Ammonia fuel power container ship leakage risk assessment method, device and equipment

Abstract

The embodiment of the invention provides a leakage risk assessment method, device and equipment for an ammonia fuel power container ship. The assessment method comprises the following steps: acquiring multi-source parameter data of the ammonia fuel power container ship; inputting the multi-source parameter data into a joint coding layer of an ammonia gas concentration distribution prediction model for feature coding to obtain a coding feature vector; inputting the coding feature vector into a decoding network layer of an ammonia concentration distribution prediction model for vector mapping to obtain a three-dimensional ammonia concentration distribution field vector; and obtaining an ammonia fuel leakage risk assessment result according to the ammonia concentration distribution field vector. According to the embodiment of the invention, the ammonia concentration values at different spatial positions can be intuitively presented through three-dimensional ammonia concentration distribution field vector output, and the quantitative evaluation efficiency and decision reliability of the ammonia fuel leakage risk are remarkably improved.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of ship safety, and particularly to a method, device and equipment for evaluating the leakage risk of an ammonia fuel-powered container ship. Background Art

[0002] Ammonia fuel is a hydrogen carrier with high energy density, and hardly emits greenhouse gases during the combustion process, having the potential to replace traditional fossil fuels. Against this background, the design and application of ships powered by ammonia have attracted much attention. However, the characteristics of ammonia fuel itself and the resulting safety risks are the main bottlenecks for its wide application. Ammonia has strong toxicity and corrosiveness, and may cause relatively serious safety problems for personnel, equipment and the environment during the process of leakage and diffusion.

[0003] Traditional rule-based qualitative assessment means can only roughly describe the outline of the leakage or failure risk of the ammonia fuel system, and cannot accurately reflect the influence range and severity after an accident, thus resulting in insufficient or excessive safety protection design margins. Assessment schemes with higher precision often rely on complex numerical calculation models, which are time-consuming and difficult to adapt to rapidly changing design requirements, and cannot timely guide rapid emergency response in the actual ship operation environment. In addition, the experience and conclusions obtained from these risk assessments are only limited to the assessment object itself, and are saved in the form of scattered research reports or assessment reports, without systematically analyzing and processing the risk assessment knowledge of the same type. Summary of the Invention

[0004] The technical problem to be solved by the embodiments of the present invention is to provide a method, device and equipment for evaluating the leakage risk of an ammonia fuel-powered container ship, which can output a three-dimensional ammonia concentration distribution field vector, visually present the ammonia concentration values at different spatial positions, and significantly improve the quantification efficiency of ammonia fuel leakage risk assessment and decision-making reliability.

[0005] To solve the above technical problem, the technical solution of the embodiments of the present invention is as follows:

[0006] A method for evaluating the leakage risk of an ammonia fuel-powered container ship includes:

[0007] Obtain multi-source parameter data of the ammonia fuel-powered container ship;

[0008] Input the multi-source parameter data into the joint coding layer of the ammonia concentration distribution prediction model for feature coding to obtain a coded feature vector;

[0009] Input the coded feature vector into the decoding network layer of the ammonia concentration distribution prediction model for vector mapping to obtain a three-dimensional ammonia concentration distribution field vector;

[0010] Obtain the ammonia fuel leakage risk assessment result according to the ammonia concentration distribution field vector; wherein, the training process of the ammonia concentration distribution prediction model includes: acquiring multi-source training data of an ammonia fuel-powered container ship; establishing an ammonia fuel leakage risk knowledge base according to the multi-source training data; inputting the sample data in the knowledge base into the joint coding layer of a preset network model for feature coding to obtain a training coding feature vector; inputting the training coding feature vector into the decoding network layer of the preset network model for vector mapping to obtain a training ammonia concentration distribution field vector, and obtaining the ammonia concentration distribution prediction model.

[0011] Optionally, establishing an ammonia fuel leakage risk knowledge base according to the multi-source training data includes:

[0012] Clean the multi-source training data to obtain sample data;

[0013] Manage and store the sample data in the form of knowledge classification and knowledge labels to obtain a first knowledge base;

[0014] Audit the first knowledge base, and remove duplicate data and conflicting rules to obtain a second knowledge base;

[0015] Perform semantic enhancement parsing on the documents in the second knowledge base, and extract the rules and formulas in the documents after semantic enhancement parsing;

[0016] Fill the extracted rules and formulas into the rule base and mathematical formula base of the second knowledge base;

[0017] Define knowledge according to the parameter relationship in the simulation parameter library of the second knowledge base;

[0018] Inject the knowledge into the mathematical formula base and rule base of the second knowledge base to obtain an ammonia fuel leakage risk knowledge base.

[0019] Optionally, inputting the sample data in the knowledge base into the joint coding layer of a preset network model for feature coding to obtain a training coding feature vector includes:

[0020] Determine the operation record data x j [i] of each position of the ammonia fuel-powered container ship and the corresponding ammonia concentration C j [i],

[0021] wherein, x j [i] is the operation record data of the jth position of the ammonia fuel-powered container ship in the ith scenario, i = 1, 2, 3,..., M, M is the total number of scenarios, j = 1, 2, 3,..., N, N is the total number of positions, C j[i] is the ammonia concentration at the j-th position of the ammonia-fueled container ship in the i-th scenario;

[0022] Normalize the operation record data x j [i], input it into the joint encoding layer of the preset network model for feature encoding, and obtain the first training data X j [i];

[0023] According to Z j [i]=f encoder (X j [i], θ enc )Perform feature encoding on the first training data to obtain the training encoded feature vector,

[0024] Among them, Z j [i] is the training encoded feature vector, f encoder is the forward propagation function of the joint encoding layer, θ enc is the preset joint encoding layer parameter of the joint encoding layer.

[0025] Optionally, input the training encoded feature vector into the decoding network layer of the preset network model for vector mapping to obtain the training ammonia concentration distribution field vector, and obtain the ammonia concentration distribution prediction model, including:

[0026] According to Y j [i]=f decoder (Z j [i], θ dec )Perform vector mapping on the training encoded feature vector to obtain the training ammonia concentration distribution field vector,

[0027] Among them, Y j [i] is the training ammonia concentration distribution field vector, f decoder is the forward propagation function of the decoding network layer, θ dec is the preset decoding network layer parameter of the joint encoding layer;

[0028] According to , obtain the first average difference,

[0029] Among them, k i is the first average difference in the i-th scenario, Y j [i] is the training ammonia concentration distribution field vector, C j [i] is the ammonia concentration at the j-th position of the ammonia-fueled container ship in the i-th scenario;

[0030] According to , obtain the first average mean value,

[0031] Among them, h is the first average mean value under various scenarios, k iis the first average difference for the i-th scenario;

[0032] Adjust the parameters of the preset joint encoding layer and the parameters of the preset decoding network layer, and repeat the above process for a set number of times to obtain the first average mean h for each time, t where t = 1, 2, 3,..., R, and R is the set number of times;

[0033] According to ,

[0034] obtain the joint encoding layer parameters θ corresponding to the minimum first average mean, e and the decoding network layer parameters θ d ;

[0035] Take the joint encoding layer parameters θ e as the preset joint encoding layer parameters of the joint encoding layer and take the decoding network layer parameters θ d as the preset decoding network layer parameters of the joint encoding layer to obtain an embedded ammonia concentration distribution prediction model.

[0036] Optionally, input the multi-source parameter data into the joint encoding layer of the ammonia concentration distribution prediction model for feature encoding to obtain an encoded feature vector, including:

[0037] Clean the multi-source parameter data to obtain first parameter data;

[0038] Normalize the first parameter data, input it into the joint encoding layer of the ammonia concentration distribution prediction model for feature encoding to obtain second parameter data;

[0039] According to B j = f encoder (A j , θ e ) perform feature encoding on the first data to obtain a training encoded feature vector,

[0040] where B j is the training encoded feature vector, f encoder is the forward propagation function of the joint encoding layer, A j is the second parameter data, and θ e are the joint encoding layer parameters.

[0041] Optionally, input the encoded feature vector into the decoding network layer of the ammonia concentration distribution prediction model for vector mapping to obtain a three-dimensional ammonia concentration distribution field vector, including:

[0042] According to C j = f decoder (B j , θ dPerform vector mapping on the training encoded feature vectors to obtain the training ammonia concentration distribution field vectors.

[0043] Among them, C j is the training ammonia concentration distribution field vector, f decoder is the forward propagation function of the decoding network layer, B j is the training encoded feature vector, and θ d is the parameter of the decoding network layer.

[0044] Optionally, according to the ammonia concentration distribution field vector, obtain the ammonia fuel leakage risk assessment result, including:

[0045] Compare the ammonia concentration distribution field vector with the first set ammonia concentration threshold and the second set ammonia concentration threshold.

[0046] If , obtain the high-risk assessment result of ammonia fuel leakage;

[0047] If , obtain the low-risk assessment result of ammonia fuel leakage;

[0048] If , obtain the medium-risk assessment result of ammonia fuel leakage;

[0049] Among them, C j is the training ammonia concentration distribution field vector, T1 is the first set ammonia concentration threshold, and T2 is the second set ammonia concentration threshold.

[0050] An embodiment of the present invention also provides an ammonia fuel-powered container ship leakage risk assessment device, including:

[0051] An acquisition module, configured to acquire multi-source parameter data of an ammonia fuel-powered container ship; acquire multi-source training data of an ammonia fuel-powered container ship.

[0052] A processing module, configured to input the multi-source parameter data into the joint encoding layer of the ammonia gas concentration distribution prediction model for feature encoding to obtain encoded feature vectors; input the encoded feature vectors into the decoding network layer of the ammonia gas concentration distribution prediction model for vector mapping to obtain three-dimensional ammonia concentration distribution field vectors; obtain ammonia fuel leakage risk assessment results according to the ammonia concentration distribution field vectors; establish an ammonia fuel leakage risk knowledge base according to the multi-source training data; input the sample data in the knowledge base into the joint encoding layer of a preset network model for feature encoding to obtain training encoded feature vectors; input the training encoded feature vectors into the decoding network layer of the preset network model for vector mapping to obtain training ammonia concentration distribution field vectors, and obtain an ammonia gas concentration distribution prediction model.

[0053] An embodiment of the present invention also provides a computing device, including:

[0054] One or more processors;

[0055] A storage device for storing one or more programs, which when executed by the one or more processors cause the one or more processors to implement the method as described above.

[0056] An embodiment of the present invention also provides a computer-readable storage medium, in which a program is stored, and when the program is executed by a processor, the method as described above is implemented.

[0057] The above solution of the embodiment of the present invention has at least the following beneficial effects:

[0058] With the above solution of the embodiment of the present invention, through the output of the three-dimensional ammonia concentration distribution field vector, the ammonia concentration values at different spatial positions can be visually presented, the specific threat levels of leakage to personnel, equipment and the environment can be quantified, and the rapid quantitative assessment of ammonia fuel leakage risk can be realized, significantly improving the decision-making reliability in its design and operation stages.

[0059] By establishing a knowledge base for ammonia fuel leakage risk and storing multi-source training data in a structured manner, the standardized management and reuse of knowledge for similar risk assessments are realized.

[0060] The input characteristics of the multi-source parameter data enable the model to analyze the influence degrees of different factors on the leakage risk. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 is a schematic flowchart of a method for assessing ammonia fuel-powered container ship leakage risk provided by an embodiment of the present invention.

[0062] Figure 2 is a system architecture diagram of a knowledge base for ammonia fuel leakage risk assessment provided by an embodiment of the present invention.

[0063] Figure 3 is a flowchart of the use and risk assessment of a knowledge base system for ammonia fuel leakage risk assessment provided by an embodiment of the present invention.

[0064] Figure 4 is a technical framework diagram of multi-modal knowledge fusion provided by an embodiment of the present invention.

[0065] Figure 5 is a schematic flowchart of intelligent knowledge extraction provided by an embodiment of the present invention.

[0066] Figure 6 is a schematic diagram of the parsing and knowledge extraction of a specification document provided by an embodiment of the present invention.

[0067] Figure 7 is a training-inference flowchart of an agent model provided by an embodiment of the present invention.

[0068] Figure 8 It is a schematic diagram of the proxy model architecture provided by an embodiment of the present invention.

[0069] Figure 9 It is a flowchart of the intelligent hybrid inference technology provided by an embodiment of the present invention.

[0070] Figure 10 It is a schematic diagram of the modules of an ammonia fuel-powered container ship leakage risk assessment device provided by an embodiment of the present invention. Detailed implementation manners

[0071] Hereinafter, the exemplary embodiments of the present invention will be described in more detail with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present invention can be more thoroughly understood and the scope of the present invention can be fully conveyed to those skilled in the art.

[0072] As Figure 1 shown, an embodiment of the present invention provides an ammonia fuel-powered container ship leakage risk assessment method, including:

[0073] Step 11, obtaining multi-source parameter data of the ammonia fuel-powered container ship;

[0074] Step 12, inputting the multi-source parameter data into the joint encoding layer of the ammonia concentration distribution prediction model for feature encoding to obtain an encoded feature vector;

[0075] Step 13, inputting the encoded feature vector into the decoding network layer of the ammonia concentration distribution prediction model for vector mapping to obtain a three-dimensional ammonia concentration distribution field vector;

[0076] Step 14, obtaining an ammonia fuel leakage risk assessment result according to the ammonia concentration distribution field vector;

[0077] Among them, the training process of the ammonia concentration distribution prediction model includes:

[0078] Step 101, obtaining multi-source training data of the ammonia fuel-powered container ship;

[0079] Step 102, establishing an ammonia fuel leakage risk knowledge base according to the multi-source training data;

[0080] Step 103, inputting the sample data in the knowledge base into the joint encoding layer of a preset network model for feature encoding to obtain a training encoded feature vector;

[0081] Step 104: Input the training encoded feature vectors into the decoding network layer of a preset network model for vector mapping to obtain training ammonia concentration distribution field vectors and an ammonia concentration distribution prediction model.

[0082] In this example, through the output of the three-dimensional ammonia concentration distribution field vectors, the ammonia concentration values at different spatial positions can be visually presented, the specific threat levels of leaks to personnel, equipment, and the environment can be quantified, and a rapid quantitative assessment of ammonia fuel leakage risks can be achieved, significantly improving the decision-making reliability during the design and operation phases.

[0083] By establishing a knowledge base for ammonia fuel leakage risks and storing multi-source training data in a structured manner, standardized management and reuse of knowledge for similar risk assessments are achieved.

[0084] The input characteristics of the multi-source parameter data enable the model to analyze the influence degrees of different factors on leakage risks.

[0085] In an optional embodiment of the present invention, in step 101, obtaining multi-source training data for an ammonia fuel-powered container ship may include:

[0086] Step 1011: Obtain multi-source training data including ship design drawing training data, operation record training data, experimental report training data, and rule manual training data; specifically, the ship design drawing training data may include hull structure parameters (fuel tank location, size, and material, ammonia fuel pipeline network layout, ventilation system design, and safety protection device design data), three-dimensional geometric models (CAD drawings, three-dimensional modeling files);

[0087] The operation record training data may include data of ammonia fuel storage tanks (pressure, temperature, liquid level, and flow rate), operating status of pipeline systems, equipment vibration data, and environmental parameters (wind speed, wind direction, temperature, and humidity);

[0088] The experimental report training data may include ammonia leakage rates and diffusion cloud maps under different conditions;

[0089] The rule manual training data may include guiding documents on ammonia fuel safety in international codes, industry standards, and enterprise internal regulations.

[0090] Step 102: Establish a knowledge base for ammonia fuel leakage risks according to the multi-source training data, including:

[0091] Step 1021: Clean the multi-source training data to obtain sample data; specifically, it may include processing missing values (filling or removing) and detecting and correcting outliers in the multi-source training data;

[0092] Step 1022: Manage and store the sample data in the form of knowledge classification and knowledge tags to obtain a first knowledge base, which includes a document library, a rule library, a mathematical formula library, a geometric model library, a simulation parameter library, a simulation sample library, and an agent model library. Specifically, the document library is used to store guiding documents on ammonia fuel safety in international norms, industry standards, and enterprise internal regulations; the rule library is used to store rules refined based on expert experience and historical cases; the mathematical formula library is used to collect formulas related to ammonia fuel characteristics, fluid mechanics, and risk assessment; the geometric model library is used for hull structure parameters (fuel tank location, size and material, ammonia fuel pipeline network layout, ventilation system design, and safety protection device design data), three-dimensional geometric models (CAD drawings, three-dimensional modeling files); the simulation parameter library is used to store data of ammonia fuel storage tanks (pressure, temperature, liquid level, and flow rate), operating status of pipeline systems, equipment vibration data, and environmental parameters (wind speed, wind direction, temperature, and humidity); the simulation sample library is used to store the result data of past simulation experiments, including ammonia leakage rates and diffusion cloud maps under different conditions; the agent model library is used to store the predicted ammonia concentration distribution model after subsequent training.

[0093] Step 1023: Audit the first knowledge base, eliminate duplicate data and conflicting rules, and obtain a second knowledge base. Specifically, it can identify data that is exactly the same or highly similar in the first knowledge base to avoid storage redundancy; discover rules with logical contradictions or threshold conflicts in the first knowledge base to avoid ambiguity during reasoning.

[0094] Step 1024: Perform semantic enhancement parsing on the documents in the second knowledge base, and extract the rules and formulas in the semantically enhanced parsed documents.

[0095] Step 1025: Fill the extracted rules and formulas into the rule library and mathematical formula library of the second knowledge base.

[0096] Step 1026: Define knowledge according to the parameter relationships in the simulation parameter library of the second knowledge base.

[0097] Step 1027: Inject the knowledge into the mathematical formula library and rule library of the second knowledge base to obtain an ammonia fuel leakage risk knowledge base.

[0098] In this example, Step 1021 avoids model training deviation caused by data noise through missing value filling and outlier correction.

[0099] Step 1022 manages the sample data in the form of knowledge classification and knowledge tags, which can make a large amount of complex data become orderly and well-organized. When querying data in this regard, the required data can be quickly located through tags and classification. The managed sample data is classified and stored, saving storage space and facilitating data backup and recovery.

[0100] Step 1023 eliminates duplicate data and conflict rules through auditing, avoiding the emergence of contradictory information in the knowledge base (such as conflicts in the definition of safety thresholds in different specifications), and ensuring the reliability of the model inference logic.

[0101] Step 1024 performs semantic enhancement parsing on the documents in the second knowledge base, enabling a deeper understanding of the content in the documents. Traditional document parsing may simply extract text information, while semantic enhancement parsing can analyze the meaning and logical relationships behind the text. For example, for specification and guideline documents, through semantic enhancement parsing, the specific requirements and potential implicit rules for ammonia fuel leakage risk assessment can be accurately understood, avoiding missing important information, and thus fully mining the valuable content for building the knowledge base in the documents.

[0102] Step 1025 extracts the rules and formulas in the documents after semantic enhancement parsing and fills them into the rule base and the mathematical formula base, realizing the integration of knowledge scattered in various documents. These rules and formulas are important bases for ammonia fuel leakage risk assessment. After integration, the rule base and the mathematical formula base are more perfect, providing more comprehensive and accurate knowledge support for subsequent assessment work. For example, integrating the rules on ammonia fuel safety concentration standards in different specification documents into the rule base facilitates unified reference during assessment.

[0103] Step 1026 defines knowledge based on the parameter relationships in the simulation parameter library, enabling relevant knowledge to be refined from actual simulation parameters. There are complex relationships between different ammonia fuel leakage scenario parameters (such as leakage location, intensity, and time) and operating condition parameters (such as environmental wind speed, wind direction, temperature, and humidity). By analyzing the knowledge defined from these relationships, the influencing factors and laws of ammonia fuel leakage risk can be more accurately described. For example, discovering the relationship knowledge between ammonia fuel leakage intensity and diffusion range under a specific environmental wind speed and injecting it into the knowledge base helps to more precisely assess the risk.

[0104] Step 1027 injects the defined knowledge into the mathematical formula base and the rule base, further enriching the content of the knowledge base. The mathematical formula base and the rule base are important components of the knowledge base. The injection of these new knowledge can provide more calculation bases and judgment criteria for risk assessment, and enable the knowledge between different libraries to be shared and associated with each other, improving the utilization efficiency of knowledge.

[0105] Establish semantic associations among the knowledge in each library, so that the knowledge in the knowledge base forms an organic whole. Through semantic associations, the internal relationships between different pieces of knowledge can be clearly shown, such as the associations between rules and formulas, and between knowledge of different parameter relationships. In this way, when conducting an ammonia fuel leakage risk assessment, the assessors can consider various factors more comprehensively, avoid using knowledge in isolation, thereby improving the accuracy and reliability of the assessment, and at the same time, it also helps to manage and maintain the knowledge base more deeply.

[0106] In an alternative embodiment of the present invention, in step 103, the sample data in the knowledge base is input into the joint encoding layer of a preset network model for feature encoding to obtain a training encoded feature vector, including:

[0107] Step 1031: Determine the operation record data x j [i] and the corresponding ammonia concentration C j [i] at each position of the ammonia fuel-powered container ship according to the sample data in the knowledge base.

[0108] Where x j [i] is the operation record data at the jth position of the ammonia fuel-powered container ship in the ith scenario, i = 1, 2, 3,..., M, where M is the total number of scenarios, j = 1, 2, 3,..., N, where N is the total number of positions, and C j [i] is the ammonia concentration at the jth position of the ammonia fuel-powered container ship in the ith scenario;

[0109] Step 1032: Normalize the operation record data x j [i], and input it into the joint encoding layer of the preset network model for feature encoding to obtain the first training data X j [i]. Specifically, normalization can be performed according to X j [i] = (x j [i] - x min ) / (x max - x min ),

[0110] Where X j [i] is the first training data, x j [i] is the operation record data, x min is the minimum value in the operation record data, and x max is the maximum value in the operation record data;

[0111] Step 1033: According to Z j [i] = f encoder (X j [i], θ enc(0) Feature-encode the first training data to obtain a training encoded feature vector.

[0112] where Z j [i] is the training encoded feature vector, and f encoder is the forward propagation function of the joint encoding layer, and θ enc is the preset joint encoding layer parameter of the joint encoding layer. Specifically, the preset joint encoding layer parameter may include a first preset weight matrix, a first preset bias vector, and a first preset embedding layer parameter.

[0113] In this example, multi-source data such as ship design parameters (such as the position of the fuel tank), operating status (such as the pressure in the storage tank), and environmental conditions (such as wind speed) are integrated into a unified feature vector.

[0114] Compress the operation record data with different dimensions to the interval [0, 1] to avoid model training deviation caused by feature scale differences. The joint encoding layer can simultaneously process numerical data (such as the pressure in the storage tank and environmental wind speed) and spatial data (such as the coordinate of the fuel tank position), and automatically learn the correlation weights of different features through the weight matrix to generate a low-dimensional dense feature vector Z j [i].

[0115] In an alternative embodiment of the present invention, in step 104, input the training encoded feature vector into the decoding network layer of the preset network model for vector mapping to obtain a training ammonia concentration distribution field vector, and obtain an ammonia concentration distribution prediction model, including:

[0116] Step 1041: According to Y j [i]=f decoder (Z j [i], θ dec ) perform vector mapping on the training encoded feature vector to obtain a training ammonia concentration distribution field vector.

[0117] where Y j [i] is the training ammonia concentration distribution field vector, f decoder is the forward propagation function of the decoding network layer, and θ dec is the preset decoding network layer parameter of the joint encoding layer;

[0118] Step 1042: According to , obtain the first average difference.

[0119] where k i is the first average difference of the i-th scenario, Y j [i] is the training ammonia concentration distribution field vector, and C j [i] is the ammonia concentration at the j-th position of the ammonia fuel-powered container ship in the i-th scenario;

[0120] Step 1043. According to , the first average mean value is obtained.

[0121] where h is the first average mean value in various scenarios, and k i is the first average difference in the i-th scenario;

[0122] Step 1044. Adjust the parameters of the preset joint encoding layer and the parameters of the preset decoding network layer, and repeat the above process for the set number of times to obtain the first average mean value h t for each time, where t = 1, 2, 3,..., R, and R is the set number of times;

[0123] Step 1045. According to ,

[0124] the joint encoding layer parameters θ e and the decoding network layer parameters θ d corresponding to the minimum first average mean value are obtained;

[0125] Step 1046. Use the joint encoding layer parameters θ e as the preset joint encoding layer parameters of the joint encoding layer and use the decoding network layer parameters θ d as the preset decoding network layer parameters of the joint encoding layer to obtain an embedded ammonia concentration distribution prediction model.

[0126] In this example, the decoding network layer restores the low-dimensional encoded feature vector to a three-dimensional ammonia concentration distribution field vector, realizing the cross-domain conversion from the data feature space to the physical space concentration distribution. The three-dimensional concentration field vector can correspond to the concentration distributions in the transverse, longitudinal, and vertical directions of the cabin according to the ship coordinate system, accurately locating high-risk areas.

[0127] Calculate the average difference between the predicted concentration and the actual concentration at all positions within a single scenario to ensure the local accuracy of the model under specific working conditions.

[0128] Take the average value of the errors for all scenarios to force the model to take into account the generalization ability under different working conditions and avoid overfitting to a single scenario.

[0129] Synchronously update the parameters of the joint encoding layer and the decoding network layer to form a closed loop of "feature encoding → concentration prediction → error feedback → parameter correction".

[0130] If the decoding network layer has a large error in predicting the concentration at a certain position, adjust the weight matrix of the joint encoding layer to enhance the encoding intensity of the associated features at that position (such as the vibration data of nearby pipelines), thereby indirectly improving the prediction accuracy.

[0131] Prevent overfitting by limiting the number of iterations to ensure that the model converges within a reasonable time.

[0132] The model optimized by the global error can adapt to new scenarios that have not been trained. When the model is trained, it only contains the scenario data of "high temperature in summer". When real-time data of the "low temperature in winter" scenario is input, the joint encoding layer can capture the change pattern of temperature parameters through feature encoding, and the decoding network layer can reasonably extrapolate the concentration distribution in the low temperature environment based on the historical temperature-concentration correlation law.

[0133] In an alternative embodiment of the present invention, in step 11, obtaining multi-source parameter data of an ammonia fuel-powered container ship, including:

[0134] Step 111, obtaining real-time operation record data of an ammonia fuel-powered container ship. Specifically, the operation record data may include real-time data (pressure, temperature, liquid level, and flow rate) of the ammonia fuel storage tank, the real-time operation status of the pipeline system, real-time equipment vibration data, and real-time environmental parameters (wind speed, wind direction, temperature, and humidity). The operation record data can be obtained by sensors arranged on the ammonia fuel-powered container ship.

[0135] In this example, data such as the storage tank pressure and pipeline vibration are collected in real time through sensors, enabling the model to respond to real-time state changes during ship operation, achieving a closed loop of "dynamic monitoring - instant evaluation" instead of static offline analysis.

[0136] In an alternative embodiment of the present invention, in step 12, inputting the multi-source parameter data into the joint encoding layer of the ammonia concentration distribution prediction model for feature encoding to obtain an encoded feature vector, including:

[0137] Step 121, cleaning the multi-source parameter data to obtain first parameter data;

[0138] Step 122, performing normalization processing on the first parameter data, inputting it into the joint encoding layer of the ammonia concentration distribution prediction model for feature encoding to obtain second parameter data; the process of the normalization processing can refer to step 1032;

[0139] Step 123, performing feature encoding on the first data according to B j =f encoder (A j ,θ e )to obtain a training encoded feature vector,

[0140] where B j is the training encoded feature vector, f encoder is the forward propagation function of the joint encoding layer, A j is the second parameter data, and θ e is the parameter of the joint encoding layer.

[0141] In step 13, input the encoded feature vector into the decoding network layer of the ammonia concentration distribution prediction model for vector mapping to obtain a three-dimensional ammonia concentration distribution field vector, including:

[0142] Step 131: According to C j = f decoder (B j , θ d ) perform vector mapping on the training encoded feature vector to obtain a training ammonia concentration distribution field vector,

[0143] where C j is the training ammonia concentration distribution field vector, f decoder is the forward propagation function of the decoding network layer, B j is the training encoded feature vector, and θ d is the decoding network layer parameter.

[0144] In this example, real-time sensor data (such as pressure, temperature) may be affected by electromagnetic interference or equipment failures. Ensure the reliability of the input data through missing value filling and outlier detection.

[0145] Adopt the same normalization formula as in the training stage to ensure that real-time data and historical training data are in the same feature space, avoiding prediction deviation caused by scale differences.

[0146] Directly use the optimized joint encoding layer parameters in the training stage to retain the risk feature patterns learned by the model from historical data.

[0147] The decoding network layer parameters have learned the physical laws of ammonia diffusion (such as attenuation with distance, deviation with wind direction) during the training stage, enabling approximate simulation.

[0148] The joint encoding layer - decoding network layer structure can be optimized to run on edge devices (such as ship PLC controllers), with a delay from data acquisition to concentration field output < 100 ms, meeting the fast response requirements for emergency leakage scenarios.

[0149] The model can automatically capture complex interactions between multiple parameters (such as the superposition effect of "pipe vibration + pressure fluctuation + temperature rise"), avoiding the single-factor limitations of traditional threshold alarm systems.

[0150] In an optional embodiment of the present invention, in step 14, according to the ammonia concentration distribution field vector, obtain an ammonia fuel leakage risk assessment result, including:

[0151] Step 141: Compare the ammonia concentration distribution field vector with a first set ammonia concentration threshold and a second set ammonia concentration threshold.

[0152] If , obtain a high-risk assessment result for ammonia fuel leakage;

[0153] If , obtain the low-risk assessment result of ammonia fuel leakage;

[0154] If , obtain the medium-risk assessment result of ammonia fuel leakage;

[0155] wherein, C j is the training ammonia concentration distribution field vector, T1 is the first set ammonia concentration threshold, and T2 is the second set ammonia concentration threshold.

[0156] In this example, through the first set threshold and the second set threshold, the continuous ammonia concentration distribution is divided into discrete risk levels to avoid fuzzy judgment. The three-dimensional concentration field vector can match thresholds for each position of the ship (such as the fuel tank and the crew living area) respectively.

[0157] Each risk assessment result (such as time, location, concentration, risk level) is stored in the knowledge base to form a complete risk log. By analyzing historical data, potential laws can be discovered to guide the formulation of preventive maintenance plans.

[0158] Such as Figure 2 shown, the ammonia fuel leakage risk assessment knowledge base system consists of an application layer, a function layer, a service layer, and a database layer. Among them, the application layer provides a direct interaction interface for users, supports the core function management of the knowledge base system, including user permission control, knowledge management, model management, rule management, knowledge retrieval, and knowledge audit and other modules; the function layer is the specific function module designed according to the business logic, supports the knowledge definition, classification, label management and other functions of the knowledge base, and at the same time includes the management modules of geometric model library, simulation parameter library, simulation sample library, proxy model library, file library, mathematical formula library, and rule library, etc. The function layer is also responsible for basic logic operations such as knowledge retrieval and reasoning, and provides specific knowledge operation function support for the application layer; the service layer provides underlying service support to ensure the implementation of the business logic of the function layer, mainly including data persistence service, file management service, knowledge extraction service, proxy model training-inference service engine, and knowledge base hybrid inference service engine and other modules; the database layer is the data storage and management center of the system, including database resources such as MySQL and MinIO, used to store structured data (such as user information, knowledge classification) and unstructured data (such as model files, guide documents, etc.), ensure the security, integrity and efficient retrieval of data, and at the same time support operations such as data backup and recovery.

[0159] Through the knowledge base system designed by the above structure, it can provide fast and accurate risk assessment and decision support for the design and operation of ammonia fuel-powered container ships.

[0160] Figure 3Shows the construction of the risk assessment knowledge base and the risk assessment process: First, the user defines knowledge classification and knowledge tags through the above knowledge base system platform, and then uploads relevant multi-source data to the knowledge base system, including geometric models, simulation parameter data, simulation sample data, guide specification documents, etc. Based on the multi-source data fusion technology, unified data management is carried out, and structured and unstructured data are stored in the database to support efficient knowledge retrieval and knowledge auditing. For document data, the invention provides intelligent knowledge extraction technology to further form the system's knowledge and rule base, including logical rules and mathematical formula rules, etc. Using the simulation database of the knowledge base system, the user can, based on the high-precision surrogate model training-inference technology provided by the invention, train the surrogate model and perform fast inference based on the surrogate model. Combining the intelligent hybrid inference technology, the knowledge base system will quickly evaluate the risk status and give response suggestions based on the dynamic fusion mechanism of rule-based reasoning and case-based reasoning.

[0161] Figure 4 Shows the multi-modal knowledge fusion technology framework. The multi-source data involved in the risk assessment process of ammonia fuel-powered container ships includes geometric model libraries, simulation sample libraries, simulation parameter libraries, document libraries, mathematical formula libraries, rule libraries, and surrogate model libraries, etc. The simulation parameter library stored in the knowledge base mainly includes ammonia fuel leakage scenario parameters, operating condition parameters, etc. These parameters will be associated with the hull model files in the geometric model library and the numerical simulation sample data under specific ship types and specific operating condition scenarios. Using the above geometric model library, simulation parameter library, and simulation sample library, the user can establish a surrogate model, train a quantitative result prediction model for ammonia fuel leakage under different simulation parameter conditions, and form a surrogate model library in the system. The user's document data, including specifications, guides, risk assessment reports, etc., will use intelligent knowledge extraction technology in the knowledge base to form the mathematical formula library and rule library in the knowledge base. At the same time, the user can also directly define relevant knowledge through the parameter relationships in the simulation parameter library and inject it into the formula library and rule library. The association relationships between the above various types of data will be maintained in the form of knowledge classification, knowledge tags, etc., and stored in the structured database (MySQL) in the form of metadata, while the geometric models, simulation samples, and files such as guides and reports associated with the metadata are stored in the unstructured database (MinIO) in the form of unstructured data. In the knowledge retrieval or knowledge auditing stage, the user's retrieval or auditing operations in the multi-source knowledge fusion system will perform fuzzy searches through information such as tags and classifications in the metadata, locate the structured data, and associate the relevant knowledge in the unstructured database, quickly return the relevant metadata and original file information and display it to the user, realizing the dynamic retrieval and intelligent recommendation of cross-modal data.

[0162] Such as Figure 5As shown in the figure, the intelligent knowledge extraction process efficiently extracts knowledge rules from user document data (including specifications, guidelines, risk assessment reports, etc.). First, based on an improved semantic-enhanced document parsing method, semantic-preserving document paging or chunking is performed on document content such as specifications and guidelines, followed by text extraction and word segmentation of the text. For the parsed document data, through custom keywords such as "ammonia fuel", "leakage", "PPM", etc., combined with semantic logic, knowledge extraction is performed on the unstructured document data and automatically converted into rules and formulas in the knowledge base.

[0163] Figure 6 It shows the parsing and knowledge extraction results of a specification document for an ammonia fuel-powered container ship, and quickly corresponds and displays the extraction results and the document content.

[0164] As Figure 7 shown, the agent model training-inference flow chart is based on the simulation sample data stored in the knowledge base and the corresponding simulation parameters. An agent model for the ammonia concentration distribution in the ammonia fuel leakage scenario is trained and used for rapid prediction. First, data extraction and preprocessing are performed on the simulation data to eliminate damaged data in storage and abnormal or unreasonable data points in the flow field data; then operations such as standardization and normalization are performed on the data to improve the efficiency and stability of training; the processed data set is divided into a training set, a validation set, and a test set for model training, tuning, and evaluation; subsequently, model parameter optimization is performed, that is, an agent model with prediction ability can be obtained. When new test parameters are input, the agent model can quickly predict the ammonia concentration distribution after leakage under this condition.

[0165] Figure 8 It shows the design of the encoder-decoder network architecture used in the agent model construction process. First, the input multi-source parameters (leakage location, leakage intensity, environmental wind speed, etc.) are jointly encoded through the encoder network to generate a latent vector containing spatio-temporal features, and then the latent vector is mapped to a three-dimensional ammonia concentration field through the decoder network to complete efficient agent modeling.

[0166] As Figure 9 shown, the intelligent hybrid inference technology process is based on the dynamic fusion mechanism of rule-based reasoning and case-based reasoning to quickly evaluate the risk situation and give responses. For the working conditions to be evaluated, the knowledge base system will call the matching agent model to quickly predict the quantitative ammonia concentration distribution results and perform quick rule matching in the rule base. The hybrid inference engine realizes "rule-data" dual-driven reasoning based on the prediction results of the agent model and the knowledge rules and formula rules in the rule base. Through the knowledge base system, the prediction results are post-processed and displayed, and a risk assessment report is quickly generated, giving emergency measures to assist decision-making responses in the ship type design and actual operation stages.

[0167] The present invention integrates multi-source data such as ship design drawings (such as fuel tank layout), operation records (pressure decoding network layer / decode network layer vibration data), experimental reports (leakage diffusion cloud map), and rule manuals (international specifications), covering all influencing factors of leakage risk.

[0168] Through data cleaning (missing value filling, outlier correction) and normalization, noise and dimension differences are eliminated to ensure the quality of model input.

[0169] The joint encoding layer automatically extracts the correlation weights of multi-modal features (numerical decoding network layer + decode network layer spatial type), generates low-dimensional dense vectors, and the decode network layer restores the three-dimensional concentration field, realizing the cross-domain mapping of the decode network layer from "data space → physical space" in the decode network layer to accurately locate high-risk areas.

[0170] The data is classified and stored in seven sub-libraries such as the file library and rule library, and an organic whole is formed through label and semantic association, supporting the standardized management and cross-project reuse of knowledge for similar risk assessments.

[0171] The semantic enhancement parsing technology extracts implicit rules in the document (such as safety thresholds in the specifications), defines knowledge in combination with simulation parameter relationships, and injects it into the rule library and formula library to improve the accuracy and practicality of the knowledge base.

[0172] Duplicate data and conflicting rules are eliminated to ensure the logical reliability of the knowledge base and avoid evaluation ambiguity.

[0173] The sensor real-time acquisition data drives the model, with a decoding network layer delay < 100ms, supporting the closed-loop of the decoding network layer for "dynamic monitoring decoding network layer - decoding network layer instant evaluation" in the decoding network layer; the model is optimized through global error and can adapt to un-trained scenarios (such as extrapolating from high temperature to low temperature conditions).

[0174] The three-dimensional concentration field is combined with double thresholds (high decoding network layer / middle decoding network layer / low risk in the decoding network layer) to achieve risk quantification and grading; the evaluation results are stored in the log, supporting historical pattern analysis and preventive maintenance.

[0175] The geometric model, simulation data, and document rules are associated through metadata, supporting cross-modal retrieval; the proxy model (encoding decoding network layer - decoding network layer decoding architecture) combines rule reasoning and case reasoning to achieve the double-driven evaluation of the decoding network layer from "data decoding network layer - decoding network layer rules" in the decoding network layer.

[0176] The model can be deployed on edge devices (such as decoding network layer PLC decoding network layer controllers) to meet the real-time monitoring needs of ships; the open architecture supports the addition of sensor data or specification files to adapt to technological upgrades.

[0177] Such asFigure 10 As shown in Figure 10 , an embodiment of the present invention further provides an ammonia fuel-powered container ship leakage risk assessment device 20, including:

[0178] An acquisition module 21, configured to acquire multi-source parameter data of an ammonia fuel-powered container ship; acquire multi-source training data of the ammonia fuel-powered container ship;

[0179] A processing module 22, configured to input the multi-source parameter data into the joint encoding layer of the ammonia concentration distribution prediction model for feature encoding to obtain an encoded feature vector; input the encoded feature vector into the decoding network layer of the ammonia concentration distribution prediction model for vector mapping to obtain a three-dimensional ammonia concentration distribution field vector; obtain an ammonia fuel leakage risk assessment result according to the ammonia concentration distribution field vector; establish an ammonia fuel leakage risk knowledge base according to the multi-source training data; input the sample data in the knowledge base into the joint encoding layer of a preset network model for feature encoding to obtain a training encoded feature vector; input the training encoded feature vector into the decoding network layer of the preset network model for vector mapping to obtain a training ammonia concentration distribution field vector, and obtain an ammonia concentration distribution prediction model.

[0180] Optionally, establishing an ammonia fuel leakage risk knowledge base according to the multi-source training data includes:

[0181] Performing data cleaning on the multi-source training data to obtain sample data;

[0182] Managing and storing the sample data in the form of knowledge classification and knowledge labels to obtain a first knowledge base;

[0183] Auditing the first knowledge base to eliminate duplicate data and conflicting rules to obtain a second knowledge base;

[0184] Performing semantic enhancement parsing on the documents in the second knowledge base, and extracting rules and formulas in the documents after semantic enhancement parsing;

[0185] Filling the extracted rules and formulas into the rule base and mathematical formula base of the second knowledge base;

[0186] Defining knowledge according to the parameter relationship in the simulation parameter library of the second knowledge base;

[0187] Injecting the knowledge into the mathematical formula base and rule base of the second knowledge base to obtain an ammonia fuel leakage risk knowledge base.

[0188] Optionally, inputting the sample data in the knowledge base into the joint encoding layer of a preset network model for feature encoding to obtain a training encoded feature vector includes:

[0189] Determine the operation record data x of each position of the ammonia fuel-powered container ship according to the sample data in the knowledge base j [i] and the corresponding ammonia concentration C j [i],

[0190] where x j [i] is the operation record data of the j-th position of the ammonia fuel-powered container ship in the i-th scenario, i = 1, 2, 3,..., M, M is the total number of scenarios, j = 1, 2, 3,..., N, N is the total number of positions, C j [i] is the ammonia concentration at the j-th position of the ammonia fuel-powered container ship in the i-th scenario;

[0191] Normalize the operation record data x j [i], input it into the joint encoding layer of the preset network model for feature encoding, and obtain the first training data X j [i];

[0192] According to Z j [i]=f encoder (X j [i], θ enc )Perform feature encoding on the first training data to obtain the training encoded feature vector,

[0193] where Z j [i] is the training encoded feature vector, f encoder is the forward propagation function of the joint encoding layer, θ enc is the preset joint encoding layer parameter of the joint encoding layer.

[0194] Optionally, input the training encoded feature vector into the decoding network layer of the preset network model for vector mapping to obtain the training ammonia concentration distribution field vector, and obtain the ammonia concentration distribution prediction model, including:

[0195] According to Y j [i]=f decoder (Z j [i], θ dec )Perform vector mapping on the training encoded feature vector to obtain the training ammonia concentration distribution field vector,

[0196] where Y j [i] is the training ammonia concentration distribution field vector, f decoder is the forward propagation function of the decoding network layer, θ dec is the preset decoding network layer parameter of the joint encoding layer;

[0197] According to , obtain the first average difference,

[0198] where ki is the first average difference for the i-th scenario, Y j [i] is the training ammonia concentration distribution field vector, C j [i] is the ammonia concentration at the j-th position of the ammonia fuel-powered container ship in the i-th scenario;

[0199] According to , the first average mean value is obtained,

[0200] where h is the first average mean value under various scenarios, k i is the first average difference for the i-th scenario;

[0201] Adjust the preset joint coding layer parameters and preset decoding network layer parameters, and repeat the above process for a set number of times to obtain the first average mean value h for each time t , t = 1, 2, 3,..., R, where R is the set number of times;

[0202] According to ,

[0203] the joint coding layer parameters θ e and the decoding network layer parameters θ d corresponding to the minimum first average mean value are obtained;

[0204] Take the joint coding layer parameters θ e as the preset joint coding layer parameters of the joint coding layer and take the decoding network layer parameters θ d as the preset decoding network layer parameters of the joint coding layer to obtain an embedded ammonia concentration distribution prediction model.

[0205] Optionally, input the multi-source parameter data into the joint coding layer of the ammonia concentration distribution prediction model for feature encoding to obtain an encoded feature vector, including:

[0206] Clean the multi-source parameter data to obtain first parameter data;

[0207] Normalize the first parameter data, input it into the joint coding layer of the ammonia concentration distribution prediction model for feature encoding to obtain second parameter data;

[0208] According to B j = f encoder (A j , θ e ) perform feature encoding on the first data to obtain a training encoded feature vector,

[0209] where B j is the training encoded feature vector, f encoder is the forward propagation function of the joint coding layer, A j is the second parameter data, θe For the combined encoding layer parameters.

[0210] Optionally, input the encoded feature vector into the decoding network layer of the ammonia concentration distribution prediction model for vector mapping to obtain a three-dimensional ammonia concentration distribution field vector, including:

[0211] According to C j =f decoder (B j , θ d ) perform vector mapping on the training encoded feature vector to obtain a training ammonia concentration distribution field vector,

[0212] where C j is the training ammonia concentration distribution field vector, f decoder is the forward propagation function of the decoding network layer, B j is the training encoded feature vector, and θ d is the combined encoding layer parameters.

[0213] Optionally, obtain an ammonia fuel leakage risk assessment result based on the ammonia concentration distribution field vector, including:

[0214] Compare the ammonia concentration distribution field vector with a first set ammonia concentration threshold and a second set ammonia concentration threshold,

[0215] If , obtain a high risk assessment result for ammonia fuel leakage;

[0216] If , obtain a low risk assessment result for ammonia fuel leakage;

[0217] If , obtain a medium risk assessment result for ammonia fuel leakage;

[0218] where C j is the training ammonia concentration distribution field vector, T1 is the first set ammonia concentration threshold, and T2 is the second set ammonia concentration threshold.

[0219] It should be noted that this device corresponds to the above method, and all implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0220] An embodiment of the present invention also provides a computing device, including:

[0221] One or more processors;

[0222] A storage device for storing one or more programs, which when executed by the one or more processors, cause the one or more processors to implement the method as described above.

[0223] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when run on a computing device, cause the computing device to perform the method described above. All implementation manners in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0224] Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed in the present invention can be implemented by electronic hardware, or a combination of computing device software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0225] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated here.

[0226] In the embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division, and there can be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces, and the indirect couplings or communication connections of the devices or units can be in an electrical, mechanical, or other form.

[0227] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0228] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit.

[0229] When the above-mentioned functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a storage medium readable by a computing device. Based on such understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. The software product of this computing device is stored in a storage medium and includes several instructions to enable a computing device (which can be a personal computing device, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, ROM, RAM, magnetic disks, or optical discs that can store program codes.

[0230] In addition, it should be noted that in the devices and methods of the present invention, obviously, each component or each step can be decomposed and / or recombined. These decompositions and / or recombinations should be regarded as equivalent solutions of the present invention. And, the steps of performing the above series of processes can naturally be executed in chronological order according to the described order, but it is not necessary to execute them in chronological order. Some steps can be executed in parallel or independently of each other. For those of ordinary skill in the art, it is understandable that all or any steps or components of the methods and devices of the present invention can be implemented in any computing device (including processors, storage media, etc.) or in a network of computing devices in the form of hardware, firmware, software, or a combination thereof, which can be achieved by those of ordinary skill in the art using basic programming skills after reading the description of the present invention.

[0231] Therefore, the object of the present invention can also be achieved by running a program or a set of programs on any computing device. The computing device can be a well-known general-purpose device. Therefore, the object of the present invention can also be achieved only by providing a program product containing program codes for implementing the method or device. That is to say, such a program product also constitutes the present invention, and the storage medium storing such a program product also constitutes the present invention. Obviously, the storage medium can be any well-known storage medium or any storage medium developed in the future. It should also be noted that in the devices and methods of the present invention, obviously, each component or each step can be decomposed and / or recombined. These decompositions and / or recombinations should be regarded as equivalent solutions of the present invention. And, the steps of performing the above series of processes can naturally be executed in chronological order according to the described order, but it is not necessary to execute them in chronological order. Some steps can be executed in parallel or independently of each other.

[0232] The above is the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle described in the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A method for assessing the leakage risk of an ammonia-fueled container ship, characterized in that, Including: Obtaining multi-source parameter data of an ammonia fuel-powered container ship; Inputting the multi-source parameter data into the joint encoding layer of an ammonia concentration distribution prediction model for feature encoding to obtain an encoded feature vector; Inputting the encoded feature vector into the decoding network layer of the ammonia concentration distribution prediction model for vector mapping to obtain a three-dimensional ammonia concentration distribution field vector; Obtaining an ammonia fuel leakage risk assessment result according to the ammonia concentration distribution field vector; wherein, the training process of the ammonia concentration distribution prediction model includes: obtaining multi-source training data of an ammonia fuel-powered container ship; establishing an ammonia fuel leakage risk knowledge base according to the multi-source training data; inputting the sample data in the knowledge base into the joint encoding layer of a preset network model for feature encoding to obtain a training encoded feature vector; inputting the training encoded feature vector into the decoding network layer of the preset network model for vector mapping to obtain a training ammonia concentration distribution field vector, and obtaining an ammonia concentration distribution prediction model.

2. The ammonia fuel-powered container ship leakage risk assessment method according to claim 1, wherein, Establishing an ammonia fuel leakage risk knowledge base according to the multi-source training data, including: Performing data cleaning on the multi-source training data to obtain sample data; Managing and storing the sample data in the form of knowledge classification and knowledge labels to obtain a first knowledge base; Auditing the first knowledge base to remove duplicate data and conflicting rules to obtain a second knowledge base; Performing semantic enhancement parsing on the documents in the second knowledge base, and extracting the rules and formulas in the documents after semantic enhancement parsing; Filling the extracted rules and formulas into the rule base and mathematical formula library of the second knowledge base; Defining knowledge according to the parameter relationship in the simulation parameter library of the second knowledge base; Injecting the knowledge into the mathematical formula library and rule base of the second knowledge base to obtain an ammonia fuel leakage risk knowledge base.

3. The ammonia fuel-powered container ship leakage risk assessment method according to claim 1, characterized in that Inputting the sample data in the knowledge base into the joint encoding layer of a preset network model for feature encoding to obtain a training encoded feature vector, including: Determine the operation record data x of each position of the ammonia fuel-powered container ship according to the sample data in the knowledge base j [i] and the corresponding ammonia concentration C j [i], Among them, x j [i] is the operation record data of the j-th position of the ammonia fuel-powered container ship in the i-th scenario, where i = 1, 2, 3,..., M, M is the total number of scenarios, and j = 1, 2, 3,..., N, N is the total number of positions, C j [i] is the ammonia concentration at the j-th position of the ammonia fuel-powered container ship in the i-th scenario; For the operation record data x j [i] perform normalization processing, input it into the joint encoding layer of the preset network model for feature encoding, and obtain the first training data X j [i]; According to Z j [i]=f encoder (X j [i], θ enc ), perform feature encoding on the first training data to obtain a training encoded feature vector. Among them, Z j [i] is the training encoded feature vector, f encoder is the forward propagation function of the joint encoding layer, θ enc is the preset joint encoding layer parameter of the joint encoding layer.

4. The ammonia fuel-powered container ship leakage risk assessment method according to claim 3, wherein Inputting the training encoded feature vector into the decoding network layer of the preset network model for vector mapping to obtain a training ammonia concentration distribution field vector, and obtaining an ammonia concentration distribution prediction model, including: According to Y j [i] = f decoder (Z j [i], θ dec )to perform vector mapping on the training encoded feature vectors to obtain the training ammonia concentration distribution field vectors Among them, Y j [i] is the training ammonia concentration distribution field vector, f decoder is the forward propagation function of the decoding network layer, θ dec is the preset decoding network layer parameter of the joint encoding layer; According to , a first average difference is obtained. where k i is the first average difference of the i-th scenario, and Y j [i] is the training ammonia concentration distribution field vector, and C j [i] is the ammonia concentration at the j-th position of the ammonia fuel-powered container ship in the i-th scenario; According to , the first average mean value is obtained. Among them, h is the first average mean in various scenarios, and k i is the first average difference in the i-th scenario; Adjust the parameters of the preset joint coding layer and the parameters of the preset decoding network layer, and repeat the above process for the set number of times to obtain the first average mean h for each time t , where t = 1, 2, 3, ..., R, and R is the set number of times; According to , Obtain the smallest first average mean and the corresponding joint coding layer parameter θ e and the decoding network layer parameter θ d ; Take the combined coding layer parameter θ e as the preset combined coding layer parameter of the combined coding layer and take the decoding network layer parameter θ d as the preset decoding network layer parameter of the combined coding layer, to obtain an embedded ammonia concentration distribution prediction model.

5. The ammonia fuel-powered container ship leakage risk assessment method according to claim 1, wherein Inputting the multi-source parameter data into the joint encoding layer of the ammonia concentration distribution prediction model for feature encoding to obtain an encoded feature vector, including: Performing data cleaning on the multi-source parameter data to obtain first parameter data; Performing normalization processing on the first parameter data, and inputting it into the joint encoding layer of the ammonia concentration distribution prediction model for feature encoding to obtain second parameter data; According to B j =f encoder (A j , θ e ) perform feature encoding on the first data to obtain a training encoded feature vector Among them, B j is the training encoded feature vector, f encoder is the forward propagation function of the joint encoding layer, A j is the second parameter data, θ e is the parameter of the joint encoding layer.

6. The ammonia fuel-powered container ship leakage risk assessment method according to claim 5, wherein, Inputting the encoded feature vector into the decoding network layer of the ammonia concentration distribution prediction model for vector mapping to obtain a three-dimensional ammonia concentration distribution field vector, including: According to C j =f decoder (B j ,θ d ), perform vector mapping on the training encoded feature vectors to obtain the training ammonia concentration distribution field vectors. Among them, C j is the training ammonia concentration distribution field vector, f decoder is the forward propagation function of the decoding network layer, B j is the training encoded feature vector, θ d is the parameter of the decoding network layer.

7. The ammonia fuel-powered container ship leakage risk assessment method according to claim 1, characterized in that Obtaining an ammonia fuel leakage risk assessment result according to the ammonia concentration distribution field vector, including: Comparing the ammonia concentration distribution field vector with a first set ammonia concentration threshold and a second set ammonia concentration threshold; If , obtain the high-risk assessment result of ammonia fuel leakage; If , obtain the low-risk assessment result of ammonia fuel leakage; If , obtain the medium risk assessment result of ammonia fuel leakage; Among them, C j is the training ammonia concentration distribution field vector, T1 is the first set ammonia concentration threshold, and T2 is the second set ammonia concentration threshold.

8. An ammonia fuel-powered container ship leakage risk assessment device, characterized in that, Including: An acquisition module for obtaining multi-source parameter data of an ammonia fuel-powered container ship; Obtaining multi-source training data of an ammonia fuel-powered container ship; A processing module, configured to input the multi-source parameter data into a joint encoding layer of an ammonia concentration distribution prediction model for feature encoding to obtain an encoded feature vector; input the encoded feature vector into a decoding network layer of the ammonia concentration distribution prediction model for vector mapping to obtain a three-dimensional ammonia concentration distribution field vector; obtain an ammonia fuel leakage risk assessment result according to the ammonia concentration distribution field vector; establish an ammonia fuel leakage risk knowledge base according to the multi-source training data; input sample data in the knowledge base into a joint encoding layer of a preset network model for feature encoding to obtain a training encoded feature vector; input the training encoded feature vector into a decoding network layer of the preset network model for vector mapping to obtain a training ammonia concentration distribution field vector, and obtain an ammonia concentration distribution prediction model.

9. A computing device, characterized in that, Comprising: One or more processors; A storage device for storing one or more programs, which when executed by the one or more processors cause the one or more processors to implement the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, A program is stored in the computer-readable storage medium, and when the program is executed by a processor, the method according to any one of claims 1-7 is implemented.

Images