Physically-guided decoupling parallel diagnosis method and device for solar heat pump coupling system

Abstract

The invention discloses a decoupling parallel diagnosis method and device for a physically guided solar heat pump coupling system. The method comprises the following steps: acquiring monitoring data in real time; performing three-dimensional dynamic decoupling on the monitoring data to obtain a feature sequence corresponding to each subsystem; inputting each feature sequence into a parallel diagnosis model of a dynamic structure to obtain a fault probability of each subsystem, and generating a fault positioning report; and based on the weight and the deviation degree of each feature and the fault level corresponding to the fault probability, calculating the health index of each subsystem, and outputting a fault diagnosis result in combination with each fault positioning report. Through the above mode, the method can reduce the influence of a complex operation mode on a fault diagnosis and identification effect through three-dimensional dynamic decoupling, and guarantees that each feature sequence has physical significance and is relatively independent. Fault diagnosis can be simultaneously and independently carried out on each subsystem through the parallel diagnosis model, so that the misdiagnosis rate and the missed diagnosis rate are reduced; and through a parallel diagnosis model of a dynamic structure, the adaptability in different environments is improved.

Description

Technical Field

[0001] This application relates to the field of intelligent operation and maintenance technology for new energy equipment, and in particular to a physically guided decoupling parallel diagnostic method and device for a solar heat pump coupling system. Background Technology

[0002] Solar-Assisted Air-Source Heat Pump (SAASHP) systems, also known as solar heat pump coupled systems, achieve efficient hot water supply and heating by coupling solar collectors with an air-source heat pump. However, such complex systems with multiple coupled subsystems face numerous challenges in operation and maintenance, mainly in the following aspects: The coupling of multiple subsystems makes it easy to misdiagnose or miss faults using a single model. A coupled solar heat pump system typically consists of a solar collector (panel) subsystem, an air-source heat pump cycle subsystem, and a hot water storage tank system. The coupling of these subsystems involves the exchange of mass, energy, and signals, leading to serious misdiagnosis problems and making it difficult for maintenance personnel to accurately pinpoint the root cause of faults in a timely manner. When a system failure occurs, a single fault may trigger abnormal signals in multiple subsystems, and multiple faults may also overlap, making it difficult for traditional single models to accurately distinguish the source of the fault. The misdiagnosis rate of traditional single models increases significantly in multi-subsystem coupling scenarios.

[0003] Significant differences in meteorological radiation conditions across regions can lead to poor adaptability of single-model fault diagnosis performance in different areas, hindering effective model transfer between regions. Climate conditions (especially solar radiation heat gain) vary considerably across regions. For example, solar irradiance in Wuhan is typically higher than in Chongqing. These climatic differences cause variations in the runtime of different subsystems and result in the same fault exhibiting different characteristics in different regions, leading to a decline in fault diagnosis performance when a fault diagnosis model trained for one region is transferred to another. Statistics show that regional climate differences can significantly increase the model's false diagnosis rate. Furthermore, traditional transfer learning methods often require collecting large amounts of data in the target region for model fine-tuning, resulting in high transfer costs and low efficiency.

[0004] While data-driven machine learning models have achieved some success in fault diagnosis, these models typically lack physical interpretability, are difficult to incorporate domain expert knowledge, and have limited generalization ability to different working conditions and environments. Summary of the Invention

[0005] This application provides a physical-guided decoupling parallel diagnostic method and apparatus for solar heat pump coupling systems, in order to solve the problems of high false diagnosis rate and high false negative rate caused by the lack of consideration for data coupling in existing solar heat pump coupling system diagnostic models, as well as the poor adaptability of the models across regional environments.

[0006] To address the aforementioned technical problems, this application provides a technical solution: a physically guided decoupling and parallel diagnostic method for a solar heat pump coupling system. This method includes: Real-time acquisition of monitoring data; The monitoring data is dynamically decoupled in three dimensions to obtain the feature sequences corresponding to each subsystem; The feature sequences are input into a parallel diagnostic model with a dynamic structure to obtain the fault probability of each subsystem and generate a fault location report. Based on the weights and deviations of each feature and the fault level corresponding to the fault probability, the health index of each subsystem is calculated, and the fault diagnosis result is output in conjunction with each fault location report.

[0007] In one optional embodiment of this application, the step of performing three-dimensional dynamic decoupling on the monitoring data to obtain the feature sequences corresponding to each subsystem includes: The monitoring data is decoupled based on the time dimension to obtain the solar collector period data and the heat pump period data; The solar heat pump coupling system is decoupled based on the subsystem dimension, and the heat collection period data is allocated to the heat collection subsystem and the heat pump period data is allocated to the heat pump subsystem. Based on the physical rule dimension, the features of the monitoring data are decoupled, and the heat collection period data and the heat pump period data are combined to obtain the heat collection feature sequence and the heat pump feature sequence, respectively.

[0008] In an optional embodiment of this application, the step of inputting each of the feature sequences into a parallel diagnostic model with a dynamic structure to obtain the failure probability of each of the subsystems includes: In response to the monitoring data belonging to the source domain, each of the feature sequences is input into the parallel diagnostic model of the initial structure to obtain the failure probability of each of the subsystems; wherein, the parallel diagnostic model of the initial structure includes a parallel heat collector diagnostic model and a heat pump diagnostic model; In response to the monitoring data belonging to the target domain, each of the feature sequences is input into the parallel diagnostic model after structural update to obtain the failure probability of each of the subsystems; wherein, the parallel diagnostic model after structural update further includes a dual-branch domain adversarial neural network.

[0009] In one optional embodiment of this application, both the heat collector fault model and the heat pump fault model include a feature extractor and a label classifier. The parallel diagnostic model of the initial structure is trained in the following ways: The source domain training set containing the labels is dynamically decoupled in three dimensions to obtain the heat collection feature sequence and the heat pump feature sequence; The heat collector feature sequence is input into the heat collector diagnostic model, and the heat pump feature sequence is input into the heat pump diagnostic model; Based on the feature extractor, fault-related features are extracted and forward propagated to the label classifier to obtain the heat collector failure probability and the heat pump failure probability, and the first classification loss is calculated. Based on the first classification loss, the feature extractor and the label classifier are updated through backpropagation, and the parallel diagnostic model with the initial structure is obtained through iterative training.

[0010] In one optional implementation of this application's embodiments, it includes: At least one of the heat collector diagnostic model and the heat pump diagnostic model is introduced into the dual-branch domain adversarial neural network as the feature extractor to obtain the parallel diagnostic model with updated structure; wherein, the dual-branch domain adversarial neural network includes a dual-branch feature extractor, a domain classifier, a feature fusion layer and a gradient inversion layer.

[0011] In one optional embodiment of this application, the parallel diagnostic model after structural update is trained in the following manner: The unlabeled target domain training set and the labeled source domain training set are dynamically decoupled in three dimensions to obtain the heat collection feature sequence and the heat pump feature sequence. The heat collector feature sequence is input into the heat collector diagnostic model, and the heat pump feature sequence is input into the heat pump diagnostic model; Based on the dual-branch feature extractor, one branch extracts region-sensitive features and the other branch extracts physically invariant features, which are then input into the feature fusion layer to obtain fused features; The fused features are forward propagated to the label classifier to obtain the heat collector failure probability and the heat pump failure probability, and the second classification loss and orthogonality loss are calculated. The dual-branch feature extractor and the label classifier are then backpropagated to update them. The physically invariant features are input into the domain classifier to obtain the distinction between the source domain and the target domain, and the domain classification loss is calculated. The domain invariant feature extractor in the dual-branch feature extractor is updated through backpropagation of the gradient reversal layer, and the parallel diagnostic model with updated structure is obtained through iterative training.

[0012] In an optional embodiment of this application, before generating the fault location report, the method further includes: When the probability of the heat collector failure and the probability of the heat pump failure meet a preset conflict condition, a parallel discrimination mechanism is invoked to perform diagnosis, obtain the fault determination result, and generate the fault location report.

[0013] In one optional embodiment of this application, the calculation function for the health index is: , formula 1; in, The aforementioned health index; The weights of the features; The current value of the feature; The reference value for the aforementioned feature; Dynamic fault threshold; The deviation of the feature; The fault level; This is the scaling factor.

[0014] To solve the above-mentioned technical problems, another technical solution adopted in this application is: to provide a decoupling parallel diagnostic device for a solar heat pump coupling system, characterized in that it includes: The data acquisition module is used to acquire monitoring data in real time; The data decoupling module is used to perform three-dimensional dynamic decoupling of the monitoring data to obtain the feature sequences corresponding to each subsystem; The parallel diagnostic module is used to input the feature sequences into the dynamic parallel diagnostic model to obtain the fault probability of each subsystem and generate a fault location report. The results output module calculates the health index of each subsystem based on the weight and deviation of each feature and the fault level corresponding to the fault probability, and outputs the fault diagnosis result in conjunction with each fault location report.

[0015] To solve the above-mentioned technical problems, another technical solution adopted in this application is: to provide a computer device, including a memory, a processor and a computer program stored in the memory, characterized in that the processor executes the computer program to implement the steps of the above-mentioned physically guided solar heat pump coupling system decoupling parallel diagnosis method.

[0016] The beneficial effects of this application are as follows: Unlike existing technologies, this application discloses a physically guided decoupling parallel diagnostic method and apparatus for a solar heat pump coupling system. This method acquires monitoring data and performs three-dimensional dynamic decoupling to obtain the characteristic sequences corresponding to each subsystem. This reduces the impact of complex operating modes on fault diagnosis and identification, ensuring that the characteristic sequence of each subsystem has a clear physical meaning and is relatively independent from the characteristics of other subsystems. Through a parallel diagnostic model, fault diagnosis of each subsystem can be performed simultaneously and independently, avoiding confusion caused by a single model fitting all subsystems, improving diagnostic efficiency, avoiding feature confusion, reducing false diagnosis and false negative rates, and achieving accurate fault location. By designing a dynamically structured parallel diagnostic model, fault diagnosis can be performed on monitoring data from any region, improving the transferability and adaptability of the parallel diagnostic model in different regional environments. By calculating the health index of each subsystem, a quantitative basis is provided for comparing the status of different regions and different systems and for maintenance decisions. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is a flowchart illustrating an embodiment of the physically guided decoupling and parallel diagnostic method for a solar heat pump coupling system provided in this application; Figure 2 This is a schematic diagram of the heat collector fault model and the heat pump fault model of an embodiment of the physical-guided decoupling and parallel diagnosis method for solar heat pump coupling system provided in this application. Figure 3 This is a schematic diagram of the structure of the updated heat pump diagnostic model, which is an embodiment of the physically guided decoupling and parallel diagnostic method for solar heat pump coupling systems provided in this application. Figure 4 This is a schematic diagram of the training process of the parallel diagnostic model after structural update in an embodiment of the physical-guided decoupling parallel diagnostic method for a solar heat pump coupling system provided in this application. Figure 5 This is a schematic flowchart of source domain data diagnosis, representing an embodiment of the physically guided decoupling parallel diagnostic method for a solar heat pump coupling system provided in this application. Figure 6 This is a schematic diagram of the structure of an embodiment of the decoupling parallel diagnostic device for a solar heat pump coupling system provided in this application; Figure 7This is a schematic diagram of the structure of an embodiment of the storage medium provided in this application; Figure 8 This is a schematic diagram of the structure of an embodiment of the computer device provided in this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0019] The terms "first," "second," and "third" used in the embodiments of this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0020] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0021] This application provides a physically guided decoupling and parallel diagnostic method for a solar heat pump coupling system, see reference. Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the physically guided decoupling and parallel diagnostic method for a solar heat pump coupling system provided in this application. The physically guided decoupling and parallel diagnostic method for a solar heat pump coupling system includes: S10: Acquire monitoring data in real time.

[0022] For a running solar heat pump coupled system, multi-source monitoring data recorded by sensors at each key point are collected in real time. For example, monitoring data corresponding to 23 operating variables (i.e., characteristic variables, hereinafter referred to as characteristics) are usually collected, including monitoring data such as temperature, power of heat pump and water pump, flow rate, and solar irradiance at each key point.

[0023] S20: Perform three-dimensional dynamic decoupling on the monitoring data to obtain the characteristic sequences corresponding to each subsystem.

[0024] In this application, three-dimensional dynamic decoupling mainly involves decoupling the raw monitoring data obtained in step S10 from three dimensions: time, subsystem, and physical rules.

[0025] In this application, the monitoring data is dynamically decoupled in three dimensions to obtain the feature sequences corresponding to each subsystem, including: S21: Decouple the monitoring data based on the time dimension to obtain data on the heat collection period and the heat pump period.

[0026] Due to the solar-assisted operation strategy in the solar heat pump coupling system, the continuous operation data (i.e., real-time monitoring data) obtained from the complete solar heat pump coupling system can be divided into time period data for different operation strategies by using equipment start-up and shutdown information.

[0027] Based on the equipment start / stop signals, the continuous operating data stream is sliced ​​over time to obtain data for each time period. Specifically, when the power of the heat collection circulation pump W... p When the value is greater than 0, the system is in the heat collection operation period, and the data during this period is marked as heat collection period data; when the heat pump power W... d When the value is greater than 0, the system is in the heat pump operation phase, and the data during this phase is marked as heat pump period data; when the power of the heat collection circulation pump W... p >0 and heat pump power W d When the value is greater than 0, the solar heat pump coupling system is in a period where the heat collector and heat pump operate simultaneously. Data during this period is simultaneously labeled as heat collector period data and heat pump period data. Furthermore, data from the hot water storage tank system within the solar heat pump coupling system is monitored throughout the entire process and temporarily labeled as tank period data.

[0028] By decoupling and slicing the monitoring data in a time dimension, the operating data of different subsystems in the solar heat pump coupling system at different times are separated, reducing the impact of complex operating modes on the fault diagnosis and identification effect.

[0029] S22: Decouple the solar heat pump coupling system based on the subsystem dimension, allocate the heat collection period data to the heat collection subsystem, and allocate the heat pump period data to the heat pump subsystem.

[0030] Based on the energy flow and structural information of a complete solar heat pump coupled system, the entire system can be physically divided into several independent subsystems. For example, a series-connected solar heat pump water heating system can be divided into subsystems according to the energy flow loop: a solar collector subsystem (referred to as the collector subsystem, including collectors and circulation pumps, etc.), an air source heat pump circulation subsystem (referred to as the heat pump subsystem, including compressors, evaporators, and condensers, etc.), and a hot water storage tank system (referred to as the water tank system). This decoupling process decouples the collector subsystem loop from the heat pump subsystem loop. Each subsystem can be defined as an independent diagnostic unit, monitoring its key components and characteristic variables during operation to establish a diagnostic model. This application mainly uses the division of a solar heat pump coupled system into a collector subsystem, a heat pump subsystem, and a water tank system as an example for illustration.

[0031] The solar collector time period data and heat pump time period data obtained in step S21 above are respectively assigned to the corresponding solar collector subsystem and heat pump subsystem. This can be understood as the solar collector time period data being the operating data of the decoupled solar collector subsystem, and the heat pump time period data being the operating data of the decoupled heat pump subsystem. As for the water tank time period data, since the water tank system only has 1-2 temperature sensors and heaters, the probability of failure is low and easy to diagnose. Therefore, separate fault diagnosis for the water tank system is not considered, and the water tank time period data can be simultaneously assigned to both the solar collector subsystem and the heat pump subsystem.

[0032] Correspondingly, the various fault types that may occur in the solar collector subsystem and the heat pump subsystem can be categorized into corresponding solar collector fault types (including collector coverage faults and collector water flow leakage faults) and heat pump fault types (including heat pump condenser blockage and heat pump water flow leakage faults). In addition, there is a type of fault that does not occur in either the solar collector or the heat pump, which can be called common fault types. These include faults such as water tank temperature sensor bias, living room temperature sensor bias, and radiator flow leakage. Common fault types can be diagnosed simultaneously by the corresponding diagnostic models of both the solar collector subsystem and the heat pump subsystem.

[0033] In some embodiments, a diagnostic model can also be established for the decoupled water storage tank system. The model structure can refer to the corresponding diagnostic models of the heat collection subsystem and the heat pump subsystem, which will not be elaborated here.

[0034] Decoupling the solar heat pump coupling system from a subsystem perspective not only helps to more accurately locate fault points but also improves the efficiency and accuracy of fault diagnosis. This decoupling method allows each subsystem to independently monitor data and build models, thereby avoiding diagnostic errors caused by mutual interference between systems.

[0035] S23: Based on the physical rule dimension, the features of the monitoring data are decoupled, and the heat collection period data and heat pump period data are combined to obtain the heat collection feature sequence and the heat pump feature sequence, respectively.

[0036] In this application, the decoupling of the physical rule dimension can be achieved from the directions of the law of conservation of mass, the law of conservation of energy, and the Fourier law of heat transfer. Based on the basic principles of thermodynamics and the system energy balance equation, a set of characteristic variables with clear physical meaning and mutual independence are selected for each subsystem.

[0037] Based on the fundamental principles of thermodynamics and the system energy balance equation, the characteristic sequences of each subsystem are extracted, and key characteristic variables that are sensitive to and independent of the faults of each subsystem are selected. At the same time, these key characteristic variables should be easy to measure in practical applications.

[0038] 1) Energy transfer process of the solar collector (collector heat gain ≈ effective utilized heat + heat loss): , formula 2; in, For solar collectors, the heat migration factor (non-characteristic variable) is used. It is the product of the effective transmittance and absorption rate of the transparent cover (a non-characteristic variable). Solar radiation intensity (is a characteristic variable and is easy to measure); The area of ​​the light-collecting surface of the solar collector (non-characteristic variable); The water flow rate at the collector outlet (is a characteristic variable, but flow sensors are expensive and difficult to measure). and These are the water temperatures at the inlet and outlet of the solar collector, respectively (which are characteristic variables and easy to measure). The specific heat capacity of water under constant pressure passing through the solar collector (a non-characteristic variable); The total heat loss coefficient of the solar collector (non-characteristic variable); The average surface temperature of the collector (is a characteristic variable, but surface temperature is difficult to measure). The outdoor ambient temperature is a characteristic variable and is easy to measure.

[0039] Therefore, the characteristic variable ultimately decoupled from the energy transfer process of the solar collector is: solar radiation intensity. collector inlet temperature collector outlet temperature and outdoor ambient temperature .

[0040] 2) Energy transfer process of air source heat pump (condenser heat dissipation ≈ compressor heat dissipation + evaporator heat absorption): , formula 3; in, The heat pump power (is a characteristic variable, but this characteristic variable is mainly used for time-dimension decoupling). The inlet water flow rate of the heat pump (is a characteristic variable, but flow sensors are expensive and difficult to measure); The airflow of the heat pump fan (is a characteristic variable, but flow sensors are expensive and difficult to measure); The specific heat capacity of the inlet water of the heat pump at constant pressure (a non-characteristic variable); The specific heat capacity of outdoor air at constant pressure (a non-characteristic variable); The average temperature of the water tank (is a characteristic variable and is easy to measure); The outlet water temperature of the heat pump (is a characteristic variable and is easy to measure); The exhaust temperature of the heat pump is a characteristic variable that is easy to measure.

[0041] Therefore, the characteristic variable ultimately decoupled from the energy transfer process of an air source heat pump is: the average temperature of the water tank. Heat pump outlet water temperature and heat pump exhaust temperature .

[0042] 3) Energy transfer process of the radiator (radiator heat dissipation ≈ convective heat transfer): , formula 4; in, The water flow rate through the radiator (is a characteristic variable, but flow sensors are expensive and difficult to measure). The average temperature of the water tank (is a characteristic variable and is easy to measure); The outlet water temperature of the radiator (is a characteristic variable and is easy to measure); The specific heat capacity of water under constant pressure passing through the radiator (a non-characteristic variable); For convective heat transfer coefficient (non-characteristic variable); The heat dissipation area of ​​the radiator (non-characteristic variable); The average surface temperature of the radiator (is a characteristic variable, but surface temperature is difficult to measure). The temperature is the living room temperature (which is a characteristic variable and easy to measure).

[0043] Therefore, the characteristic variable ultimately decoupled from the energy transfer process of the radiator is: the average temperature of the water tank. Radiator outlet water temperature and living room temperature .

[0044] In summary, the solar collection characteristic sequence corresponding to the solar collector subsystem is: solar radiation intensity collector inlet temperature collector outlet temperature and outdoor ambient temperature The heat pump characteristic sequence corresponding to the heat pump subsystem is: heat pump outlet water temperature. Heat pump exhaust temperature and outdoor ambient temperature The water tank characteristic sequence corresponding to the water tank system is: average water tank temperature. Radiator outlet water temperature and living room temperature .

[0045] In some embodiments, besides using the energy conservation equation for explicit feature selection, physical rule-based feature decoupling can also employ unsupervised learning (such as principal component analysis (PCA) and autoencoders) for implicit feature decoupling, automatically learning the feature subspace representing the state of each subsystem.

[0046] By decoupling the original 23 features from the physical rule dimension, nine features were filtered out, resulting in a feature sequence for each subsystem. This ensures that the feature sequence of each subsystem has a clear physical meaning and is relatively independent of the features of other subsystems. This feature decoupling method not only simplifies the complexity of the system, making the operating state of each subsystem clearer to understand and analyze, but also helps improve the accuracy and efficiency of system fault diagnosis.

[0047] It is understandable that different brands or models of solar heat pump coupling systems may have differences in structure or application environment. The number of features obtained by key point sensors may be 22 or 24, and the number of features selected after decoupling through physical rule dimensions may be 8 or 10, depending on the actual application scenario. In this application, we will take the selection of the above 9 dimensions of features from 23 dimensions as an example for illustration.

[0048] S30: Input each feature sequence into the parallel diagnostic model of the dynamic structure to obtain the fault probability of each subsystem and generate a fault location report.

[0049] In this application, each feature sequence is input into a parallel diagnostic model with a dynamic structure to obtain the failure probability of each subsystem, including: S31: In response to the monitoring data belonging to the source domain, each feature sequence is input into the parallel diagnostic model of the initial structure to obtain the failure probability of each subsystem. The parallel diagnostic model of the initial structure includes a parallel heat collector diagnostic model and a heat pump diagnostic model.

[0050] In the field of transfer learning, the source domain is a domain that already possesses a large amount of high-quality, labeled data, and a model has been trained on the source domain training set composed of this data. The target domain is a domain where the model needs to be transferred, but where data is scarce, labels are insufficient (or even unlabeled). This application primarily considers model transfer across different geographical environments; therefore, both the source and target domains are geographical. This application uses Wuhan as the source domain and Chongqing as the target domain as an example for illustration.

[0051] In some embodiments, in addition to geographical environment, migration to other fields can also be considered. When there are differences in the data performance of the solar heat pump coupling system in the source domain and the target domain, model migration can be performed through the parallel diagnostic model with dynamic structure provided in this application.

[0052] In this application, when the monitoring data obtained in step S10 is source domain data, the parallel diagnostic model with the initial structure already trained on the source domain training set can be used for fault diagnosis.

[0053] In this application, the solar heat pump coupling system is decoupled into two independently diagnosable subsystems (the heat collector subsystem and the heat pump subsystem). In order to improve the diagnostic efficiency, the diagnostic models of the two subsystems (the heat collector diagnostic model and the heat pump diagnostic model) can be implemented using a parallel diagnostic architecture.

[0054] Specifically, the initial parallel diagnostic model includes a parallel heat collector diagnostic model and a heat pump diagnostic model. The parallel diagnostic architecture can be implemented using two parallel convolutional neural network (CNN) models to achieve simultaneous and independent fault diagnosis of each subsystem.

[0055] In this application, both the heat collector fault model and the heat pump fault model include a feature extractor and a label classifier.

[0056] See Figure 2 , Figure 2 This is a schematic diagram of the collector fault model and the heat pump fault model, representing an embodiment of the physically guided decoupling parallel diagnostic method for a solar heat pump coupling system provided in this application. Both the collector fault model and the heat pump fault model consist of a feature extractor implemented using a convolutional neural network and a label classifier implemented using a fully connected layer. The feature extractor extracts features from the input multidimensional features that have a positive effect on fault classification; the label classifier classifies the faults in the training data corresponding to the features transmitted by the feature extractor.

[0057] The initial parallel diagnostic model was trained in the following ways: S311: Perform three-dimensional dynamic decoupling on the source domain training set containing labels to obtain the heat collection feature sequence and the heat pump feature sequence.

[0058] The data in the source domain training set comes from a large amount of operational data acquired by key sensors of the solar heat pump coupling system. Furthermore, this data has been labeled with fault classification based on the actual fault conditions of the solar heat pump coupling system, thus generating a source domain training set containing the labels.

[0059] For the data in the source domain training set, the same three-dimensional dynamic decoupling method described in steps S21 to S23 is used to decouple the data to obtain the heat collection feature sequence, heat pump feature sequence, and water tank feature sequence. This will not be elaborated here.

[0060] S312: Input the heat collector feature sequence into the heat collector diagnostic model, and input the heat pump feature sequence into the heat pump diagnostic model.

[0061] In this application, the heat collector feature sequence, heat pump feature sequence, and water tank feature sequence are input into the parallel diagnostic model. Each feature sequence contains the data values ​​corresponding to each feature variable. Furthermore, since fault diagnosis of the water tank system is not considered separately, the water tank feature sequence is simultaneously input into both the heat collector diagnostic model and the heat pump diagnostic model.

[0062] For ease of explanation, the solar collector characteristic sequence can be interpreted as including solar radiation intensity. collector inlet temperature collector outlet temperature Outdoor ambient temperature Average temperature of water tank Radiator outlet water temperature and living room temperature The feature sequence comprises 7 dimensions; the heat pump feature sequence can be interpreted as including the heat pump outlet water temperature. Heat pump exhaust temperature Outdoor ambient temperature Average temperature of water tank Radiator outlet water temperature and living room temperature The feature sequence has a total of 6 dimensions.

[0063] S313: Extract fault-related features based on the feature extractor, propagate forward to the label classifier, obtain the heat collector failure probability and the heat pump failure probability respectively, and calculate the first classification loss.

[0064] For the solar collector diagnostic model, after inputting the solar collector feature sequence into the model, the feature extractor in the model extracts the solar radiation intensity from the input. collector inlet temperature collector outlet temperature Outdoor ambient temperature Average temperature of water tank Radiator outlet water temperature and living room temperature From the 7-dimensional feature sequence, features that positively contribute to fault classification (i.e., fault-related features) are extracted and forward-propagated to the label classifier. The label classifier classifies the operational data based on the fault-related features and outputs the probability P of the heat collector failure. col (For example, the probability of collector cover failure).

[0065] For heat pump diagnostic models, after inputting the heat pump feature sequence into the model, the feature extractor in the model extracts the heat pump outlet water temperature from the input. Heat pump exhaust temperature Outdoor ambient temperature Average temperature of water tank Radiator outlet water temperature and living room temperature From the feature sequence spanning six dimensions, features that positively contribute to fault classification (i.e., fault-related features) are extracted and forward-propagated to the label classifier. The label classifier classifies the operational data based on the fault-related features and outputs the heat pump fault probability P. hp (For example, the probability of heat pump refrigerant leakage failure).

[0066] During the training of the solar collector diagnostic model and the heat pump diagnostic model, the model loss is calculated for each diagnostic model. At this point, the model loss equals the first classification loss. The purpose of the first classification loss is to encourage the solar collector diagnostic model to correctly classify faults. The calculation formula is as follows: , formula 5; in, This is the first category loss; The number of samples in the source domain training set; The true fault label for the i-th sample ; Let be the predicted probability of the i-th sample in the c-th type of fault.

[0067] S314: Based on the first classification loss, backpropagation updates the feature extractor and label classifier, and iterative training yields a parallel diagnostic model with the initial structure.

[0068] During the training of the heat collector diagnostic model and the heat pump diagnostic model, for each diagnostic model, the first classification loss is backpropagated to update the feature extractor and the label classifier. This prompts the feature extractor to extract features that have a positive effect on fault classification and the label classifier to correctly classify the fault. The training is iterated until the model converges, resulting in a parallel diagnostic model with a well-trained initial structure.

[0069] In this application, when the monitoring data is source domain data, a parallel diagnostic model with an initial structure already trained on the source domain training set can be used for fault diagnosis. The monitoring data is input into the parallel diagnostic model with the initial structure, specifically the heat collector feature sequence and the heat pump feature sequence are input into the heat pump diagnostic model. After feature extraction and label classification, the heat collector fault probability P is output respectively. col And the probability of heat pump failure P hp Among them, the probability of heat collector failure P col This includes the probability of each fault type, including both collector fault types and common fault types; the heat pump fault probability P. hp It includes the probability of each fault corresponding to the heat pump fault type and the common fault type.

[0070] Unlike existing technologies for complex systems like solar-heat pump coupled systems, which often rely on a single model and are prone to misdiagnosis and missed diagnosis, and struggle to accurately identify the source of faults, this application presents a physical-guided decoupling parallel diagnostic method for solar-heat pump coupled systems. By constructing parallel fault diagnosis models, it achieves independent and simultaneous diagnosis of the collector and heat pump subsystems. This parallel diagnostic architecture not only improves diagnostic efficiency but also significantly enhances the accuracy of fault location. Because the diagnostic models for each subsystem are independent, a fault in one subsystem will not interfere with the diagnostic results of other subsystems, effectively avoiding misdiagnosis and missed diagnosis.

[0071] S32: In response to the monitoring data belonging to the target domain, the feature sequences are input into the parallel diagnostic model after structural update to obtain the failure probability of each subsystem. The parallel diagnostic model after structural update also includes a two-branch domain adversarial neural network.

[0072] In this application, when the monitoring data does not belong to the source domain but to the target domain of the model transfer, it is necessary to transfer the parallel diagnostic model with dynamic structure to achieve cross-regional fault diagnosis.

[0073] To improve the adaptability of parallel diagnostic models in different geographical environments, cross-regional migration of parallel diagnostic models can be achieved through physically guided feature decoupling and regional adaptation. This physically guided migration mechanism can be based on the idea of ​​separating "physically invariant features" and "regionally sensitive features," implicitly (or explicitly) decomposing the original features into two parts: those that are not geographically related and those that are geographically related.

[0074] Specifically, the input feature sequence is decoupled again to separate physically invariant features and geographically sensitive features. For example, a 6-dimensional heat pump feature sequence is further decoupled into a 2-dimensional geographically sensitive feature sequence and a 4-dimensional physically invariant feature sequence. This feature decoupling method can be explicitly divided according to traditional physics theory: physically invariant features X phy This includes quantities that reflect the inherent thermodynamic behavior of the system, such as the temperature difference ΔT between the inlet and outlet of the solar collector. c (T) clo -T cli ), temperature difference ΔT between inlet and outlet of heat pump heat exchanger h (T) hpo -T tank These physically invariant characteristics, such as X, have similar physical meanings and value ranges in different regions, and are theoretically applicable to the general part of diagnosing all faults in solar heat pump coupling systems; region-sensitive characteristic X reg This includes quantities related to local climate conditions, such as ambient temperature T. a Solar irradiance I t These regionally sensitive characteristics vary significantly with location and time, requiring adaptive adjustments based on the actual operating location of the solar heat pump coupling system. Through feature decoupling, the patterns that the parallel diagnostic model needs to learn are divided into universal and regionally specific parts.

[0075] Simultaneously, achieving model transfer requires constructing a transfer learning architecture, consisting of a shared base model and a region adapter. Through this "shared + adapted" transfer learning architecture, the model can utilize the knowledge of the source domain (training region) to achieve performance adaptation in the target domain (new region) without requiring a large amount of labeled data.

[0076] In this application, the base model uses the physically invariant feature X. phy The model is trained to learn a fault detection pattern applicable across regions, and its parameters are frozen during transfer to ensure that the core discrimination capability is not affected by region. The regional adapter targets region-sensitive features X. reg Modeling can be performed, for example using a convolutional neural network model, to generate a climate compensation factor α based on local climate parameters. α is used to adjust the diagnostic threshold or weights to compensate for the impact of regional differences on the diagnostic results. For example, in cold regions, the failure threshold of heat pumps can be appropriately increased to avoid false alarms caused by low ambient temperatures.

[0077] In this application, in addition to explicit partitioning according to traditional physics theories, implicit partitioning can also be performed according to data analysis and identification methods (such as convolutional neural network methods). Specifically, the structure of the parallel diagnostic model with dynamic structure can be updated. Based on the initial structure of the parallel diagnostic model, a two-branch domain adversarial neural network is introduced to simultaneously achieve feature decoupling and transfer architecture.

[0078] In this application, at least one of the heat collector diagnostic model and the heat pump diagnostic model is introduced into a bi-branch domain adversarial neural network as a feature extractor to obtain a parallel diagnostic model with updated structure; wherein, the bi-branch domain adversarial neural network includes a bi-branch feature extractor, a domain classifier, a feature fusion layer and a gradient inversion layer.

[0079] Specifically, the parallel diagnostic model after the structure update may have several forms: it consists of a heat collection diagnostic model with a dual-branch domain adversarial neural network as a feature extractor and a heat pump diagnostic model with the initial structure; it consists of a heat collection diagnostic model with the initial structure and a heat pump diagnostic model with a dual-branch domain adversarial neural network as a feature extractor; and it consists of a heat collection diagnostic model with a dual-branch domain adversarial neural network as a feature extractor and a heat pump diagnostic model with a dual-branch domain adversarial neural network as a feature extractor.

[0080] The specific structure update method used can be determined based on the data differences between the target domain and the source domain. By comparing the differences in the 9-dimensional feature data between the target and source domains, if the difference in data for features belonging to the heat collector sequence exceeds a preset value, the diagnostic performance of the heat collector diagnostic model in the target domain is considered to have degraded, requiring a structure update for model migration. If the difference in data for features belonging to the heat pump sequence exceeds a preset value, the diagnostic performance of the heat pump diagnostic model in the target domain is considered to have degraded, requiring a structure update for model migration. If the features with data differences exceeding a preset value in both the target and source domains belong to both heat collector and heat pump feature sequences, the diagnostic performance of both diagnostic models in the target domain is considered to have degraded, requiring a structure update for model migration in both cases.

[0081] In this application, taking Wuhan as the source domain and Chongqing as the target domain as an example, the heat collection diagnostic model can be directly applied to Chongqing without any degradation in its diagnostic performance. However, the heat pump diagnostic model cannot be applied directly and needs to be migrated across regions before it can be used for cross-regional diagnosis. That is, the parallel diagnostic model after the structure update consists of the heat collection diagnostic model with the initial structure and the heat pump diagnostic model with the introduction of a dual-branch domain adversarial neural network as a feature extractor.

[0082] See Figure 3 , Figure 3This is a schematic diagram of the structure of the updated heat pump diagnostic model, an embodiment of the physically guided decoupling parallel diagnostic method for solar heat pump coupling systems provided in this application. The updated heat pump diagnostic model includes a label classifier and a two-branch domain adversarial neural network. The two-branch domain adversarial neural network includes a two-branch feature extractor, a domain classifier, a feature fusion layer, and a gradient inversion layer. It replaces the feature extractor in the initial heat pump diagnostic model, but its function is not limited to feature extraction; it also includes feature fusion and adversarial training.

[0083] The dual-branch feature extractor consists of a classification discriminative branch (branch A) and a domain-invariant branch (branch B). Branch A is constructed using a Multi-Layer Perceptron (MLP) or a one-dimensional CNN, and its core task is to focus on extracting features most relevant to fault classification, even if these features may be sensitive to the source domain (e.g., Wuhan), i.e., extracting region-sensitive features X. reg It learns the inherent physical laws of system failures, such as: a decrease in collector efficiency leading to ΔT c Abnormal decrease; refrigerant leakage in the heat pump will cause ΔT h Abnormal fluctuations. Branch B also consists of an MLP or CNN, whose core task is to extract geographically insensitive features, i.e., extract physically invariant features X. phy This will eliminate differences between Wuhan and Chongqing in areas such as climate, water quality, and user habits.

[0084] Both branch A and branch B are connected to a feature fusion layer, which concatenates the feature vectors output from branch A and branch B to form a fused feature. This fused feature is then fed into a label classifier for final fault type determination.

[0085] The Gradient Reversal Layer (GRL) connects branch B and the domain discriminator. The features output by branch B (physically invariant features X) phy The data is fed into the domain discriminator, which distinguishes which mixed data come from the source domain and which come from the target domain based on the features. During backpropagation, the gradient inversion layer multiplies the data by a negative coefficient -λ, so that the features extracted by branch B will confuse the domain discriminator, making it unable to distinguish between the source domain and the target domain, thus prompting branch B to extract physically invariant features.

[0086] In this application, when the monitoring data obtained in step S10 is target domain data, the parallel diagnostic model with updated structure that has been trained on the source domain training set and the target domain training set can be used for fault diagnosis.

[0087] In this application, the parallel diagnostic model with updated structure is trained in the following ways: S321: Perform three-dimensional dynamic decoupling on the unlabeled target domain training set and the labeled source domain training set to obtain the heat collection feature sequence and the heat pump feature sequence.

[0088] Multi-source monitoring data from both the target and source domains are collected. The source domain monitoring data is labeled with corresponding fault types, forming a source domain training set containing fault type labels and an unlabeled target domain training set. The sample sizes of the source and target domain training sets can be nearly balanced, or a small number of target domain samples (enough to cover all fault types) can be used to achieve unsupervised adversarial training. The source and target domain training sets are dynamically decoupled in three dimensions to obtain heat collection feature sequences and heat pump feature sequences, respectively.

[0089] S322: Input the heat collector feature sequence into the heat collector diagnostic model, and input the heat pump feature sequence into the heat pump diagnostic model.

[0090] The feature sequences are input into the corresponding diagnostic models. Taking the model transfer of only the heat pump diagnostic model as an example, it is not necessary to transfer the trained initial structure heat collection diagnostic model. Only the heat pump feature sequences need to be input into the heat pump diagnostic model after the structure is updated.

[0091] S323: Based on a dual-branch feature extractor, one branch extracts regionally sensitive features, and the other branch extracts physically invariant features. These are input into a feature fusion layer to obtain fused features.

[0092] For the input heat pump feature sequence, in the dual-branch feature extractor, branch A extracts the region-sensitive feature X. reg Branch B extracts physically invariant features X phy The region-sensitive features and physically invariant features are input into the feature fusion layer, and the two types of features are concatenated into vectors to obtain the fused features.

[0093] S324: Propagate the fused features forward to the label classifier to obtain the heat collector failure probability and the heat pump failure probability, and calculate the second classification loss and orthogonality loss. Backpropagate to update the dual-branch feature extractor and the label classifier.

[0094] The fused features obtained from the heat pump feature sequence are input into a label classifier to obtain the heat pump failure probability. The label classifier has the same structure as the label classifier in the initial heat pump diagnostic model.

[0095] The loss function in this process is calculated, including the second classification loss and the orthogonality loss. The purpose of the second classification loss is to encourage the heat pump diagnostic model to correctly classify faults. The calculation formula is the same as that of the first classification loss, referring to Formula 5 above, and will not be repeated here.

[0096] The purpose of orthogonality loss is to encourage the feature vectors extracted by the two branches to tend to be orthogonal in the feature space, thereby promoting the extraction of as many different features as possible from branch A and branch B, reducing redundancy, and clearly defining the division of labor. The calculation formula is as follows: , Formula 6; in, Orthogonality loss; For branch A, pair of samples Extracted feature vectors; For branch B, pair of samples Extracted feature vectors; Represents the L2 norm; The number of samples in the source domain training set; Number of training set samples in the target domain.

[0097] By backpropagating the second classification loss and orthogonality loss, the bi-branch feature extractor and label classifier are updated, prompting the bi-branch feature extractor to extract features that have a positive effect on fault classification, and the features extracted by branch A and branch B are as different as possible, so as to prompt the label classifier to correctly classify faults.

[0098] S325: Input the physically invariant features into the domain classifier to obtain the distinction between the source domain and the target domain, and calculate the domain classification loss. Update the domain invariant feature extractor in the dual-branch feature extractor through backpropagation of the gradient reversal layer, and iteratively train to obtain the parallel diagnostic model after structural update.

[0099] For the physically invariant feature X extracted from branch B phy The data is fed into the feature fusion layer and the domain classifier at the same time. The domain classifier distinguishes between the source domain and the target domain data and outputs the result of the distinction between the source domain and the target domain.

[0100] The adversarial loss is calculated during this process. The domain discriminator tries its best to distinguish the data sources. The role of the adversarial loss is to encourage the feature extractor (especially branch B) to try to confuse the domain discriminator through GRL, thereby forcing branch B to extract physically invariant features X. phy The formula for calculating adversarial losses is as follows: , Formula 7; in, To combat the losses; The number of samples in the source domain training set; Number of samples in the target domain training set; This represents the model's predicted probability that sample xi belongs to the target domain; This represents the model's predicted probability that sample xj belongs to the source domain.

[0101] λ is the gradient reversal adaptive coefficient, which gradually increases from 0 according to the number of training epochs to smoothly initiate adversarial training. The calculation formula is as follows: , formula 8; in, α is the maximum adversarial weight coefficient; t is the current training round number; T is the total number of training rounds; and α is the scheduling rate parameter.

[0102] Adversarial loss is countered by backpropagation through a gradient reversal layer to update the neighborhood-invariant feature extractor in the dual-branch feature extractor, prompting branch B to extract physically invariant features X. phy This makes it difficult for the domain discriminator to distinguish between target domain and source domain data.

[0103] During the adversarial training of the entire restructured heat pump diagnostic model, the formula for the total model loss function is as follows: , formula 9; in: Total loss; This is the second category loss; To combat the losses; λ is the orthogonality loss; λ is the gradient reversal adaptive coefficient; β is the cross-validation loss coefficient.

[0104] By backpropagating the total loss of the model, iterative training is performed until the model converges, resulting in a well-trained heat pump diagnostic model with updated structure.

[0105] In this application, steps S324 and S325 are not executed sequentially, but in parallel. Furthermore, the two branches in the structurally updated heat pump diagnostic model can be trained collaboratively or in stages.

[0106] In some embodiments, the structurally updated heat pump diagnostic model can be trained in three stages: the first stage training branch B begins adversarial training to learn the physically invariant features X. phy The second stage introduces branch A and orthogonality loss to begin learning the regionally sensitive feature X. reg The three-stage, dual-branch collaborative training process adaptively adjusts λ until the model converges, resulting in a well-trained, structurally updated heat pump diagnostic model.

[0107] In this application, the trained, structurally updated heat pump diagnostic model and the initial structural heat collector diagnostic model, which can be directly used in the target domain without migration, together constitute the trained, structurally updated parallel diagnostic model. When the monitoring data obtained from step S10 belongs to the target domain, the trained, structurally updated parallel diagnostic model can be used for fault diagnosis. The heat collector feature sequence and heat pump feature sequence obtained through three-dimensional dynamic decoupling are input into the structurally updated parallel diagnostic model, and the heat collector fault probability P is output.col And the probability of heat pump failure P hp .

[0108] Unlike existing technologies where diagnostic models are typically trained on source domain climate data, performance degrades significantly when deployed to target domains with drastically different climates due to drastic changes in key features such as ambient temperature, solar irradiance, and humidity. Furthermore, traditional transfer learning often requires collecting large amounts of new labeled data in the target domain for fine-tuning, which is costly, time-consuming, and hinders rapid deployment in engineering practice, severely restricting the large-scale application of the technology. This application presents a physical-guided decoupling parallel diagnostic method for solar heat pump coupling systems. Through this dynamically structured parallel diagnostic model, it not only performs excellently on source domain data but also achieves efficient fault diagnosis on target domain data after structural updates. This eliminates the need for large-scale labeled data collection and model fine-tuning in the target domain, significantly reducing costs and time in engineering practice and providing strong support for the large-scale application of the technology.

[0109] In one embodiment, see Figure 4 , Figure 4 This is a schematic diagram of the training process of the parallel diagnostic model after structural update, according to an embodiment of the physical-guided decoupling parallel diagnostic method for a solar heat pump coupling system provided in this application. First, the target domain and source domain data are dynamically decoupled in three dimensions to obtain decoupled source domain data (labeled) and decoupled target domain data (unlabeled). These data are then input into the structurally updated parallel diagnostic model. Next, the regionally sensitive feature X is extracted through branches A and B, respectively. reg and physically invariant characteristic X phy The output features of the two branches are input into the feature fusion layer, which outputs fused features. The fused features are then input into the label classifier to predict the fault type. At the same time, the output features of branch B are input into the domain discriminator to predict the domain type. Finally, the classification loss, adversarial loss, and orthogonality loss in the process are calculated, and the two-branch feature extractor and label classifier are updated by backpropagation. The training is iterated until the model converges, and the trained parallel diagnostic model with updated structure is obtained.

[0110] In this application, before generating the fault location report, the following are also included: When the probability of heat collector failure and the probability of heat pump failure meet the preset conflict conditions, the parallel discrimination mechanism is invoked to perform diagnosis, obtain the fault determination result, and generate a fault location report.

[0111] In this application, a parallel fault discrimination mechanism is set up to prevent diagnostic conflicts when diagnosing multiple subsystems simultaneously. Specifically, the parallel discrimination mechanism is triggered when the probability of a heat collector failure and the probability of a heat pump failure meet a preset conflict condition, i.e., when both contain the probability of the same common failure. The parallel discrimination mechanism will make a judgment based on the subsystem with the higher probability of the common failure and output the fault judgment result of the parallel diagnosis, ensuring the reliability of diagnosis under complex operating conditions. For example, for a water tank temperature sensor bias failure, the probability of this failure in the heat collector failure probability is 0.95, and the probability of this failure in the heat pump failure probability is 0.3. Since 0.95 > 0.3, the probability of the water tank temperature sensor bias failure is considered to be 0.95.

[0112] If, at any given time, the common fault only presents as either a collector fault probability or a heat pump fault probability, then parallel discrimination is unnecessary; the fault determination result is directly given based on the output collector fault probability or heat pump fault probability. Specifically, a specific fault location report is output based on the set fault occurrence threshold (e.g., a fault probability greater than 0.8 is considered a fault). For example, if P... col If the fault occurrence threshold is exceeded, a report of "Coverage fault in the solar collector subsystem" will be submitted; if P hp If the fault threshold is exceeded, a report of "refrigerant leakage fault in the heat pump subsystem" will be generated. In cases of diagnostic conflicts arising from simultaneous diagnosis of multiple subsystems, the parallel discrimination mechanism will also output a final fault location report. For example, if, after parallel discrimination, the probability of a water tank temperature sensor bias fault is determined to be 0.95, and since 0.95 > 0.8, a report of "water tank temperature sensor bias fault" will be generated. The fault location report clearly indicates the fault location and possible causes, helping maintenance personnel to quickly take repair measures.

[0113] S40: Based on the weights and deviations of each feature and the fault level corresponding to the fault probability, calculate the health index of each subsystem, and output the fault diagnosis results in conjunction with each fault location report.

[0114] In this application, based on the completion of fault probability diagnosis, physical rules and interpretable machine learning technology are further integrated to calculate the health index HI of the equipment, which is used to quantify the overall health status of the solar heat pump coupling system.

[0115] The Health Index (HI) calculation comprehensively considers the failure probability, failure severity, and the degree to which key characteristic parameters deviate from the normal range for each subsystem, providing a quantitative basis for comparing the status of different regions and systems and making maintenance decisions. Specifically, it includes three steps: characteristic importance assessment, failure severity classification, and health index calculation.

[0116] Feature importance assessment: Using interpretability methods such as Layer-wise Relevance Propagation (LRP), the contribution weight w of each monitored feature to the final diagnostic result is calculated. i (can be abbreviated as weight w) i ), weight w i This reflects the extent to which this abnormal characteristic affects the health of the system.

[0117] Specifically, an algorithm utilizing local correlation propagation interpretability can be used to analyze the contribution of each key feature to the final diagnostic result, thereby obtaining the weight w. i The calculation formula for the local correlation propagation interpretability algorithm is as follows: , formula 10; in, This is the relevance score of neuron k in the previous layer (closer to the output layer); This is the relevance score that neuron j in the current layer (closer to the input layer) will receive; This represents the contribution of neuron j to the activation of neuron k, and , This represents the activation value of neuron j during forward propagation. Indicates connection weights; the LRP-β rule will activate the term. It is decomposed into positive and negative components, and a penalty of scaling factor β is applied to the negative components, where , .

[0118] Fault Severity Classification: Assign a fault level S to the identified fault type. i (For example, level 1 represents a minor fault, and level 4 represents a serious fault). Fault levels can be determined based on expert experience or historical data, reflecting the degree of impact of the fault on system functionality and safety.

[0119] Health index calculation: Based on the deviation and weight of each feature, as well as the fault level, a comprehensive health index (HI) is calculated. For example, for each key feature parameter, the degree of deviation of its current value from the normal range is calculated and multiplied by the corresponding weight w. i and fault level S i The total health loss is obtained by summing the values, and then the total loss is subtracted from 100 to obtain the HI value (HI is represented by 0~100, and the higher the value, the healthier the person).

[0120] In this application, the function for calculating the health index is: , formula 1; in, As a health index; The weights of the features; The current value of the feature; The baseline value for the feature; Dynamic fault threshold; The degree of deviation of the feature; Fault level; This is the scaling factor.

[0121] By introducing physical rules (such as normal parameter ranges and fault impact models) and interpretable weights, the calculation process of the Health Index (HI) has clear physical meaning, making it easy for operations and maintenance personnel to understand and trust.

[0122] This application provides corresponding maintenance recommendations based on the HI value and fault location report. For example, an HI value above 90 indicates a good system health condition, and regular inspections are recommended; an HI value between 70 and 90 indicates a minor abnormality, and it is recommended to check the relevant components within 1 to 2 weeks; an HI value below 70 indicates a serious fault or performance degradation, and it is recommended to arrange repairs as soon as possible.

[0123] The final output of this application is maintenance decision-making information for operations and maintenance personnel, including fault location reports and health scores and maintenance decision recommendations. The health scores and maintenance decision recommendations combine quantitative indicators and expert experience to provide clear guidance for operations and maintenance decisions.

[0124] In one embodiment, see Figure 5 , Figure 5 This is a schematic flowchart illustrating the source domain data diagnosis process of an embodiment of the physical-guided decoupling parallel diagnosis method for a solar heat pump coupling system provided in this application. First, multi-source sensor data (such as temperature, power, irradiance, and flow rate) is acquired. Second, the multi-source sensor data undergoes three-dimensional dynamic decoupling, including time-dimension decoupling, subsystem-dimension decoupling, and physical rule-dimension decoupling, resulting in a collector feature sequence and a heat pump feature sequence. Then, the decoupling features are input into a parallel diagnosis model, including inputting the 7-dimensional collector feature sequence into the collector feature model and outputting the collector failure probability P. col The 6-dimensional heat pump feature sequence is input into the heat pump feature model, and the heat pump failure probability P is output. hp Then, determine whether the failure probability of the heat collector and the failure probability of the heat pump meet the preset conflict conditions. If so, trigger the parallel discrimination mechanism to perform diagnosis and output the failure judgment result. Otherwise, directly output the failure probability. Finally, output the failure location report (failure location and possible causes) and calculate the health index HI.

[0125] In one embodiment, a solar heat pump coupling system in Wuhan is used as the object to verify the diagnostic effect of a parallel diagnostic model trained on a source domain training set in a source domain environment.

[0126] This solar-heat pump coupling system is equipped with a flat-plate solar collector, an air-source heat pump unit, and a 150L hot water storage tank. It adopts a conventional control strategy (starting the water tank heating temperature T). start =45℃, stop water tank heating temperature T stop =50℃). During the experiment, a shading material was covered on the surface of the collector to simulate a collector shading failure (COL). cover ), reducing the flow rate of the collector loop pump to simulate collector flow leakage fault (COL) leak ), simulating a heat pump condenser clogging fault by covering the heat pump condenser with a baffle plate (HP). block ), reduce heat pump loop water pump flow to simulate heat pump flow leakage fault (HP) leak ), baseline enhancement of water tank temperature sensor to simulate water tank temperature sensor bias fault (TTS) offset ), simulating living room temperature sensor baseline improvement to simulate living room temperature sensor bias fault (RTS) offset ), reduce radiator loop pump flow to simulate radiator flow leakage fault (USE) leak The system records sensor data from the solar heat pump coupling system. Specific parameter settings are shown in Table 1 below.

[0127]

[0128] Table 1 Parameter Setting Table First, data acquisition was conducted. Twenty-three operational variables were collected in real time, including temperature at key points, power of the heat pump and water pump, flow rate, and solar irradiance.

[0129] Secondly, three-dimensional dynamic decoupling is performed on the multi-source detection data. Time dimension: Based on the start / stop signals of the water pump and compressor, the data is divided into a solar collector cycle operation period (Wp>0) and a heat pump operation period (Wd>0). In this case, the water pump starts during the day when there is sunlight, entering the solar collector cycle and collecting data during this period; at night or when the water temperature is insufficient, the heat pump starts, entering the heat pump cycle and collecting data during this period. Subsystem dimension: Data from the solar collector cycle is assigned to the solar collector subsystem, and data from the heat pump cycle is assigned to the heat pump subsystem, and both are processed independently. Physical rule dimension: Thermodynamic rules are applied to select features; the solar radiation intensity is selected for the solar collector channel. collector inlet temperature collector outlet temperature Outdoor ambient temperature Average temperature of water tank Radiator outlet water temperature and living room temperature The feature sequence consists of 7 dimensions, with the heat pump channel selecting the heat pump outlet water temperature. Heat pump exhaust temperature Outdoor ambient temperature Average temperature of water tank Radiator outlet water temperature and living room temperature There are a total of 6 dimensions of feature sequences. These features can sensitively reflect the operating status of their respective subsystems. The specific features of these 9 dimensions after decoupling are shown in Table 2 below.

[0130]

[0131] Table 2 Characteristic parameters after decoupling Then, parallel diagnostics are performed on the decoupled heat collector characteristic sequence and heat pump characteristic sequence. The heat collector characteristic sequence is input into the heat collector diagnostic model, and the heat collector failure probability P is output. col Input the heat pump characteristic sequence into the heat pump diagnostic model, and output the heat pump failure probability P. hp Taking the common fault of the living room temperature sensor bias as an example, when the heat collection subsystem and the heat pump subsystem are running simultaneously, the heat collection diagnostic model outputs P for this fault. col =0.92, indicating a high probability of this common fault occurring; while the heat pump diagnostic model outputs P hp =0.50 indicates that the common fault has a moderate probability of occurrence. The parameter settings for the parallel diagnostic model are shown in Table 3 below.

[0132] serial number parameter set up 1 Learning rate 0.0005 2 Optimizer Adam 3 Number of neurons 512 4 Excitation function Softsign 5 kernel size 1×1 66 Convolution kernel number 32,64,32 7 Number of convolutional layers 3 8 Fully connected layers 2 Table 3 Parameter settings for the parallel diagnostic model Then, a parallel discriminant decision is made regarding the probability of heat collector failure and the probability of heat pump failure. Because P col >P hp The parallel discrimination mechanism determines that the probability of a bias fault occurring in the living room temperature sensor is P. col =0.92. Because P col =0.92 is greater than the fault occurrence threshold of 0.8. Therefore, the fault determination result is that the living room temperature sensor bias fault has occurred.

[0133] Finally, the health index HI is calculated based on the fault diagnosis results. The weights of each feature, such as T, are obtained using the LRP algorithm. rdo weight w i The value is approximately 0.3, indicating a significant impact on the diagnostic results. The fault level S for the living room temperature sensor bias fault is... i It was set to Level 2 (moderate severity). When calculating HI, T was taken into account. rdo The current value is 43℃, while the normal operating temperature should be 40-42℃. i The average temperature is 41℃, and the fault threshold is... If the temperature is 42℃, then the deviation is (43-41) / (42-41)=2. The scaling factor K=15. Combining the weight and fault level, the HI loss caused by this fault is approximately 15*0.3*(2)*2=18. Assuming that other indicators are normal and there is no loss, the final HI≈100-18=82. The system outputs a health score of 82 and recommends "checking the room temperature sensor components within 1 week".

[0134] Maintenance personnel checked the output fault diagnosis report and found that the room temperature sensor was indeed biased. After replacing the sensor, the system returned to normal. It can be seen that the physical-guided decoupling parallel diagnosis method for solar heat pump coupling systems provided in this application accurately detected and located the fault, avoiding the missed diagnoses or misjudgments that may occur with traditional methods.

[0135] The trained parallel diagnostic model was used for fault diagnosis of a large amount of data from a solar heat pump coupled system in the source domain. The average accuracy, false diagnosis rate, and false negative rate of the model in the source domain were calculated. Simultaneously, comparative experiments were conducted using other single-model diagnostic methods on the same monitoring data. The specific comparison of fault diagnosis results for different models is shown in Table 4 below.

[0136] Diagnostic model Average diagnostic accuracy misdiagnosis rate Misdiagnosis rate Parallel diagnostic model 95.3% 4% 6% Convolutional Neural Networks 77.3% 66% 12% Decision Tree 71.7% 71% 14% Support Vector Machine 77.2% 66% 13% Random Forest 74.2% 70% 14% Nearest neighbor model 61.2% 77% 21% Table 4. Comparison of Fault Diagnosis Results for Different Models Applied to the Source Domain Experiments show that the parallel diagnostic model can achieve a diagnostic accuracy of over 95% in the source domain, with a false diagnosis rate of only 4% and a false negative rate of only 6%. Compared with other single models such as convolutional neural networks, decision trees, and support vector machines, the diagnostic accuracy is significantly improved, and the false diagnosis rate and false negative rate are significantly reduced. At the same time, it provides better interpretability and provides reliable technical support for the intelligent operation and maintenance of solar heat pump coupling systems.

[0137] In another embodiment, taking the application of a parallel diagnostic model trained in the source domain (Wuhan) to a similar solar heat pump coupling system in the target domain (Chongqing) as an example, due to differences in some characteristic parameters between the two locations, the parallel diagnostic model needs to be transferred when applied to the target domain. Therefore, the transfer capability of the dynamic parallel diagnostic model under different regional environments (Chongqing has lower solar irradiance than Wuhan) is verified. The specific comparison of the differences in characteristic parameters between Wuhan and Chongqing is shown in Table 5 below.

[0138] Place Average solar radiation (difference) Average temperature (difference) at the collector outlet Wuhan 350W / m² (-) 18.1℃（-） Chongqing 183W / m² (-48%) 16.7℃（-8%） Table 5 Comparison of characteristic parameters between Wuhan and Chongqing The parallel diagnostic model with dynamic structure is updated. Based on the initial parallel diagnostic model, a dual-branch domain adversarial neural network is introduced to simultaneously achieve feature decoupling and transfer architecture.

[0139] The updated parallel diagnostic model consists of two branches: branch A (classification discriminative branch) and branch B (physical invariance branch). The geographically sensitive features and physically invariant features extracted from the two branches are fused and then input into the label classifier for fault classification.

[0140] Input the key features after three-dimensional dynamic decoupling into the parallel diagnostic model after structural update (7 features for the heat collector diagnostic model and 6 features for the heat pump diagnostic model).

[0141] Branch A uses a multilayer perceptron (MLP) or CNN, focusing on extracting the features most relevant to fault classification from the input (i.e., the geographically sensitive feature X). reg Even though these features may be domain-sensitive; branch B uses another MLP or CNN and connects it to the domain discriminator via a gradient inversion layer, focusing on extracting physically invariant features X. phy The features output from branch A and branch B are concatenated to form a fused feature. The label classifier classifies faults based on the fused feature; the domain discriminator classifies the domain (source domain or target domain) based on the features output from branch B.

[0142] The training process based on the source and target domain training sets includes iterative training of the model via forward and backward propagation. Forward propagation: The key features of the input 3D dynamically decoupled model are processed by a two-branch feature extractor, outputting feature f. A and f B After fusion, the fusion feature f is obtained. fused And input it into the classifier output fault prediction, f B The data is then fed into the domain discriminator to predict the output domain, and the classification loss and adversarial loss are calculated. Backpropagation: The gradient of the adversarial loss is backpropagated to branch B of the two-branch feature extractor through a gradient reversal layer, while the classification loss is backpropagated normally.

[0143] In this physics-guided transfer training, branch A eliminates regional differences, including climate influences (temperature and humidity differences between the hot and humid Wuhan and the foggy Chongqing), air quality influences (different rates of fan clogging due to air quality in the two locations), and operating habits (regional characteristics of user usage patterns); branch B learns physical laws, including decreased collector efficiency → abnormal temperature differences between collector inlet and outlet; and heat pump performance degradation → heat pump exhaust temperature T. air Abnormality; refrigerant leakage → temperature difference ΔT between heat pump inlet and outlet h Abnormal fluctuations.

[0144] The training process consists of three stages: Stage 1: Training branch B to begin adversarial training and learn physically invariant features; Stage 2: Introducing branch A and orthogonality loss to begin learning geo-sensitive features; Stage 3: Dual-branch collaborative training, adaptively adjusting λ, iteratively training until the model converges, and obtaining a parallel diagnostic model with a trained dual-branch DANN (Domain Adversarial Network) structure.

[0145] The trained parallel diagnostic model with a dual-branch DANN structure was used for fault diagnosis of a similar solar heat pump coupled system in the target domain (Chongqing). The average accuracy and false alarm rate in the target domain were calculated. Based on the difference in solar radiation parameters, the climate compensation factor α = 0.1 * (350 - 183) / 350 ≈ 0.05. Therefore, the fault occurrence threshold for collector coverage faults is 0.8 + 0.05 = 0.85, which can avoid false alarms caused by local building obstruction. Simultaneously, for the same monitoring data, after decoupling and physical guidance, the parallel diagnostic model without structural updates and the parallel diagnostic model with structural updates but not a dual-branch DANN structure were compared. The specific comparison of the fault diagnosis results of the different models is shown in Table 6 below.

[0146] method Target domain average accuracy (improved) Target domain misdiagnosis rate (improved) Parallel diagnostic model without structural updates 88.9%（-） 51%（-） Parallel diagnostic model based on DANN architecture 90.6%（1.7%） 16%（35%） Parallel diagnostic model with dual-branch DANN structure 92.2%（3.3%） 5%（46%） Table 6. Comparison of Fault Diagnosis Results for Different Models Applied to the Target Domain Experiments show that after structural updates, the parallel diagnostic model with a dual-branch DANN structure can achieve a cross-regional diagnostic accuracy of over 92%, with a false diagnosis rate of only 5%, which is about 11% higher than that of the DANN structure. It also provides better interpretability and provides reliable technical support for the intelligent operation and maintenance of solar heat pump coupling systems.

[0147] The experimental data in Tables 4 and 6 above show that, compared with traditional methods, the physical-guided decoupling parallel diagnostic method for solar heat pump coupling systems provided in this application has significant advantages in diagnostic accuracy, cross-regional adaptability, and lightweight model. Specific analysis is as follows: The diagnostic accuracy is significantly improved: the average diagnostic accuracy of the physical-guided decoupling parallel diagnostic method for solar heat pump coupling systems provided in this application reaches 95.3%, which is much higher than the 77.2% of the traditional method. This is mainly due to the decoupling parallel diagnostic method reducing misdiagnosis and the adaptive adjustment avoiding missed diagnosis.

[0148] Excellent cross-domain migration performance: In cross-regional migration applications, the parallel diagnostic model with a dual-branch DANN structure provided in this application has a false diagnosis rate of only 5%, compared to 51% for the parallel diagnostic model without structure updates and 16% for the DANN structure parallel diagnostic model, representing an improvement of more than 10% over the DANN structure. The physically guided migration mechanism ensures that the parallel diagnostic model maintains high accuracy in the new regional environment.

[0149] Lightweight Model: The physical-guided decoupling parallel diagnostic method for solar heat pump coupling systems provided in this application can compress the model size to a certain extent through feature decoupling and shared model architecture. The lightweight model facilitates deployment on edge devices, reduces computing and storage costs, and improves online real-time diagnostic efficiency.

[0150] In summary, the physical-guided decoupling and parallel diagnostic method for solar heat pump coupling systems provided in this application effectively solves the problem of fault diagnosis in different regions of solar heat pump coupling systems through three-dimensional coupling, parallel diagnosis, and physical-guided model transfer, and significantly improves the reliability and operation and maintenance efficiency of solar heat pump coupling systems.

[0151] This application provides a decoupling and parallel diagnostic device for a solar heat pump coupling system, see reference. Figure 6 , Figure 6 This is a schematic diagram of an embodiment of the solar heat pump coupling system decoupling parallel diagnostic device provided in this application. The solar heat pump coupling system decoupling parallel diagnostic device includes: Data acquisition module 10 is used to acquire monitoring data in real time; The data decoupling module 20 is used to perform three-dimensional dynamic decoupling of monitoring data to obtain the feature sequences corresponding to each subsystem; Parallel diagnostic module 30 is used to input each feature sequence into a dynamic parallel diagnostic model to obtain the fault probability of each subsystem and generate a fault location report. The results output module 40 calculates the health index of each subsystem based on the weight and deviation of each feature and the fault level corresponding to the fault probability, and outputs the fault diagnosis results in conjunction with each fault location report.

[0152] The data acquisition module 10, data decoupling module 20, parallel diagnosis module 30 and result output module 40 interact to realize the process of decoupling and parallel diagnosis of the solar heat pump coupling system. You can refer to the specific description of steps S10 to S40 above. The repeated parts will not be repeated here.

[0153] See Figure 7 , Figure 7 This is a schematic diagram of an embodiment of the storage medium provided in this application.

[0154] The storage medium 700 stores program data 710, which, when executed by the processor, implements, as follows: Figure 1 The steps of the physical-guided decoupling parallel diagnostic method for solar heat pump coupling systems are described.

[0155] The program data 710 is stored in a storage medium 700 and includes several instructions for causing a network device (which may be a router, personal computer, server, or other network device) or processor to execute all or part of the steps of the methods described in the various embodiments of this application.

[0156] Optionally, the storage medium 700 can be any medium that can store program data, such as a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), disk, or optical disc.

[0157] See Figure 8 , Figure 8 This is a schematic diagram of the structure of an embodiment of the computer device provided in this application.

[0158] The device 800 includes a processor 820 and a memory 810 connected to each other. The memory 810 stores a computer program. When the processor 820 executes the computer program, it implements the steps of the physically guided solar heat pump coupling system decoupling parallel diagnostic method described above.

[0159] Unlike existing technologies, this application discloses a physically guided decoupling parallel diagnostic method and apparatus for a solar heat pump coupling system. This method acquires monitoring data and performs three-dimensional dynamic decoupling to obtain the characteristic sequences corresponding to each subsystem. This reduces the impact of complex operating modes on fault diagnosis and identification, ensuring that the characteristic sequence of each subsystem has a clear physical meaning and is relatively independent from the characteristics of other subsystems. The parallel diagnostic model enables simultaneous and independent fault diagnosis of each subsystem, avoiding confusion caused by a single model fitting all subsystems, improving diagnostic efficiency, preventing feature confusion, reducing false and false diagnosis rates, and achieving accurate fault location. By designing a dynamically structured parallel diagnostic model, fault diagnosis can be performed on monitoring data from any region, improving the transferability and adaptability of the parallel diagnostic model in different geographical environments. By calculating the health index of each subsystem, a quantitative basis is provided for comparing the status of different regions and systems and for maintenance decisions.

[0160] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the storage medium embodiments and computer device embodiments are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0161] This application can be used in a wide range of general-purpose or special-purpose computing system environments or configurations. For example: personal computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, network PCs, minicomputers, distributed computing environments including any of the above systems or devices, etc.

[0162] In the several embodiments provided in this application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device embodiments described above are merely illustrative; multiple units or components may be combined or integrated into another system, or some features may be omitted or not performed.

[0163] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0164] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0165] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A physically guided decoupling and parallel diagnostic method for a solar heat pump coupling system, characterized in that, include: Real-time acquisition of monitoring data; The monitoring data is dynamically decoupled in three dimensions to obtain the feature sequences corresponding to each subsystem; The feature sequences are input into a parallel diagnostic model with a dynamic structure to obtain the fault probability of each subsystem and generate a fault location report. Based on the weights and deviations of each feature and the fault level corresponding to the fault probability, the health index of each subsystem is calculated, and the fault diagnosis result is output in conjunction with each fault location report.

2. The physical-guided decoupling and parallel diagnostic method for a solar heat pump coupling system according to claim 1, characterized in that, The three-dimensional dynamic decoupling of the monitoring data to obtain the feature sequences corresponding to each subsystem includes: The monitoring data is decoupled based on the time dimension to obtain the solar collector period data and the heat pump period data; The solar heat pump coupling system is decoupled based on the subsystem dimension, and the heat collection period data is allocated to the heat collection subsystem and the heat pump period data is allocated to the heat pump subsystem. Based on the physical rule dimension, the features of the monitoring data are decoupled, and the heat collection period data and the heat pump period data are combined to obtain the heat collection feature sequence and the heat pump feature sequence, respectively.

3. The physical-guided decoupling and parallel diagnostic method for a solar heat pump coupling system according to claim 2, characterized in that, The step of inputting each of the aforementioned feature sequences into a parallel diagnostic model with a dynamic structure to obtain the failure probability of each of the aforementioned subsystems includes: In response to the monitoring data belonging to the source domain, each of the feature sequences is input into the parallel diagnostic model of the initial structure to obtain the failure probability of each of the subsystems; wherein, the parallel diagnostic model of the initial structure includes a parallel heat collector diagnostic model and a heat pump diagnostic model; In response to the monitoring data belonging to the target domain, each of the feature sequences is input into the parallel diagnostic model after structural update to obtain the failure probability of each of the subsystems; wherein, the parallel diagnostic model after structural update further includes a dual-branch domain adversarial neural network.

4. The physical-guided decoupling and parallel diagnostic method for a solar heat pump coupling system according to claim 3, characterized in that, Both the heat collector fault model and the heat pump fault model include a feature extractor and a label classifier. The parallel diagnostic model of the initial structure is trained in the following ways: The source domain training set containing the labels is dynamically decoupled in three dimensions to obtain the heat collection feature sequence and the heat pump feature sequence; The heat collector feature sequence is input into the heat collector diagnostic model, and the heat pump feature sequence is input into the heat pump diagnostic model; Based on the feature extractor, fault-related features are extracted and forward propagated to the label classifier to obtain the heat collector failure probability and the heat pump failure probability, and the first classification loss is calculated. Based on the first classification loss, the feature extractor and the label classifier are updated through backpropagation, and the parallel diagnostic model with the initial structure is obtained through iterative training.

5. The physical-guided decoupling and parallel diagnostic method for a solar heat pump coupling system according to claim 3, characterized in that, include: At least one of the heat collector diagnostic model and the heat pump diagnostic model is introduced into the dual-branch domain adversarial neural network as the feature extractor to obtain the parallel diagnostic model with updated structure; wherein, the dual-branch domain adversarial neural network includes a dual-branch feature extractor, a domain classifier, a feature fusion layer and a gradient inversion layer.

6. The physical-guided decoupling and parallel diagnostic method for a solar heat pump coupling system according to claim 5, characterized in that, The parallel diagnostic model after structural update is trained in the following ways: The unlabeled target domain training set and the labeled source domain training set are dynamically decoupled in three dimensions to obtain the heat collection feature sequence and the heat pump feature sequence. The heat collector feature sequence is input into the heat collector diagnostic model, and the heat pump feature sequence is input into the heat pump diagnostic model; Based on the dual-branch feature extractor, one branch extracts region-sensitive features and the other branch extracts physically invariant features, which are then input into the feature fusion layer to obtain fused features; The fused features are forward propagated to the label classifier to obtain the heat collector failure probability and the heat pump failure probability, and the second classification loss and orthogonality loss are calculated. The dual-branch feature extractor and the label classifier are then backpropagated to update them. The physically invariant features are input into the domain classifier to obtain the source domain and target domain distinction results, and the domain classification loss is calculated. The domain invariant feature extractor in the dual-branch feature extractor is updated through backpropagation of the gradient reversal layer, and the parallel diagnostic model with updated structure is obtained through iterative training.

7. The physical-guided decoupling and parallel diagnostic method for a solar heat pump coupling system according to claim 3, characterized in that, Before generating the fault location report, the following steps are also included: When the probability of the heat collector failure and the probability of the heat pump failure meet a preset conflict condition, a parallel discrimination mechanism is invoked to perform diagnosis, obtain the fault determination result, and generate the fault location report.

8. The physical-guided decoupling and parallel diagnostic method for a solar heat pump coupling system according to claim 1, characterized in that, The function for calculating the health index is: , Formula 1; in, The aforementioned health index; The weights of the features; The current value of the feature; The reference value for the aforementioned feature; Dynamic fault threshold; The deviation of the feature; The fault level; This is the scaling factor.

9. A physically guided decoupling parallel diagnostic device for a solar heat pump coupling system, characterized in that, include: The data acquisition module is used to acquire monitoring data in real time; The data decoupling module is used to perform three-dimensional dynamic decoupling of the monitoring data to obtain the feature sequences corresponding to each subsystem; The parallel diagnostic module is used to input the feature sequences into the dynamic parallel diagnostic model to obtain the fault probability of each subsystem and generate a fault location report. The results output module calculates the health index of each subsystem based on the weight and deviation of each feature and the fault level corresponding to the fault probability, and outputs the fault diagnosis result in conjunction with each fault location report.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the physically guided parallel diagnostic method for decoupling a solar heat pump coupling system as described in any one of claims 1-8.

Images