module: used to preprocess the first and second datasets and perform feature extraction, dimensionality reduction and fusion processing;

[0204] Initial model building module: used to build multilayer perceptron regression and classification models, and to train the initial model based on the second dataset;

[0205] Transfer learning module: Used to fine-tune the initial model using the first dataset to adapt it to the target site scenario, employing a transfer learning method with a continuous label smoothing strategy.

[0206] Real-time prediction module: It is used to receive environmental parameters and equipment status data collected by sensors in real time through edge computing nodes, and after feature processing, input them into the fine-tuned model to output the equipment failure probability and the optimized target value of environmental parameters.

[0207] Feedback adjustment module: used for dynamic adjustment and graded early warning response based on the target value of equipment failure probability and environmental parameters.

[0208] like Figure 3 As shown, the structural extension specifically includes: an acquisition module, a feature processing module, an initial model construction module, a transfer learning module, a real-time prediction module, and a feedback adjustment module. It may further include: a storage module, a system reliability assurance module, and an early warning and decision support module. The functions of each module are as follows:

[0209] The data acquisition module consists of a sensor network and data acquisition terminals deployed around the substation cabin and key equipment. The sensor types include temperature and humidity sensors, vibration sensors, gas detectors, and infrared thermal imagers. The data acquisition terminals support communication protocols such as RS485, LoRa, and 5G. They collect environmental parameters, equipment status data, and equipment operation data in real time and transmit the data to the feature processing module and storage module.

[0210] Feature processing module: The original data is denoised and normalized using methods such as sliding window technology, Z-score normalization, and wavelet transform. Multidimensional features are extracted and then processed by PCA dimensionality reduction.

[0211] Storage module: It adopts a combination of local storage on edge computing nodes and distributed storage in the cloud. Local storage stores raw data collected in real time and preprocessed data (storage period of 3 months), while the cloud stores historical data, the first dataset, the second dataset and model training parameters (long-term storage), supporting fast data query and retrieval;

[0212] Initial model building module: used to build multilayer perceptron regression and classification models; the model training unit is deployed in the cloud.

[0213] Transfer learning module: The model supports initial model training based on the second dataset and fine-tuning through transfer learning based on the first dataset. It integrates Bayesian optimization algorithm for model parameter search and introduces L2 regularization and Dropout techniques to prevent overfitting.

[0214] Real-time prediction module: Loads the trained intelligent early warning model, receives the feature vectors output by the feature processing module and the storage module in real time, completes the prediction of equipment failure probability and environmental parameter optimization target value, and transmits the results to the feedback adjustment module and the early warning and decision support module;

[0215] Feedback control module: Composed of control equipment controller and PID control unit: The control equipment controller is connected to environmental control equipment such as fresh air unit, dehumidifier, air conditioner, and exhaust fan, and receives the environmental parameter optimization target value output by the real-time prediction module; The PID control unit calculates the adjustment amount and generates control commands based on the deviation between the current environmental parameters and the optimization target value, controls the operating status of the control equipment, and collects the adjusted environmental parameters in real time and feeds them back to the acquisition module to form closed-loop control;

[0216] Early warning and decision support module: This module includes an early warning unit and a decision support unit. The early warning unit triggers corresponding audible and visual alarms based on the fault probability output by the real-time prediction module and the three-level early warning threshold, and notifies maintenance personnel via SMS, APP push, etc. The decision support unit generates equipment health reports, including fault probability trends, environmental parameter compliance rates, and environmental regulation effect analysis reports. This provides data support for predictive maintenance, such as equipment maintenance cycle formulation and regulation strategy optimization, such as PID parameter adjustment. It also supports data interaction with the offshore booster station operation and maintenance management platform.

[0217] System reliability assurance module: includes dual redundant communication links, UPS power supply and offline emergency unit: The dual redundant communication links adopt fiber optic + 5G communication to ensure the stability of data transmission and control command issuance; the UPS power supply provides backup power for sensors, edge computing nodes and key adjustment equipment to ensure the normal operation of basic monitoring and adjustment functions in the event of a sudden power outage; the offline emergency unit has preset emergency adjustment strategies, such as automatically starting the dehumidifier when humidity exceeds the standard, starting the exhaust fan when SF6 concentration exceeds the standard, and automatically triggering the emergency strategy when the edge computing node is disconnected from the cloud, improving system robustness.

[0218] The present invention provides an intelligent early warning method for the operating environment of electrical equipment based on MLP, applicable to offshore substations. Specific application examples are as follows:

[0219] (1) At a certain offshore substation (target site), a sensor network is deployed in multiple controlled terminals such as electrical equipment room 1 and electrical equipment room 2. Environmental parameters (temperature -10℃~45℃, humidity 20%~90%RH), equipment vibration data (sampling frequency 500Hz), SF6 concentration data (0~1000ppm), and equipment operation data are collected for one month by the acquisition device 23 to form the first dataset (sample size 5000). The PLC device 21 is used to transmit the signals collected by the acquisition device 23 to the control terminal 1 through a Category 6 network cable. The control terminal 1 receives the transmitted signals through the network switch 11. The central control room 13 is used to output control signals according to the collected information. The monitoring workstation 12 is used to perform visual display according to the control signals.

[0220] Specifically, the central control room 13 selects data of the same dimensions from the offshore wind farm environmental dataset OWC to form a second dataset (sample size 20,000). Statistical features (such as mean temperature and humidity variance), frequency domain features (vibration signal FFT amplitude), and time series features (temperature autocorrelation coefficient) are extracted and dimensionality reduced by PCA to obtain a 20-dimensional feature vector.

[0221] MLP model training: The model parameters were determined using the Bayesian optimization algorithm (2 hidden layers, 128 and 64 neurons respectively, learning rate 0.005, batch size 64). The initial model was trained based on the second dataset, and then smooth transfer learning was performed using continuous labels (the label interval was divided into 10, and the loss was adjusted after weight calculation). The model was then fine-tuned using the first dataset.

[0222] Model validation: The test set uses an additional 1,000 data points collected from the target site. The model's fault prediction accuracy reached 92%, and the predicted RMSE of the optimized environmental parameters was less than 2 (temperature RMSE < 2℃, humidity RMSE < 5%RH). Compared with the traditional fixed threshold early warning method, the false alarm rate was reduced by 60%, and the early warning time was up to 1 hour.

[0223] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0224] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for intelligent early warning of the operating environment of electrical equipment based on MLP, applicable to offshore substations, characterized in that, include: Collect data from the target site to construct the first dataset, and filter publicly available datasets to construct the second dataset; The first and second datasets are preprocessed and subjected to feature extraction, dimensionality reduction, and fusion. A multilayer perceptron regression and classification model was constructed, and the initial model was trained based on the second dataset. A transfer learning method with a continuous label smoothing strategy is adopted, and the initial model is fine-tuned using the first dataset to adapt to the target site scenario; The edge computing node receives environmental parameters and equipment status data collected by sensors in real time, processes them for features, and inputs them into the fine-tuned model to output the equipment failure probability and the optimized target value of environmental parameters. Dynamic adjustment and graded early warning response are implemented based on the optimization target values ​​of equipment failure probability and environmental parameters.

2. The intelligent early warning method for the operating environment of electrical equipment based on MLP, applicable to offshore substations, as described in claim 1, is characterized in that... The collection of target site data to construct the first dataset includes: Data is collected through a sensor network deployed at offshore substations. The sensor types include temperature and humidity sensors, vibration sensors, differential pressure sensors, gas detectors, and infrared thermal imagers. At the same time, the operating status data of environmental control equipment is also collected. The collected raw data were preprocessed, and time series data were extracted using the sliding window technique. Z-score normalization was used to eliminate the influence of dimensions, and wavelet transform was used to remove noise interference from vibration signals and gas concentration signals. Finally, the first dataset was formed, in which the features include preprocessed environmental parameters, equipment status parameters, and regulating equipment operating parameters, and the labels are equipment fault status and environmental parameter optimization target values.

3. The intelligent early warning method for the operating environment of electrical equipment based on MLP, applicable to offshore substations, as described in claim 1, is characterized in that... The process of filtering the public dataset to construct the second dataset includes: We selected publicly available datasets related to the operating environment of electrical equipment at offshore substations and filtered out data that matched the sensor bands and measurement dimensions of the target sites. The filtered public data is preprocessed using the same normalization and denoising methods as the first dataset. At the same time, the data is labeled according to the equipment failure records and the optimal operating range of environmental parameters in the public dataset to form the second dataset.

4. The intelligent early warning method for the operating environment of electrical equipment based on MLP, applicable to offshore substations, as described in claim 1, is characterized in that... The preprocessing and feature extraction of the first and second datasets includes extracting statistical features, frequency domain features, time series features, and linkage features of regulating devices from the preprocessed first and second datasets.

5. The intelligent early warning method for the operating environment of electrical equipment based on MLP, applicable to offshore substations, as described in claim 1, is characterized in that... The construction of the multilayer perceptron regression and classification model includes: Construct an MLP model that combines classification and regression functions. The model structure includes an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is equal to the dimension of the feature vector after dimensionality reduction. The hidden layers are set to p layers. The number of neurons in each layer is determined by Bayesian optimization algorithm. The activation function is ReLU. The output layer adopts a dual-output structure. The first is the device failure probability, which is output through the Sigmoid activation function and used for classification tasks. The second is the environmental parameter optimization target value, which is output through the linear activation function and used for regression tasks.

6. The intelligent early warning method for the operating environment of electrical equipment based on MLP, applicable to offshore substations, as described in claim 1, is characterized in that... The initial model training based on the second dataset includes: The MLP model was trained using the second dataset as training data. The model performance was evaluated using 5-fold cross-validation. The prediction accuracy of the environmental parameter optimization target value was evaluated using root mean square error. The equipment fault classification performance was evaluated using accuracy and recall. The Bayesian optimization algorithm is used to search for the optimal model parameters. The optimization objective is to minimize the RMSE. The parameters to be optimized include the number of hidden layer neurons, learning rate, batch size, and regularization coefficient. During training, L2 regularization and Dropout techniques are introduced to prevent overfitting. The Adam optimizer is used to minimize the loss function until the model converges, thus obtaining the initial model for intelligent early warning of electrical equipment operating environment.

7. The intelligent early warning method for the operating environment of electrical equipment based on MLP, applicable to offshore substations, as described in claim 1, is characterized in that... The transfer learning method employing a continuous label smoothing strategy fine-tunes the initial model using a first dataset to adapt it to the target site scenario, including: Based on the distribution of equipment failure probability labels and environmental parameter optimization target value labels in the first dataset, label intervals are divided. The division criterion is that there are no more than m intervals with a continuous sample count of 0. The sample count in each interval is then counted. The interval weight is calculated based on the number of samples in each interval and the weighted sum of the number of samples in its neighboring intervals, as shown in the following formula: ; in, Let be the weight of the i-th interval. Let be the number of samples in the i-th interval. When i exceeds the upper and lower limits, =0; Freeze the parameters of the initial model's lower-level networks and train only the higher-level networks. During training, adjust the loss function using interval weights, as shown in the following formula: ； in, , These are the predicted and actual values ​​of the failure probability, respectively. , These represent the predicted and actual values ​​of the environmental parameter optimization target values, respectively; α and β are weighting coefficients; and Weight is the weight of the interval where the actual label is located. The training continues until the loss value no longer decreases, resulting in an intelligent early warning model for the operating environment of electrical equipment adapted to the target site.

8. The intelligent early warning method for the operating environment of electrical equipment based on MLP, applicable to offshore substations, as described in claim 1, is characterized in that... The dynamic adjustment and graded early warning response based on the optimized target value of equipment failure probability and environmental parameters includes: When environmental parameters deviate from the optimization target value but do not reach the warning threshold, the PID algorithm generates adjustment commands to control the operating power or speed of the ventilation and air conditioning equipment. During the adjustment process, environmental parameter changes are collected in real time to form a closed loop. Three levels of early warning thresholds are set. When the probability of a fault exceeds the corresponding threshold, the system will notify the maintenance personnel through audible and visual alarms and SMS push notifications. At the same time, the system will coordinate with inspections to focus on the abnormal areas. If the early warning is triggered by abnormal environmental parameters, the adjustment actions will be strengthened first.

9. A method for intelligent early warning of the operating environment of electrical equipment based on MLP, applicable to offshore substations, as described in claim 8, characterized in that, The ventilation and air conditioning equipment includes an air conditioner, a fresh air dehumidifier, an exhaust fan, an electric heater, and a salt spray filter.

10. An intelligent early warning system for the operating environment of electrical equipment based on MLP, applicable to offshore substations, characterized in that, include: Data collection module: Used to collect data from target sites to build the first dataset, and to filter publicly available datasets to build the second dataset; Feature processing module: used to preprocess the first and second datasets and perform feature extraction, dimensionality reduction and fusion processing; Initial model building module: used to build multilayer perceptron regression and classification models, and to train the initial model based on the second dataset; Transfer learning module: Used to fine-tune the initial model using the first dataset to adapt it to the target site scenario, employing a transfer learning method with a continuous label smoothing strategy. Real-time prediction module: It is used to receive environmental parameters and equipment status data collected by sensors in real time through edge computing nodes, and after feature processing, input them into the fine-tuned model to output the equipment failure probability and the optimized target value of environmental parameters. Feedback adjustment module: used for dynamic adjustment and graded early warning response based on the target value of equipment failure probability and environmental parameters.

Images