Ammonia refrigeration system leakage rate prediction method based on physical model driving and multi-modal data fusion

Abstract

The invention relates to an ammonia refrigeration system leakage rate prediction method based on physical model driving and multi-modal data fusion, and belongs to the technical field of refrigeration system safety monitoring. The method comprises the steps that firstly, based on mass conservation, energy conservation and isentropic blocking equations, ammonia liquid vaporization latent heat phase change is coupled; establishing a leakage physical model and generating a multi-working-condition simulation data set; then, wavelet multi-scale feature extraction is carried out on the pressure time sequence data, and the wavelet multi-scale feature extraction is fused with physical parameters to construct a multi-modal feature vector; and finally, carrying out cooperative training by adopting a support vector machine (PSO-SVM) optimized by a particle swarm algorithm and a Transform model, fusing prediction results of the PSO-SVM and the Transform model through a dynamic weight distribution mechanism, and constructing a leakage rate prediction model. According to the method, the reliability of a physical mechanism and the adaptive capacity of a data driving model are both achieved, rapid and high-precision prediction of the ammonia leakage rate is achieved, and the safety monitoring and emergency response level of a refrigeration system is effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of refrigeration system safety monitoring technology, specifically involving a method for predicting the leakage rate of an ammonia refrigeration system based on physical model-driven and multimodal data fusion. Background Technology

[0002] Ammonia, as a highly efficient and low-cost refrigerant, is widely used in industrial refrigeration systems. However, ammonia is highly toxic, flammable, and explosive, and leaks can pose serious threats to personnel safety and the environment. Therefore, rapidly and accurately predicting ammonia leak rates is crucial for timely warnings, developing emergency response measures, and minimizing accident losses.

[0003] Currently, the prediction of ammonia leakage rates mainly relies on traditional fluid dynamics models, such as orifice leakage models based on Bernoulli's equations, or numerical simulations using computational fluid dynamics (CFD). While these methods have clear physical meaning, they typically require establishing complex governing equations and performing numerous numerical iterations, resulting in lengthy computations and making it difficult to meet the "real-time, rapid" response requirements in emergency response. Furthermore, traditional methods are mostly based on single physical models, limiting their ability to capture the dynamic changes of multiple coupled factors during actual leakage processes, leading to insufficient prediction accuracy and limited adaptability.

[0004] With the development of artificial intelligence technology, machine learning methods have shown advantages in the field of industrial system state prediction. However, simple data-driven models rely on a large amount of high-quality real data and lack physical mechanism constraints. When data is scarce or the operating conditions are beyond the training range, their extrapolation prediction ability and interpretability are often poor.

[0005] Therefore, how to combine the mechanistic reliability of physical models with the flexible learning ability of data-driven models to develop an ammonia leakage rate prediction method that combines fast response and high-precision prediction has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for predicting the leakage rate of ammonia refrigeration systems based on physical model-driven and multimodal data fusion. This method achieves rapid and accurate prediction of the leakage rate of ammonia refrigeration systems by constructing a hybrid model framework that combines physical mechanisms and artificial intelligence.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a method for predicting the leakage rate of an ammonia refrigeration system based on physical model-driven and multimodal data fusion, comprising the following steps:

[0008] S1. Constructing a physical model: Based on the mass conservation equation, energy conservation equation, and isentropic blockage equation, and coupled with the latent heat phase change process of ammonia liquid vaporization, a physical model for predicting the leakage rate of an ammonia refrigeration system is established. The physical model includes the parameters and computational domain of the high-pressure ammonia storage tank and downstream pipeline.

[0009] S2. Generate a feature dataset: Based on the physical model, multiple sets of ammonia leakage data containing different operating parameters are obtained through calculation to estimate and determine the key factors affecting the leakage rate, forming a training dataset;

[0010] S3. Collaborative Machine Learning Modeling: Using a particle swarm optimization-optimized support vector machine (PSO-SVM) model and a Transformer model, the dataset generated in step S2 is collaboratively trained and learned to construct a fusion prediction model for the leakage rate of the ammonia refrigeration system.

[0011] S4. Model Validation and Correction: The prediction model obtained in step S3 is used to perform parameter tuning, generalization ability testing and experimental validation using the validation set and test set. The model is then corrected based on the comparison between the experimental results and the simulation data to obtain the final practical prediction model.

[0012] Further, in step S1, the mass conservation equation is:

[0013]

[0014] in, This indicates the total mass flow rate of the leak. , These represent the densities of the liquid phase and the gas phase, respectively. , These represent the volumetric flow rates of the liquid phase and the gas phase, respectively.

[0015] A blocking effect correction factor ζ is introduced to characterize the local flow resistance of the leakage orifice, and its expression is:

[0016]

[0017] in, , These represent the liquid phase and gas phase flow rates, respectively.

[0018] Further, in step S1, the energy conservation equation is:

[0019]

[0020] in, , These represent the phase enthalpy of the liquid phase and the phase enthalpy of the gas phase, respectively. Indicates the enthalpy value at the leak source inlet. Indicates the latent heat of phase transition;

[0021] Among them, latent heat of phase transition Based on real-time temperature and pressure dynamic calculations, its expression is:

[0022]

[0023] in, Indicates saturation pressure Lower temperature The corresponding latent heat of vaporization, This represents the specific heat capacity at constant pressure. Indicates volume.

[0024] Further, in step S1, the isentropic blocking equation is:

[0025]

[0026] in, This indicates the stagnant pressure upstream of the leak hole. This indicates the stagnation density upstream of the leak orifice. This indicates the specific heat capacity of ammonia. Indicates the leakage flow rate;

[0027] Mass flow rate is solved using an iterative method. :

[0028]

[0029] in, Represents the flow coefficient. Indicates the area of ​​the leak hole (㎡).

[0030] Furthermore, in step S2, the range of operating parameters covered when generating the dataset includes: leakage orifice diameter 1-10mm, system pressure 0.1-2.45 MPa, and system temperature 253-423 K.

[0031] Furthermore, step S3 specifically includes the following sub-steps:

[0032] S31. Data preprocessing: Generate a dataset containing time-series pressure and temperature data based on the physical model, and inject Gaussian noise to simulate sensor error;

[0033] S32. Multi-scale feature extraction: Wavelet transform is used to decompose the pressure time series data into multiple layers, extract high-frequency and low-frequency coefficients as features, and calculate the time-domain statistical features of leakage rate;

[0034] S33. Multimodal feature fusion: fusing physical parameter features with wavelet features to form a feature vector X;

[0035] S34. Collaborative Model Training and Weight Allocation: The PSO-SVM model and the Transformer model are trained separately, and the fusion weight α is calculated through dynamic weight allocation;

[0036] S35. Fusion Prediction: The final predicted leakage rate is calculated using the following formula:

[0037]

[0038] in, This represents the predicted value of the PSO-SVM model. This represents the predicted value from the Transformer model.

[0039] Furthermore, the feature vector X is:

[0040]

[0041] in, Indicates average pressure. Indicates the total pressure difference. Indicates leakage flow rate. This indicates a specific temperature parameter for a particular scenario. Let i represent the i-th high-frequency coefficient after wavelet decomposition. Represents the flow coefficient. Indicates the area of ​​the leak. Indicates specific heat ratio. Indicates the density of the liquid. This indicates the density of the gas.

[0042] Furthermore, the method for calculating the fusion weight α through dynamic weight allocation is as follows:

[0043]

[0044] in, Indicates the PSO-SVM weights. =10.

[0045] The present invention provides a computer device, comprising: at least one processor, at least one memory, and computer program instructions stored in the memory, wherein the above-described method is implemented when the computer program instructions are executed by the processor.

[0046] The present invention provides a computer-readable storage medium having computer program instructions stored thereon, which implement the above-described method when the computer program instructions are executed by a processor.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. Mechanism and Data Fusion: This invention generates a high-quality dataset covering a wide range of working conditions through a physical model, which makes up for the scarcity of real data. At the same time, it provides physical mechanism constraints for machine learning models, enhancing the generalization ability and interpretability of the models.

[0049] 2. Multimodal feature fusion: This invention combines physical parameters, time-domain statistical features, and time-frequency wavelet features to comprehensively capture multi-dimensional and multi-scale dynamic information in the leakage process, thereby improving feature representation capabilities.

[0050] 3. Intelligent Collaborative Prediction: This invention innovatively adopts PSO-SVM and Transformer collaborative modeling and designs a dynamic weight allocation mechanism, which fully leverages the classification advantages of SVM under small samples and high-dimensional features and the advantages of Transformer in capturing long-term temporal dependencies. The prediction accuracy (R2, MSE, MAE) is better than that of a single model.

[0051] 4. Rapid Response and High Precision: The collaborative model trained by this invention can achieve rapid prediction of leakage rate while maintaining high precision, meeting the needs of real-time safety monitoring and emergency decision-making in industrial sites. Attached Figure Description

[0052] Figure 1 This is a flowchart of the leakage rate prediction method for ammonia refrigeration systems based on physical model-driven and multimodal data fusion provided in this embodiment of the invention;

[0053] Figure 2 This is a diagram of the collaborative architecture between PSO-SVM and Transformer in an embodiment of the present invention;

[0054] Figure 3 This is a block diagram for calculating the leakage rate in an embodiment of the present invention. Detailed Implementation

[0055] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0056] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0057] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0058] like Figure 1-3 As shown, this embodiment provides a method for predicting the leakage rate of an ammonia refrigeration system based on physical model-driven and multimodal data fusion, including the following steps:

[0059] Step S1: Based on the mass conservation equation, energy conservation equation and isentropic blocking equation, explicitly couple the latent heat phase change process of ammonia liquid vaporization to establish a physical model of the leakage rate of the ammonia refrigeration system.

[0060] Step S2: Generate multiple sets of ammonia leakage data based on the physical model;

[0061] Step S3: The Particle Swarm Optimization (PSO) algorithm-optimized Support Vector Machine (PSO-SVM) and Transformer model are used to perform collaborative learning on a large dataset generated by the physical model; in order to obtain a machine learning-based model for predicting the leakage rate of the ammonia refrigeration system.

[0062] Step S4: Study the adaptability of the machine learning model, revise the model, and finally determine the practical model for predicting the leakage rate of the ammonia refrigeration system.

[0063] Furthermore, step S1 specifically includes:

[0064] (1) The mass conservation equation is:

[0065] (1)

[0066] in, This represents the total mass flow rate of the leak (kg / s). , This indicates the density of the liquid / gas phase (kg / m³). , This indicates the liquid / gas phase volumetric flow rate (m³ / s).

[0067] Among them, a blocking effect correction factor is introduced. Characterizing the local flow resistance of the leakage orifice:

[0068] (2)

[0069] in, , This indicates the liquid / gas phase flow rate (m / s).

[0070] (2) The energy conservation equation is:

[0071] (3)

[0072] in, , This represents the liquid / gas phase relative enthalpy (J / kg). This indicates the enthalpy value (J / kg) at the leak source inlet. This indicates the latent heat of phase transition.

[0073] (3) Dynamic correction of latent heat of phase change: The latent heat of ammonia vaporization is calculated based on real-time temperature and pressure using a lookup table or interpolation method.

[0074] (4)

[0075] in, Indicates saturation pressure Lower temperature The corresponding latent heat of vaporization (J / kg), This represents the specific heat capacity at constant pressure (J / (kg∙K)). It represents specific volume (m³ / kg). Indicates current pressure The corresponding saturation temperature, integral term This indicates the calculation of ammonia from the saturation temperature. Rise to the current temperature The sensible heat to be absorbed, i.e., the heat of non-phase change, is a partial differential term. This represents the extra energy consumed during the phase transition due to the work done by volume expansion, i.e., the volume work correction term.

[0076] (4) Isoentropy blockage equation: The leakage mass flow rate is solved by iterative method. :

[0077] (5)

[0078] in, This represents the flow coefficient (related to the Reynolds number). Indicates the area of ​​the leak hole (㎡). This indicates the specific heat capacity of ammonia. This indicates the stagnation density upstream of the leak orifice. This indicates the stagnant pressure upstream of the leak (the total pressure when the fluid has not expanded). This indicates the dynamic pressure downstream of the leak hole (the local pressure after expansion).

[0079] (5) Phase change dynamic equilibrium constraint: When the pressure downstream of the leakage hole Below saturation pressure At that time, the table lookup method or interpolation method is used to obtain the result. and The saturation property under the formula (4) Dynamic adjustment (gas phase mass fraction) and (Liquid phase mass fraction) and iteratively solve the above system of equations until convergence.

[0080] (a) Mass conservation residual: (6)

[0081] (b) Energy conservation residual: (7)

[0082] (c) Pressure-temperature saturation constraint: (8)

[0083] The physical properties of ammonia, including pressure, temperature, enthalpy, and entropy, were calculated using Tillner-Roth & Friend (1998); the final model output is the leakage mass flow rate. .

[0084] Furthermore, in step S2, generating a large dataset of ammonia leakage characteristics based on this physical model specifically involves:

[0085] 1. Using a full factorial experimental design, a dataset was generated within the following parameter space:

[0086] (1) Leakage orifice diameter: Sampled at 0.5 mm intervals within the range of 1-10 mm (19 discrete values ​​in total). System pressure: Sampled on a logarithmic scale within the range of 0.1-2.45 MPa (ensuring uniform coverage of high and low pressure zones, 200 points in total). Temperature: Sampled linearly within the range of 253-423 K (covering the critical zone of liquid / gas phase transition, 150 points in total). Pressure time series data: Pressure fluctuation data for 60 seconds is generated synchronously during the simulation of each set of static operating points, with the sampling frequency set to 5Hz, and the actual sensor signal is simulated by superimposing white noise.

[0087] (2) For each set of sampling parameters (pressure) taken, ,temperature Add ±5% Gaussian noise to simulate actual sensor measurement error:

[0088] (9)

[0089] in, This indicates that the distribution follows a normal distribution with a mean of 0 and a standard deviation of 0.05. Indicates pressure including noise. This indicates the temperature including noise.

[0090] (3) 57,000 initial operating points (19×200×150) were generated by parameter combination, covering common leakage conditions in industrial scenarios.

[0091] 2. Dynamic simulation and feature extraction of the physical model:

[0092] (1) Dynamic simulation of physical model: input pressure containing noise ,temperature Leakage orifice diameter The leakage mass flow rate is output through the physical model established in step S1. and related parameters (blocking factor) Phase transition latent heat Gas / liquid phase density , ).

[0093] (2) Feature extraction:

[0094] The following features are extracted from each set of simulation results to form a data sample:

[0095] sample (10)

[0096] 3. Multi-scale data augmentation strategy: (1) Property perturbation: Gaussian noise of ±5% (σ=0.05) is superimposed on the pressure and temperature parameters to ensure the consistency of noise distribution. (2) Phase change simulation: In the critical region of saturated pressure Intensive sampling generates an additional 8,000 sets of data. (3) Extreme operating condition extension: Supplement 5,000 sets of data for high flow leakage scenarios with orifice diameter > 8 mm.

[0097] 4. A total of 70,000 data sets were generated, with leakage orifice diameters ranging from 1 to 10 mm, system pressures from 0.1 to 2.45 MPa, temperatures from 253 to 453 K, liquid phase densities from 602.1 to 682.7 kg / m³, gas phase densities from 1.24 to 15.89 kg / m³, blockage effect correction factors from 1.02 to 3.17, latent heat of phase change from 218.5 to 1372.4 kJ / kg, and leakage mass flow rates from 0.001 to 42.6 kg / s. 70% (49,000 data sets) were used as the training set, 15% (10,500 data sets) as the validation set, and 15% (10,500 data sets) as the test set. In an actual ammonia refrigeration system, pressure time-series data P(t) is generated in real time by installing a high-frequency pressure sensor (sampling frequency ≥ 5 Hz) upstream of the leak point. This data is used to characterize the dynamic fluctuation characteristics of the leak process, such as sudden pressure drop, oscillation decay and other transient features.

[0098] Furthermore, in step S3, the Particle Swarm Optimization (PSO)-Optimized Support Vector Machine (PSO-SVM) and Transformer model are used to perform collaborative learning on the large dataset generated by the physical model. Specifically:

[0099] 1. Multimodal data splitting and feature engineering:

[0100] (1) Static feature branch (input PSO-SVM): Extract the physical parameters of a single working point to form a static vector:

[0101] in, , , , This is only used for generating training data for physical models; for actual predictions, only directly measurable parameters need to be input. , , .

[0102] (2) Temporal feature branch (input Transformer): Performs multi-scale decomposition on the pressure fluctuation signal:

[0103] (a) The pressure time series data was decomposed into 5 levels using the Daubechies 4th order wavelet basis function (db4) to extract high-frequency coefficients. (Sensitive to leakage mutations) and low-frequency coefficients (Characterizing the steady-state background); in the original pressure time series data (a2) Define a fixed-length analysis window; based on a 5Hz sampling frequency, cover 15 consecutive data points, set the window length to 3 seconds, and the sliding step size to 0.2 seconds (move 1 data point), generate 286 windows for 60 seconds of time series data;

[0104] (b) Calculate time-domain statistics:

[0105] (b1) Average pressure within the window : (11)

[0106] in, This represents the measured value of the i-th pressure acquisition point within the window;

[0107] (b2) Standard deviation :

[0108] (12)

[0109] Where N represents the total number of valid pressure sampling points contained in the current analysis window;

[0110] (b3) Skewness:

[0111] (13)

[0112] (c) Constructing temporal feature vectors: .

[0113] 2. Heterogeneous model collaboration mechanism:

[0114] (1) Model division of labor: PSO-SVM is used to process static features. Predicting the base leakage rate Its optimization objective is to minimize the predicted leakage rate. With physical model output Mean squared error (MSE); PSO optimization parameters are set as the penalty factor for SVM. and kernel function parameters (RBF kernel selected). PSO configuration set to particle count = 50, iteration count = 100, learning factor c1 = c2 = 1.5, and inertia weight w linearly decreasing from 0.9 to 0.4. Transformer is used to process temporal features. Dynamic correction for predicted leakage rate .

[0115] (2) Dynamic weight fusion strategy: The final predicted leakage rate is:

[0116] (14)

[0117] Among them, weight Dynamically generated by the gating network:

[0118] (a) The input is a concatenated vector of static features and temporal features. .

[0119] (b) Network structure: Fully connected layer (dimension=16, ReLU activation) → Sigmoid output layer (scalar) ∈[0,1]);

[0120] (c) When static features dominate (i.e.) →1), ignoring temporal fluctuations; when leakage mutations are significant ( →0), enhancing dynamic correction.

[0121] 3. Dynamic weight fusion strategy:

[0122] (1) Data partitioning: 49,000 sets of training set were used to train PSO-SVM, Transformer and gating network, 10,500 sets of validation set were used for hyperparameter tuning and early stopping, and 10,500 sets of test set were used for final performance evaluation.

[0123] (2) Phased training: Phase 1: Independent training of PSO-SVM and Transformer. PSO-SVM: Initialization using grid search, PSO optimization. and Transformer: Adam optimizer (learning rate = 0.001, batch size = 64), early stopping threshold = 10 epochs. Stage 2: Jointly train the gating network. Fix the PSO-SVM and Transformer parameters, train only the gating network (loss function = total prediction). (MSE). Phase 3: End-to-end fine-tuning of full model parameters to optimize collaborative prediction capabilities.

[0124] Furthermore, in step S4, the adaptability of the machine learning model is studied, the model is revised, and a practical model is finally determined. Specifically:

[0125] 1. Multi-condition adaptability verification: (1) Pressure sudden change scenario: simulate valve sudden closure (pressure drops from 2.0MPa to 0.5MPa / second), test model in Predicted stability at >1.5MPa / s; (2) Temperature disturbance scenario: ±10K step temperature noise is injected into the phase change critical region (253-280K) to verify the latent heat of phase change. Robustness of computation. (3) Sensor failure scenario: Randomly shield 20% of the pressure time series data, and evaluate the model's dependence on the gating network to automatically degrade to static prediction ( →1) ability;

[0126] 2. Extreme condition generalization test: (1) Over-design parameter test: input orifice diameter 12mm (exceeding the upper limit of 10mm in the training set), pressure 0.05MPa (lower than the lower limit of 0.1MPa in the training set), record the prediction error growth rate <15%; (2) Multiphase flow interference test: inject 5% oil and sludge mixture (density deviation ±8%), and adaptively adjust the compensation error through PSO-SVM kernel function.

[0127] Furthermore, in step S5, the model is revised to finally determine the practical model. Specifically:

[0128] 1. Error compensation mechanism based on adaptive verification:

[0129] (1) Pressure change scenario correction: When the pressure change rate is detected When the pressure is >1.5MPa / s (e.g., when the valve is suddenly closed), the transient compensation function is activated:

[0130] (15)

[0131] in =0.15 is an empirical coefficient. =0.5s is the decay time constant, used to suppress oscillation noise. This indicates the change in pressure. The transient compensation amount represents the leakage rate; the corrected predicted value:

[0132]

[0133] in, This represents the predicted leakage rate of the original fusion model. This represents the corrected final predicted leakage rate.

[0134] (2) Temperature perturbation correction in the critical region of phase transition when the temperature is between 253 and 280 K and At that time, the static leakage rate of the PSO-SVM output. Add phase transition stability constraints:

[0135] (16)

[0136] in, This is the sensitivity coefficient. This represents the noise amplitude of the temperature sensor. This represents the original static leakage rate predicted by PSO-SVM. This represents the static leakage rate after temperature noise correction.

[0137] (3) Sensor failure tolerance mechanism: If the pressure time series data missing rate is ≥20%, the forced gating network output is activated. =1 (depending only on the static model), and an alarm signal is triggered simultaneously.

[0138] 2. Deployment optimization for industrial scenarios:

[0139] (1) Lightweight model compression:

[0140] (a) Convert the PSO-SVM model into an analytical expression to avoid iterative computation:

[0141] (17)

[0142] in, This represents the static leakage rate predicted by the PSO-SVM model. This indicates the number of support vectors. Let the Lagrange multiplier of the i-th support vector be denoted as . This represents the label value of the i-th support vector. The width parameter represents the RBF kernel function. Represents the input static feature vector This represents the static feature vector corresponding to the i-th support vector. Indicates the bias term;

[0143] (b) The Transformer model is converted to INT8 quantization format, and the model size is compressed to 30% of the original size.

[0144] 3. Practical model validation and output:

[0145] (1) Final performance indicators: On the test set (10,500 groups): the mean absolute error (MAE) is 0.18 kg / s, which is 62% lower than that of the pure physical model; the maximum relative error is <8%; the single prediction time is <50 ms, which meets the industrial real-time requirements;

[0146] (2) Model application scope declaration: The effective prediction domain is the leakage orifice diameter: 1–12 mm, and the extrapolation compensation mechanism supports the over-limit value; the system pressure is 0.08–2.5 MPa, and linear interpolation compensation is enabled in the low pressure zone; the temperature is 250–430 K; the failure boundary is when the input parameters exceed the above range, the physical model is automatically switched to the fallback and a safety alarm is output.

[0147] The present invention employs the above technical solutions, and the leakage rate prediction method for ammonia refrigeration systems based on physical model-driven and multimodal data fusion shows better agreement with experimental measurement results. The simulation method of the present invention is more consistent with actual conditions and is more scientific, thus significantly improving simulation accuracy.

[0148] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0149] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0150] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0151] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0152] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for predicting the leakage rate of an ammonia refrigeration system based on physical model-driven and multimodal data fusion, characterized in that, Includes the following steps: S1. Constructing a physical model: Based on the mass conservation equation, energy conservation equation, and isentropic blockage equation, and coupled with the latent heat phase change process of ammonia liquid vaporization, a physical model for predicting the leakage rate of an ammonia refrigeration system is established. The physical model includes the parameters and computational domain of the high-pressure ammonia storage tank and downstream pipeline. S2. Generate a feature dataset: Based on the physical model, multiple sets of ammonia leakage data containing different operating parameters are obtained through calculation to estimate and determine the key factors affecting the leakage rate, forming a training dataset; S3. Collaborative Machine Learning Modeling: Using a particle swarm optimization-optimized support vector machine (PSO-SVM) model and a Transformer model, the dataset generated in step S2 is collaboratively trained and learned to construct a fusion prediction model for the leakage rate of the ammonia refrigeration system. S4. Model Validation and Correction: The prediction model obtained in step S3 is used to perform parameter tuning, generalization ability testing and experimental validation using the validation set and test set. The model is then corrected based on the comparison between the experimental results and the simulation data to obtain the final practical prediction model.

2. The method for predicting the leakage rate of an ammonia refrigeration system according to claim 1, characterized in that, In step S1, the mass conservation equation is: in, This indicates the total mass flow rate of the leak. , These represent the densities of the liquid phase and the gas phase, respectively. , These represent the volumetric flow rates of the liquid phase and the gas phase, respectively. A blocking effect correction factor ζ is introduced to characterize the local flow resistance of the leakage orifice, and its expression is: in, , These represent the liquid phase and gas phase flow rates, respectively.

3. The method for predicting the leakage rate of an ammonia refrigeration system according to claim 1, characterized in that, In step S1, the energy conservation equation is: in, , These represent the phase enthalpy of the liquid phase and the phase enthalpy of the gas phase, respectively. Indicates the enthalpy value at the leak source inlet. Indicates the latent heat of phase transition; Among them, latent heat of phase transition Based on real-time temperature and pressure dynamic calculations, its expression is: in, Indicates saturation pressure Lower temperature The corresponding latent heat of vaporization, This represents the specific heat capacity at constant pressure. Indicates volume.

4. The method for predicting the leakage rate of an ammonia refrigeration system according to claim 1, characterized in that, In step S1, the isentropic blocking equation is: in, This indicates the stagnant pressure upstream of the leak hole. This indicates the stagnation density upstream of the leak orifice. This indicates the specific heat capacity of ammonia. Indicates the leakage flow rate; Mass flow rate is solved using an iterative method. : in, Represents the flow coefficient. Indicates the area of ​​the leak hole (㎡).

5. The method for predicting the leakage rate of an ammonia refrigeration system according to claim 1, characterized in that, In step S2, the range of operating parameters covered when generating the dataset includes: leakage orifice diameter 1-10mm, system pressure 0.1-2.45 MPa, and system temperature 253-423 K.

6. The method for predicting the leakage rate of an ammonia refrigeration system according to claim 1, characterized in that, Step S3 specifically includes the following sub-steps: S31. Data preprocessing: Generate a dataset containing time-series pressure and temperature data based on the physical model, and inject Gaussian noise to simulate sensor error; S32. Multi-scale feature extraction: Wavelet transform is used to decompose the pressure time series data into multiple layers, extract high-frequency and low-frequency coefficients as features, and calculate the time-domain statistical features of leakage rate; S33. Multimodal feature fusion: fusing physical parameter features with wavelet features to form a feature vector X; S34. Collaborative Model Training and Weight Allocation: The PSO-SVM model and the Transformer model are trained separately, and the fusion weight α is calculated through dynamic weight allocation; S35. Fusion Prediction: The final predicted leakage rate is calculated using the following formula: in, This represents the predicted value of the PSO-SVM model. This represents the predicted value from the Transformer model.

7. The method for predicting the leakage rate of an ammonia refrigeration system according to claim 6, characterized in that, The feature vector X is: in, Indicates average pressure. Indicates the total pressure difference. Indicates leakage flow rate. This indicates a specific temperature parameter for a particular scenario. Let i represent the i-th high-frequency coefficient after wavelet decomposition. Represents the flow coefficient. Indicates the area of ​​the leak. Indicates specific heat ratio. Indicates the density of the liquid. This indicates the density of the gas.

8. The method for predicting the leakage rate of an ammonia refrigeration system according to claim 6, characterized in that, The method for calculating the fusion weight α through dynamic weight allocation is as follows: in, Indicates the PSO-SVM weights. =10.

9. A computer device, characterized in that, include: At least one processor, at least one memory, and computer program instructions stored in the memory, which, when executed by the processor, implement the method as described in any one of claims 1-8.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by a processor, the method described in any one of claims 1-8 is implemented.

Images