Offshore wind power operation and maintenance risk assessment method and device, storage medium and electronic equipment

Abstract

The present application relates to the technical field of offshore wind power safety, and discloses a kind of offshore wind power operation risk assessment method, device, storage medium and electronic equipment, comprising: collecting operation and maintenance multi-source information and dividing into static and dynamic information to construct dataset;Corresponding information vector is obtained by targeted preprocessing;Based on vector, static and dynamic risk characteristics are respectively constructed and personalized risk characteristics are generated by fusion;Accordingly, reinforcement learning state, action space and reward function are constructed, and strategy optimization model is trained;Risk assessment is carried out using the model, and dynamic updating is carried out according to actual results.The present application fuses personnel static and dynamic multi-source information to construct personalized characteristics, and realizes model self-adaptive updating by means of reinforcement learning, effectively solves the problem of low precision and recognition lag caused by ignoring individual differences and lacking dynamic optimization in existing methods, and significantly improves the accuracy and real-time performance of evaluation.

Description

Technical Field

[0001] This invention relates to the field of offshore wind power operation and maintenance safety technology, specifically to an offshore wind power operation and maintenance risk assessment method, device, storage medium, and electronic equipment. Background Technology

[0002] With the accelerated global energy transition, offshore wind power, as an important component of clean energy, has seen its installed capacity continuously expand. However, the operation and maintenance environment of offshore wind power is complex and harsh, facing multiple challenges such as high salt spray corrosion, strong wind and wave impact, and limited working space. In addition, maintenance personnel need to carry out equipment maintenance and troubleshooting in high-risk scenarios such as high altitudes and confined spaces, resulting in significantly higher safety risks for maintenance operations compared to onshore wind power. Existing offshore wind power operation and maintenance risk assessment methods mainly rely on static risk matrices or expert experience scoring, considering only objective factors such as equipment status and environmental parameters, while ignoring the dynamic impact of individual differences in maintenance personnel on risks. At the same time, traditional assessment models lack adaptive learning capabilities, cannot dynamically optimize assessment strategies based on actual operation and maintenance feedback, and are difficult to accurately identify hidden risks caused by personnel-environment coupling, easily leading to missed risks or false alarms, and failing to meet the needs of refined and intelligent safety management for offshore wind power operation and maintenance. Summary of the Invention

[0003] The purpose of this invention is to provide a method, device, storage medium, and electronic device for risk assessment of offshore wind power operation and maintenance. It can integrate static and dynamic multi-source information from operation and maintenance personnel to construct personalized risk characteristics, and achieve adaptive dynamic updating of the risk assessment model through reinforcement learning strategy optimization. It effectively solves the problems of low assessment accuracy and delayed risk identification caused by neglecting individual differences of personnel and lack of dynamic optimization capabilities in existing assessment methods, thereby improving the accuracy and real-time performance of risk assessment for offshore wind power operation and maintenance.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] In a first aspect, the present invention provides a method for risk assessment of offshore wind power operation and maintenance based on multi-source personnel information and reinforcement learning, which includes the following steps:

[0006] S1. Collect multi-source information of personnel during offshore wind power operation and maintenance, and divide the multi-source information into static information and dynamic information according to the information change characteristics to construct a multi-source information dataset of personnel; wherein, the static information includes basic personnel information, health characteristic information and historical operation information, and the dynamic information includes personnel vital signs information, behavioral status information and operation environment information;

[0007] S2, perform targeted preprocessing on the static information and the dynamic information respectively to obtain the processed static information vector and dynamic information vector;

[0008] S3, based on the processed static information vector and dynamic information vector, construct static information risk features and dynamic information risk features respectively, and merge them to generate personalized risk features for operation and maintenance based on multi-source personnel information;

[0009] S4. Based on the personalized risk characteristics of the operation and maintenance tasks, construct the state space, action space and reward function of reinforcement learning, and obtain the reinforcement learning policy optimization model through policy network training.

[0010] S5. Based on the reinforcement learning strategy, optimize the model to perform a risk assessment on the current operation and maintenance task, and dynamically update the model according to the actual operation and maintenance task results.

[0011] Further, step S2 specifically includes: encoding and converting static information in multi-source personnel information: converting static information according to different types of static information using corresponding encoding formats, wherein unordered static information uses one-hot encoding, ordered static information uses label encoding, and constructing a static information vector by concatenating feature vectors; performing time-series processing on dynamic information in multi-source personnel information: constructing a sliding time window for each dynamic information collected in real time, performing noise reduction processing on the data within the window using a moving average filtering method, and performing normalization processing to construct a dynamic information vector.

[0012] Further, step S3 specifically includes: constructing static information risk features based on the processed static information vector through linear scoring of risk weights; constructing dynamic information risk feature vectors based on the processed dynamic information vectors by calculating the mean, standard deviation, trend, and rate of change of dynamic information within the window, and constructing personnel-environment coupled risk features through a nonlinear mapping function based on the deviation between personnel vital sign information and its corresponding benchmark parameters and the deviation between environmental information and its corresponding safety threshold; and fusing the static information risk features, dynamic information risk feature vectors, and personnel-environment coupled risk features to construct personalized risk features for operation and maintenance based on multi-source personnel information.

[0013] Further, step S4 specifically includes: constructing a reinforcement learning state space based on personalized risk characteristics of operation and maintenance tasks: concatenating the personalized risk characteristics of the operation and maintenance tasks at the current moment with the historical risk characteristics within the historical information backtracking step size, as the current state of reinforcement learning; constructing a risk assessment action space: using the risk assessment value as the action output; constructing a reward function based on risk assessment feedback: the reward function is composed of a weighted combination of risk missed detection penalty terms and risk false alarm penalty terms; constructing a reinforcement learning policy optimization model: using the current state as input, generating the corresponding risk assessment value through the policy network, and optimizing the policy network by maximizing the cumulative reward objective function according to the reward function, wherein the cumulative reward objective function introduces a reward discount factor to weighted sum the rewards over multiple rounds.

[0014] Further, step S5 specifically includes: inputting the personalized risk characteristics of the current operation and maintenance task into the optimized strategy network, and outputting the risk assessment value of the current operation and maintenance task; mapping the continuous risk assessment value into discrete risk levels through a preset risk level mapping function based on the risk assessment value; calculating the reinforcement learning reward function value based on the actual operation and maintenance task process results, and iteratively updating the strategy network parameters to achieve adaptive dynamic optimization of the risk assessment model.

[0015] Secondly, the present invention provides a device for assessing the operation and maintenance risks of offshore wind power using multi-source personnel information and reinforcement learning, comprising:

[0016] The multi-source information acquisition module is used to collect multi-source information from offshore wind power operation and maintenance personnel during operation and maintenance work, and to divide the multi-source information into static information and dynamic information according to the information change characteristics, thereby constructing a multi-source information dataset for personnel; wherein, the static information includes basic personnel information, health characteristic information and historical operation information, and the dynamic information includes personnel vital sign information, behavioral status information and operation environment information;

[0017] The information preprocessing module is used to perform targeted preprocessing on the static information and the dynamic information respectively to obtain the processed static information vector and dynamic information vector;

[0018] The risk feature construction module is used to construct static information risk features and dynamic information risk features based on the processed static information vector and dynamic information vector, respectively, and merge them to generate personalized risk features for operation and maintenance based on multi-source personnel information.

[0019] The strategy optimization model module is used to construct the state space, action space, and reward function of reinforcement learning based on the personalized risk characteristics of the operation and maintenance operation, and obtain the reinforcement learning strategy optimization model through policy network training; the risk assessment and update module is used to perform risk assessment on the current operation and maintenance operation based on the reinforcement learning strategy optimization model, and dynamically update the model according to the actual operation and maintenance operation process results.

[0020] Furthermore, the information preprocessing module includes:

[0021] The static information encoding unit is used to convert static information according to different types of static information using the corresponding encoding format. Unordered static information is encoded using one-hot encoding, while ordered static information is encoded using label encoding. Static information vectors are constructed by concatenating feature vectors.

[0022] The dynamic information timing processing unit is used to construct a sliding time window for each dynamic information acquired in real time, perform noise reduction processing using the moving average filtering method, and perform normalization processing using the min-max normalization method to construct a dynamic information vector.

[0023] Furthermore, the strategy optimization model module includes:

[0024] The state space construction unit is used to combine the personalized risk characteristics of the current operation and maintenance task with the historical risk characteristics within the historical information backtracking step to construct the current state of reinforcement learning.

[0025] Action space construction unit, used to take risk assessment value as action output;

[0026] The reward function construction unit is used to construct a reward function consisting of a weighted combination of risk missed detection penalty and risk false alarm penalty.

[0027] The policy network optimization unit is used to take the current state as input, generate the corresponding risk assessment value through the policy network, and optimize and train the policy network by maximizing the cumulative reward objective function according to the reward function.

[0028] Thirdly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described offshore wind power operation and maintenance risk assessment method.

[0029] Fourthly, the present invention provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-mentioned offshore wind power operation and maintenance risk assessment method.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] 1. This invention breaks through the limitations of traditional assessment methods that only focus on objective factors of equipment and environment. It deeply integrates multi-source information of static basic attributes and dynamic physiological behavior of operation and maintenance personnel to construct personalized risk characteristics that cover the coupling effect of personnel and environment. It can accurately identify the differentiated risk levels of different personnel in the same work scenario and effectively avoid individual risk misjudgment caused by "one-size-fits-all" assessment.

[0032] 2. This invention introduces a reinforcement learning strategy optimization mechanism, using feedback from actual operation and maintenance as a reward signal to drive the iterative update of the strategy network. This enables the risk assessment model to continuously learn the risk evolution patterns in real operation scenarios, overcoming the shortcomings of traditional static models that cannot adapt to the complex and ever-changing environment of offshore wind power, and significantly improving the stability and timeliness of assessments during long-term operation.

[0033] 3. This invention designs a composite reward function consisting of a weighted penalty for missed detection and a penalty for false alarms. During model training, it simultaneously constrains the two types of evaluation biases. Combined with sliding window temporal feature extraction and nonlinear coupling mapping, it avoids frequent false alarms that interfere with normal operation and maintenance due to excessive conservatism, and also prevents the missed detection of hidden risks caused by insufficient feature capture, thus ensuring the reliability and practicality of offshore wind power operation and maintenance safety management. Attached Figure Description

[0034] 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 embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention.

[0035] Figure 1 The diagram shows a flowchart of the offshore wind power operation and maintenance risk assessment method based on multi-source personnel information and reinforcement learning disclosed in this invention.

[0036] Figure 2 A schematic diagram of the structure of the offshore wind power operation and maintenance risk assessment device based on multi-source personnel information and reinforcement learning disclosed in this invention is shown.

[0037] Explanation of reference numerals in the attached figures:

[0038] 100 - Multi-source information acquisition module, 200 - Information preprocessing module, 300 - Risk feature construction module, 400 - Strategy optimization model module. Detailed Implementation

[0039] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0040] Before providing a detailed description of the present invention, a brief overview of the prior art involved in the present invention will be given first.

[0041] In one embodiment, this application discloses a method for assessing the operation and maintenance risks of offshore wind power using multi-source personnel information and reinforcement learning, which includes the following steps:

[0042] S1. Collect multi-source information of personnel during offshore wind power operation and maintenance, and divide the multi-source information into static information and dynamic information according to the information change characteristics to construct a multi-source information dataset of personnel; wherein, the static information includes basic personnel information, health characteristic information and historical operation information, and the dynamic information includes personnel vital signs information, behavioral status information and operation environment information.

[0043] Specifically, basic personnel information includes age, gender, length of service, type and level of certification, training records, etc.; health characteristics information includes physical examination report indicators, past medical history, physical fitness test results, etc.; historical work information includes the number of historical work assignments, historical violation records, historical accident records, work completion rate, etc.; personnel vital signs information includes real-time monitoring data such as heart rate, blood pressure, blood oxygen saturation, and body temperature; behavioral status information includes the type of work action, fatigue status, and compliance with operating procedures, etc.; and work environment information includes wind speed, wind direction, temperature, humidity, wave height, and salt spray concentration, etc.

[0044] In this step, static information refers to information that remains essentially unchanged or changes very infrequently during the operation and maintenance cycle, while dynamic information refers to information that changes in real time or is updated frequently during the operation and maintenance process. This distinction provides a foundation for subsequent differentiated processing and feature construction.

[0045] S2, perform targeted preprocessing on the static information and the dynamic information respectively to obtain the processed static information vector and dynamic information vector.

[0046] The specific implementation process includes the following sub-steps.

[0047] S21, Encoding and conversion processing of static information in multi-source personnel information: The conversion is performed using appropriate encoding formats according to different static information types, specifically as follows: .

[0048] In the formula, C iRepresents the i-th static information, Hot(·) represents one-hot encoding, and Lobel(·) represents tag encoding; C nominal Static information representing unordered categories, such as gender; C ordinal Static information representing ordered categories, such as certificate level.

[0049] Further integrating the static information transformation results, a static information vector is constructed by concatenating feature vectors, represented as:

[0050] In the formula, X s This represents the processed static information. This represents the feature vector concatenation operator, where n represents the amount of static information.

[0051] S22, Time-series processing of dynamic information from multiple sources of personnel information: Constructing a sliding time window for each piece of dynamic information collected in real time, applying a moving average filtering method to reduce noise for the data within the window, and then normalizing the data to construct a dynamic information vector.

[0052] Specifically, firstly, for the real-time acquisition of the j-th dynamic information d j Construct a sliding time window of length L, denoted as: W j,t =[d j,t−L+1 ,d j,t−L+2 ,...,d j,t ]; where t represents time.

[0053] For the data within the window, a moving average filtering method is used for noise reduction, as shown below: .

[0054] Further processing using the min-max normalization method eliminates the influence of dimensions, and is expressed as:

[0055] In the formula, d max and d min They represent the first The maximum and minimum values ​​of each dynamic information.

[0056] Finally, the dynamic information vector is constructed and represented as follows: In the formula, X d This represents the processed dynamic information, where m represents the number of dynamic information items.

[0057] The static information of unordered categories is encoded using one-hot encoding, while the static information of ordered categories is encoded using label encoding. The static information vector is constructed by concatenating feature vectors.

[0058] S3, based on the processed static information vector and dynamic information vector, constructs static information risk features and dynamic information risk features respectively, and merges them to generate personalized risk features for operation and maintenance based on multi-source personnel information.

[0059] The specific implementation process includes the following sub-steps.

[0060] S31, Based on the processed static information vector, a static information risk feature is constructed through linear scoring of risk weights; denoted as: F s =W s X s In the formula, F s As a static information risk characteristic, W s Let X be the risk feature weight matrix. s This is the processed static information vector constructed in step S21. Through linear scoring with risk weights, basic personnel information, health characteristics, and historical work information are assigned corresponding risk weights to each dimension, achieving a quantitative mapping from static information to static risk characteristics.

[0061] S32, based on the processed dynamic information vector, constructs a dynamic information risk feature vector by calculating the mean, standard deviation, trend, and rate of change of the dynamic information within the window, expressed as: F d =h(X d In the formula, F d Let X be the dynamic information risk feature vector, h(·) be the feature mapping function, and X be the feature vector. d The dynamic information vector constructed in step S22.

[0062] Specifically, the feature mapping function h(·) calculates the mean, standard deviation, trend, and rate of change for each dimension of dynamic information within a sliding time window:

[0063] Mean: ;

[0064] Standard deviation: ;

[0065] Trend of change (slope of linear regression): ;

[0066] Rate of change: ;

[0067] The dynamic information risk feature vector F is constructed by concatenating the above statistics. d = .

[0068] Based on the deviations between personnel vital sign information and their corresponding baseline parameters, and the deviations between environmental information and their corresponding safety thresholds, a personnel-environment coupled risk characteristic is constructed through a nonlinear mapping function; expressed as:

[0069] ;

[0070] In the formula, Represents a nonlinear mapping function. Indicates the weight of the relationship between people and the environment. and Let represent the p-th vital sign information and its corresponding baseline parameters for the individual. and Let q represent the q-th environmental information and its corresponding safety threshold, respectively. This coupling feature captures the synergistic amplification effect between human physiological state and environmental deterioration through the product of vital sign deviation and environmental deviation.

[0071] S33, by integrating the static information risk features, dynamic information risk feature vectors, and personnel-environment coupled risk features, a personalized risk feature for operation and maintenance based on multi-source personnel information is constructed, represented as: In the formula, F represents the personalized risk characteristics of operation and maintenance based on multi-source personnel information, and ⊕ represents the feature vector concatenation operator. s F d F o These are static information risk characteristics, dynamic information risk characteristic vectors, and personnel-environment coupling risk characteristics, respectively.

[0072] S4. Based on the personalized risk characteristics of the operation and maintenance tasks, construct the state space, action space, and reward function of reinforcement learning, and obtain the reinforcement learning policy optimization model through policy network training.

[0073] The specific implementation process includes the following sub-steps.

[0074] S41, construct a reinforcement learning state space based on the personalized risk characteristics of operation and maintenance tasks. This involves concatenating the personalized risk characteristics of the operation and maintenance task at the current moment with the historical risk characteristics within the historical information backtracking step size, and using this as the current state for reinforcement learning, represented as: In the formula, S t Let F(t) represent the current state, F(t) be the personalized risk characteristics of the operation and maintenance work at time t constructed in step S33, and k represent the historical information backtracking step size. By introducing historical risk characteristics within the historical information backtracking step size, the state space contains the temporal evolution information of risk characteristics, providing a basis for the policy network to capture the dynamic changing trend of risks.

[0075] S42, Construct the risk assessment action space: that is, use the risk assessment value as the action output; represented as: In the formula, This represents the risk assessment value. The closer the value is to 1, the higher the risk of the operation and maintenance work. By using a continuous action space of [0,1], it is possible to achieve fine-grained quantitative output of risk level.

[0076] S43, Construct a reward function based on risk assessment feedback: The reward function is a weighted combination of risk missed detection penalty and risk false alarm penalty, expressed as: In the formula, and This represents the weighting coefficient.

[0077] The penalty for missed detection of risks is specifically expressed as follows: In the formula, This represents the safety risk boundary. When the actual risk characteristic F(t) exceeds the safety risk boundary, the assessed value α... t When the value is too low (i.e., a false negative occurs), the penalty increases exponentially with the degree of excess, and increases from (1−α) / ( ... t Linear amplification is used to strengthen the punishment for missed detections.

[0078] The penalty for false positives is specifically expressed as follows: In the formula, This represents the indicator function. When the actual risk characteristic F(t) is below the safety risk boundary (i.e., in a safe state), the assessed value α... t When the value is too high (i.e., a false alarm occurs), the penalty is proportional to the evaluation value, thereby constraining false alarm behavior. Furthermore, by setting γ1>γ2, the evaluation principle of "safety first" can be reflected, simultaneously constraining both missed detections and false alarms while prioritizing the minimization of the risk of missed detections.

[0079] S44, Construct a reinforcement learning policy optimization model: Using the current state as input, generate the corresponding risk assessment value through the policy network, represented as: In the formula, This represents the policy network to be optimized. These are the parameters of the policy network. In this embodiment, the policy network can adopt a fully connected neural network structure, and its output layer uses the Sigmoid activation function to constrain the output to the [0,1] interval, which matches the action space.

[0080] Further, the policy network is optimized and trained by maximizing the cumulative reward objective function according to the reward function described in step S43. The cumulative reward objective function introduces a reward discount factor to perform a weighted summation of rewards across multiple rounds. This is expressed as: In the formula, This represents the reward discount factor. Indicates the number of rounds. Representation strategy The expected value of the following.

[0081] The cumulative reward objective function introduces a reward discount factor to perform a weighted summation of rewards over multiple rounds. J(θ) is maximized through a policy gradient optimization method, and the policy network parameter θ is iteratively updated to obtain a reinforcement learning policy optimization model, enabling the model to continuously learn the risk evolution patterns in real-world operational scenarios.

[0082] The state space constructed by backtracking and piecing together historical information enables the policy network to perceive the temporal evolution of risks, overcoming the one-sidedness of single-moment feature assessment. Fine-grained quantification of risk levels is achieved through a continuous action space of [0,1]. A composite reward function consisting of a weighted combination of risk omission penalty and risk false alarm penalty is used to simultaneously constrain two types of assessment biases during model training. The omission penalty grows rapidly in an exponential form as the risk exceeds the limit, which avoids frequent false alarms that interfere with normal operation and maintenance due to excessive conservatism, and also prevents the omission of hidden risks due to insufficient feature capture. By introducing a cumulative reward objective function with a reward discount factor, the model takes into account both immediate feedback and long-term benefits, overcoming the shortcomings of traditional static models that cannot adapt to the complex and ever-changing environment of offshore wind power, and significantly improving the stability and timeliness of assessments during long-term operation.

[0083] S5. Based on the reinforcement learning strategy, optimize the model to perform a risk assessment on the current operation and maintenance task, and dynamically update the model according to the actual operation and maintenance task results.

[0084] Specifically, it includes the following sub-steps.

[0085] S51, personalize the risk characteristics of the current operation and maintenance work (i.e., the current state S constructed in step S41). t Input the optimized policy network, and output the current operation and maintenance risk assessment value, specifically expressed as: In the formula, Indicates the current operational risks, α t The risk assessment value output by the policy network. This represents the optimized policy network. θ ∗ P represents the policy network parameters obtained through optimization training in step S44. t The closer it is to 1, the higher the risk of the current operation and maintenance.

[0086] S52, based on the risk assessment value, the continuous risk assessment value is mapped to a discrete risk level using a preset risk level mapping function; expressed as: In the formula, L represents the risk level mapping function. tThis indicates a discrete risk level.

[0087] Furthermore, the risk level mapping function described in this embodiment can be set as a segmented threshold mapping, for example:

[0088] ;

[0089] In the formula, , This is a preset risk level threshold. Based on the classification results, corresponding safety control measures such as graded early warning, work permit control, and monitoring level adjustment can be triggered, providing intuitive decision support for on-site safety management of offshore wind power operation and maintenance.

[0090] S53 calculates the reinforcement learning reward function value based on the actual operation and maintenance process results, iteratively updates the policy network parameters, and realizes adaptive dynamic optimization of the risk assessment model.

[0091] Specifically, after the maintenance operation is completed at time t, the reward function value is calculated according to the actual maintenance operation results (including the measured risk characteristics F(t) and whether any security anomalies or false alarms occurred), based on the reward function defined in step S43: .

[0092] The calculated reward function value is then used to iteratively update the policy network parameters, as shown below:

[0093] ;

[0094] In the formula, Let J(θ) represent the learning rate, and J(θ) be the cumulative reward objective function described in step S44. Through the aforementioned "evaluation-feedback-update" closed-loop mechanism, the policy network can continuously absorb new operation and maintenance operation samples, achieve adaptive dynamic optimization of the risk assessment model, and continuously update the optimized policy network. This is used for risk assessment in the next moment.

[0095] The optimized strategy network outputs continuous risk assessment values ​​online, which are then divided into discrete risk levels via a risk level mapping function. This achieves a unified approach of fine-grained quantification of risk levels and intuitive hierarchical control, facilitating differentiated safety management on-site. By calculating reward function values ​​based on actual operation and maintenance process results and iteratively updating the strategy network parameters, the model can continuously learn the risk evolution patterns in real-world operation scenarios. This overcomes the shortcomings of traditional static models that cannot adapt to the complex and ever-changing environment of offshore wind power, significantly improving the stability and timeliness of assessments during long-term operation. Simultaneously, the closed-loop feedback mechanism and the composite reward function work together to continuously constrain two types of deviations: missed detections and false alarms, ensuring the reliability and practicality of offshore wind power operation and maintenance safety management.

[0096] In one embodiment, see Figure 2 This application discloses a personnel multi-source information and reinforcement learning-based offshore wind power operation and maintenance risk assessment device, used to implement the method described in Embodiment 1 above. The device includes:

[0097] The multi-source information acquisition module 100 is used to collect multi-source information from offshore wind power operation and maintenance personnel during operation and maintenance work, and divide the multi-source information into static information and dynamic information according to the information change characteristics to construct a personnel multi-source information dataset; wherein, the static information includes basic personnel information, health characteristic information and historical operation information, and the dynamic information includes personnel vital sign information, behavioral status information and operation environment information;

[0098] The information preprocessing module 200 is used to perform targeted preprocessing on the static information and the dynamic information respectively to obtain the processed static information vector and dynamic information vector;

[0099] The risk feature construction module 300 is used to construct static information risk features and dynamic information risk features based on the processed static information vector and dynamic information vector, respectively, and to merge and generate personalized risk features for operation and maintenance based on multi-source personnel information.

[0100] The strategy optimization model module 400 is used to construct the state space, action space, and reward function of reinforcement learning based on the personalized risk characteristics of the operation and maintenance operation, and obtain the reinforcement learning strategy optimization model through policy network training; the risk assessment and update module is used to perform risk assessment on the current operation and maintenance operation based on the reinforcement learning strategy optimization model, and dynamically update the model according to the actual operation and maintenance operation process results.

[0101] As an optional implementation of this application, the information preprocessing module 200 includes:

[0102] The static information encoding unit is used to convert static information according to different types of static information using the corresponding encoding format. Unordered static information is encoded using one-hot encoding, while ordered static information is encoded using label encoding. The static information vector is constructed by concatenating feature vectors.

[0103] The dynamic information timing processing unit is used to construct a sliding time window for each dynamic information acquired in real time, perform noise reduction processing using the moving average filtering method, and perform normalization processing using the minimum-maximum normalization method to construct a dynamic information vector.

[0104] As an optional implementation of this application, the strategy optimization model module 400 includes:

[0105] The state space construction unit is used to combine the personalized risk characteristics of the current operation and maintenance task with the historical risk characteristics within the historical information backtracking step to construct the current state of reinforcement learning.

[0106] Action space construction unit, used to take risk assessment value as action output;

[0107] The reward function construction unit is used to construct a reward function consisting of a weighted combination of risk missed detection penalty and risk false alarm penalty.

[0108] The policy network optimization unit is used to generate a corresponding risk assessment value through the policy network using the current state as input, and to optimize and train the policy network by maximizing the cumulative reward objective function according to the reward function.

[0109] This application achieves closed-loop management of the entire process from data perception to decision output through the collaborative work of five modules: multi-source information acquisition, preprocessing, feature construction, strategy optimization, and evaluation and updating. The functions of each module are decoupled, which facilitates system maintenance and upgrades.

[0110] The static information encoding unit and the dynamic information time-series processing unit adopt targeted preprocessing strategies (one-hot / label encoding, sliding window filtering normalization) for the characteristics of heterogeneous data, which effectively solves the problems of inconsistent dimensions of multi-source information, large noise interference, and difficulty in calculating non-numerical data, and provides a high-quality data foundation for subsequent feature construction.

[0111] The strategy optimization model module 400 integrates four units: state, action, reward, and network optimization. In particular, the reward function construction unit introduces a dual penalty mechanism for missed detections and false alarms, and the strategy network optimization unit optimizes the objective function based on cumulative rewards. This enables the device to have self-learning and dynamic adjustment capabilities, adapt to the complex and ever-changing operation and maintenance environment of offshore wind power, and significantly improve the accuracy and robustness of risk assessment.

[0112] In one embodiment, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the personnel multi-source information and reinforcement learning offshore wind power operation and maintenance risk assessment method described in the above embodiments.

[0113] Specifically, when the processor executes this computer program, it performs the following steps in sequence:

[0114] Multi-source information collected by personnel during offshore wind power operation and maintenance is divided into static and dynamic information to construct a dataset.

[0115] Static information is encoded and converted (one-hot / tag encoding) and concatenated, while dynamic information is filtered and normalized using a sliding window to obtain static and dynamic information vectors.

[0116] Static risk features are constructed based on static information vectors through linear scoring of risk weights, dynamic risk features are constructed based on dynamic information vectors through statistical feature mapping, and personnel-environment coupled risk features are constructed and fused to obtain personalized risk features.

[0117] Based on personalized risk characteristics, a state space, an action space, and a reward function containing penalties for missed detections / false alarms are constructed. The policy network is trained by maximizing the cumulative reward objective function.

[0118] The optimized policy network outputs risk values, which are then classified into risk levels using a mapping function. The policy network parameters are iteratively updated based on the reward values ​​calculated from the actual operation results.

[0119] The computer-readable storage medium may be, but is not limited to, read-only memory (ROM), random access memory (RAM), magnetic disk, optical disk, flash memory, USB flash drive, portable hard drive, solid-state drive (SSD), etc. This storage medium may exist independently of the electronic device or may be integrated within the electronic device.

[0120] By embedding the program code of the above method in a storage medium, the method can be easily migrated and deployed across different computing platforms (such as servers, personal computers, and embedded terminals). Particularly in offshore wind power operation and maintenance scenarios, the storage medium can be pre-installed in the field edge computing gateway or handheld terminal, enabling rapid localized risk assessment in offline or weak network environments. This reduces reliance on cloud computing power and improves system response speed and engineering applicability.

[0121] In one embodiment, this application provides an electronic device including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the personnel multi-source information and reinforcement learning offshore wind power operation and maintenance risk assessment method described in the above embodiments.

[0122] In this embodiment, the processor may be a central processing unit (CPU), graphics processing unit (GPU), digital signal processor (DSP), field-programmable gate array (FPGA), application-specific integrated circuit (ASIC) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0123] The memory may include high-speed random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device, flash memory device or other volatile solid-state storage device.

[0124] In specific application scenarios, this electronic device can take the following forms:

[0125] Cloud server: Deployed in an onshore data center, it receives massive amounts of operation and maintenance data from multiple offshore wind farms, uses high-performance GPU clusters to train large-scale policy networks and optimize global models, and then distributes the optimized models to on-site terminals.

[0126] Edge computing gateway: Deployed at offshore wind power booster stations or submarine cable landing points, it has strong local computing power and is responsible for real-time collection of multi-source information, execution of lightweight inference (risk assessment) and local model updates to meet low latency response requirements.

[0127] Smart wearable / handheld terminals: worn or held by maintenance personnel, these terminals collect vital signs and behavioral data in real time, run lightweight assessment algorithms, and provide immediate feedback on individual risk levels and early warning information.

[0128] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention. Where there is no conflict, the above embodiments and features of the embodiments of the present invention can be combined with each other.

Claims

1. A method for assessing operational and maintenance risks of offshore wind power, characterized in that, Includes the following steps: S1. Collect multi-source information of personnel during offshore wind power operation and maintenance, and divide the multi-source information into static information and dynamic information according to the information change characteristics to construct a multi-source information dataset of personnel; wherein, the static information includes basic personnel information, health characteristic information and historical operation information, and the dynamic information includes personnel vital signs information, behavioral status information and operation environment information; S2, perform targeted preprocessing on the static information and the dynamic information respectively to obtain the processed static information vector and dynamic information vector; S3, based on the processed static information vector and dynamic information vector, construct static information risk features and dynamic information risk features respectively, and merge them to generate personalized risk features for operation and maintenance based on multi-source personnel information; S4. Based on the personalized risk characteristics of the operation and maintenance tasks, construct the state space, action space and reward function of reinforcement learning, and obtain the reinforcement learning policy optimization model through policy network training. S5. Based on the reinforcement learning strategy, optimize the model to perform a risk assessment on the current operation and maintenance task, and dynamically update the model according to the actual operation and maintenance task results.

2. The method according to claim 1, characterized in that, Step S2 specifically includes: The static information in the multi-source information of personnel is encoded and converted: the corresponding encoding format is used to convert according to different static information types. The static information of unordered categories is encoded using one-hot encoding, and the static information of ordered categories is encoded using label encoding. The static information vector is constructed by concatenating feature vectors. Time-series processing is performed on dynamic information from multiple sources of personnel information: a sliding time window is constructed for each dynamic information collected in real time, and the data within the window is processed by moving average filtering to reduce noise and then normalized to construct a dynamic information vector.

3. The method according to claim 1, characterized in that, Step S3 specifically includes: Based on the processed static information vector, static information risk features are constructed through linear scoring of risk weights. Based on the processed dynamic information vector, a dynamic information risk feature vector is constructed by calculating the mean, standard deviation, trend and rate of change of the dynamic information within the window. Based on the deviation between personnel vital signs information and their corresponding benchmark parameters, and the deviation between environmental information and its corresponding safety threshold, a personnel-environment coupled risk feature is constructed through a nonlinear mapping function. By integrating the static information risk features, dynamic information risk feature vectors, and personnel-environment coupled risk features, personalized risk features for operation and maintenance operations based on multi-source personnel information are constructed.

4. The method according to claim 1, characterized in that, Step S4 specifically includes: A reinforcement learning state space is constructed based on the personalized risk characteristics of operation and maintenance tasks: the personalized risk characteristics of the operation and maintenance task at the current moment are concatenated with the historical risk characteristics within the historical information backtracking step size, which is used as the current state of reinforcement learning. Constructing a risk assessment action space: using risk assessment values ​​as action outputs; Construct a reward function based on risk assessment feedback: the reward function is composed of a weighted combination of risk missed detection penalty and risk false alarm penalty; Construct a reinforcement learning policy optimization model: using the current state as input, generate the corresponding risk assessment value through the policy network, and optimize the policy network by maximizing the cumulative reward objective function according to the reward function. The cumulative reward objective function introduces a reward discount factor to perform a weighted summation of rewards over multiple rounds.

5. The method according to claim 1, characterized in that, Step S5 specifically includes: Input the personalized risk characteristics of the current operation and maintenance task into the optimized policy network, and output the risk assessment value of the current operation and maintenance task; Based on the risk assessment value, the continuous risk assessment value is mapped to a discrete risk level through a preset risk level mapping function; The reinforcement learning reward function value is calculated based on the actual operation and maintenance process results, and the policy network parameters are iteratively updated to achieve adaptive dynamic optimization of the risk assessment model.

6. A personnel multi-source information and reinforcement learning-based offshore wind power operation and maintenance risk assessment device, characterized in that, include: The multi-source information acquisition module is used to collect multi-source information from offshore wind power operation and maintenance personnel during operation and maintenance work, and to divide the multi-source information into static information and dynamic information according to the information change characteristics, thereby constructing a multi-source information dataset for personnel; wherein, the static information includes basic personnel information, health characteristic information and historical operation information, and the dynamic information includes personnel vital sign information, behavioral status information and operation environment information; The information preprocessing module is used to perform targeted preprocessing on the static information and the dynamic information respectively to obtain the processed static information vector and dynamic information vector; The risk feature construction module is used to construct static information risk features and dynamic information risk features based on the processed static information vector and dynamic information vector, respectively, and merge them to generate personalized risk features for operation and maintenance based on multi-source personnel information. The strategy optimization model module is used to construct the state space, action space, and reward function of reinforcement learning based on the personalized risk characteristics of the operation and maintenance operation, and obtain the reinforcement learning strategy optimization model through policy network training; the risk assessment and update module is used to perform risk assessment on the current operation and maintenance operation based on the reinforcement learning strategy optimization model, and dynamically update the model according to the actual operation and maintenance operation process results.

7. The apparatus according to claim 6, characterized in that, The information preprocessing module includes: The static information encoding unit is used to convert static information according to different types of static information using the corresponding encoding format. Unordered static information is encoded using one-hot encoding, while ordered static information is encoded using label encoding. Static information vectors are constructed by concatenating feature vectors. The dynamic information timing processing unit is used to construct a sliding time window for each dynamic information acquired in real time, perform noise reduction processing using the moving average filtering method, and perform normalization processing using the min-max normalization method to construct a dynamic information vector.

8. The apparatus according to claim 6, characterized in that, The strategy optimization model module includes: The state space construction unit is used to combine the personalized risk characteristics of the current operation and maintenance task with the historical risk characteristics within the historical information backtracking step to construct the current state of reinforcement learning. Action space construction unit, used to take risk assessment value as action output; The reward function construction unit is used to construct a reward function consisting of a weighted combination of risk missed detection penalty and risk false alarm penalty. The policy network optimization unit is used to take the current state as input, generate the corresponding risk assessment value through the policy network, and optimize and train the policy network by maximizing the cumulative reward objective function according to the reward function.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the offshore wind power operation and maintenance risk assessment method as described in any one of claims 1 to 5.

10. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the offshore wind power operation and maintenance risk assessment method as described in any one of claims 1 to 5.

Images