A method, device and equipment for monitoring leakage fault of refrigerant of a ground source heat pump system

Abstract

This application discloses a method, apparatus, and equipment for monitoring refrigerant leakage faults in a ground source heat pump system. The method includes: acquiring and preprocessing sensor data for the current time period to obtain operating condition variable data and performance variable data; based on the operating condition variable data and a residual generator, obtaining a residual sequence between predicted performance variable data and the performance variable data; inputting the residual sequence into a cross-time period common feature encoder to extract potential decay features from the residual sequence; calculating health indicators based on the potential decay features; and determining the monitoring results for the current time period based on a preset dynamic monitoring method. Through this approach, this application can effectively eliminate interference from complex operating condition fluctuations, reduce false alarms and missed alarms related to leakage, and by constructing a cross-time period common feature encoder to extract physically meaningful and universally applicable potential decay features from the residual sequence, it can sensitively capture the weak decay signals at the initial stage of refrigerant leakage, achieving early warning of leakage faults.

Description

Technical Field

[0001] This application relates to the field of fault detection in ground source heat pump systems, and in particular to a method, apparatus and equipment for monitoring refrigerant leakage faults in ground source heat pump systems. Background Technology

[0002] Ground source heat pump systems, as a highly efficient and renewable energy utilization technology, are widely used in building heating and cooling. The refrigerant cycle within the ground source heat pump system, its core component, is crucial for energy conversion and transfer. Refrigerant leakage is one of the most common unit failures, leading to continuous performance degradation, compressor overheating and damage, and increased maintenance costs due to leak repair and recharging. Therefore, timely and accurate monitoring of refrigerant leakage is essential for ensuring long-term efficient and stable system operation and reducing energy consumption and maintenance costs.

[0003] Currently, the industry commonly uses threshold alarm methods based on direct measurement or performance derivation for refrigerant leak monitoring. For example, by monitoring key operating parameters such as suction or discharge pressure, evaporation or condensation temperature, and compressor power, fixed warning thresholds are set, and an alarm is triggered when the measured value exceeds the threshold. However, these existing methods still have significant limitations when practically applied to ground source heat pump systems, making it difficult to provide accurate and timely leak warnings. Summary of the Invention

[0004] This application provides a method, device, and equipment for monitoring refrigerant leakage faults in a ground source heat pump system, in order to solve the problems of low accuracy and difficulty in timely early warning of existing refrigerant leakage monitoring methods.

[0005] To solve the above-mentioned technical problems, one technical solution adopted in this application is: to provide a method for monitoring refrigerant leakage faults in a ground source heat pump system. This method includes: Acquire and preprocess sensor data for the current time period to obtain operating condition variable data and performance variable data; Based on the operating condition variable data and the residual generator, a residual sequence between the predicted performance variable data and the performance variable data is obtained; The residual sequence is input into a cross-time period common feature encoder to extract potential decay features from the residual sequence; Based on the potential decay characteristics, health indicators are calculated, and the monitoring results for the current period are determined based on a preset dynamic monitoring method.

[0006] In one optional embodiment of this application, the step of acquiring and preprocessing the sensor data for the current time period to obtain operating condition variable data and performance variable data includes: Acquire sensor data for the current time period and calculate derived variable data to obtain multiple sets of variable data; Based on the sliding window method, the multiple sets of variable data are divided into multiple sequence segments; Based on the preset stationarity determination mechanism, the corresponding statistical characteristic values ​​are calculated for each key variable that is sensitive to the operating conditions in each sequence segment. In response to any of the key variables having a statistical feature value greater than a preset statistical threshold, the corresponding sequence segment is filtered out. The filtered sequence segments are integrated and divided to obtain the operating condition variable data and the performance variable data.

[0007] In one optional embodiment of this application, the residual generator includes at least one convolutional layer, one fully connected layer, and multiple residual shrink blocks; The step of obtaining the residual sequence between the predicted performance variable data and the performance variable data based on the operating condition variable data and the residual generator includes: The operating condition variable data is input into the residual generator, and the input feature map is obtained by high-dimensional mapping through the convolutional layer. The input feature map is input into the residual shrinkage block, and the output feature map is obtained based on the adaptive threshold and soft threshold functions. The output feature map is input into the fully connected layer to obtain the predicted performance variable data layer, and the residual sequence between the predicted performance variable data and the performance variable data is calculated.

[0008] In an optional embodiment of this application, the training process of the cross-time period common feature encoder is as follows: The historical residual sequence obtained from sensor data based on historical time periods is segmented to obtain multiple historical residual segments; Based on the empirical trend function of leakage rate, a target weak label is set for each of the historical residual segments. Each of the historical residual segments is input into the cross-time period common feature encoder to extract historical potential decay features from each of the historical residual segments; The historical potential decay features are mapped to the corresponding single-time decoder to generate predicted weak labels for each residual segment. The loss value is calculated based on the target weak label and the predicted weak label. The loss value is backpropagated to optimize the cross-time common feature encoder and each of the single-time decoders to obtain the trained cross-time common feature encoder.

[0009] In an optional embodiment of this application, the empirical trend function of the leakage rate is: , formula 1; in, The target weak label of the i-th historical residual segment; This represents the initial refrigerant balance of the unit. This refers to the refrigerant leakage rate; This is the runtime of the unit from its initial state to the i-th historical residual segment.

[0010] In one optional embodiment of this application, the loss value includes at least one or more of the following: prediction loss, physical consistency loss, and monotonicity regularization term; The function for calculating the prediction loss is: , formula 2; in, The predicted loss; The target weak label for the i-th historical residual segment; The predicted weak label for the i-th historical residual segment; The calculation function for the physical consistency loss is: , formula 3; in, This refers to the physical consistency loss; The historical potential decay feature of the i-th historical residual segment; The residual feature value is obtained based on the historical potential decay characteristics; N is the number of historical residual segments corresponding to the historical potential decay characteristics; The function for calculating the monotonicity regularization term is: , formula 4; in, This refers to the monotonicity regularization term; To correct the linear unit function; This is a hyperparameter.

[0011] In an optional embodiment of this application, after extracting the historical potential attenuation features from each of the historical residual segments, the method further includes: Based on the historical potential decay characteristics, the historical health index of each historical residual segment is calculated to obtain the historical health index sequence. The historical health indicator sequence is dynamically monitored based on a preset dynamic monitoring method, and the historical leakage statistics are calculated. Based on the historical leakage statistics, a leakage warning threshold and a maintenance warning threshold are determined, wherein the leakage warning threshold is less than the maintenance warning threshold.

[0012] In an optional embodiment of this application, calculating health indicators based on the potential attenuation characteristics and determining the monitoring results for the current time period includes: Based on principal component analysis, the potential attenuation features are dimensionality reduced to obtain the dimensionality-reduced projected coordinates. The health index is calculated by weighting and combining the dimensions of the projected coordinates. Based on the health indicators and the preset dynamic monitoring method, calculate the leakage statistics for the current time period; If the leakage statistics value for the current time period is greater than the leakage warning threshold, the monitoring result is a Level 1 leakage, triggering a leakage warning; If the leakage statistics value for the current time period exceeds the maintenance warning threshold, the monitoring result is a level 2 leakage, triggering a maintenance warning.

[0013] To solve the above-mentioned technical problems, another technical solution adopted in this application is: to provide a refrigerant leakage fault monitoring device for a ground source heat pump system, the device comprising: The data processing module is used to acquire and preprocess the sensor data for the current time period to obtain operating condition variable data and performance variable data; The residual generation module is used to obtain the residual sequence between the predicted performance variable data and the performance variable data based on the operating condition variable data and the residual generator. The feature extraction module is used to input the residual sequence into the cross-time common feature encoder and extract potential decay features from the residual sequence; The indicator calculation module is used to calculate health indicators based on the potential decay characteristics and determine the monitoring results for the current period based on a preset dynamic monitoring method.

[0014] 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 ground source heat pump system refrigerant leakage fault monitoring method.

[0015] The beneficial effects of this application are as follows: Unlike existing technologies, this application discloses a method, apparatus, and equipment for monitoring refrigerant leakage faults in a ground source heat pump system. This method divides sensor data into operating condition variable data and performance variable data. Through a residual generator, it constructs a residual sequence between the predicted performance variable data obtained from the operating condition variable data and the actual performance variable data. This effectively eliminates the interference of complex operating condition fluctuations, improves the accuracy of subsequent analysis, and fundamentally reduces the false alarm and false alarm rates of leakage. By constructing a cross-time period common feature encoder, it extracts physically meaningful and cross-time period universal potential attenuation features from the residual sequence, which can sensitively capture the weak attenuation signal at the initial stage of refrigerant leakage, improving the timeliness of leakage fault early warning and achieving early warning of leakage faults. Attached Figure Description

[0016] 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 refrigerant leakage fault monitoring method for a ground source heat pump system provided in this application; Figure 2 This is a schematic diagram of the residual generator calculating the residual sequence in an embodiment of the refrigerant leakage fault monitoring method for a ground source heat pump system provided in this application; Figure 3 This is a schematic diagram of the process of extracting historical potential attenuation features using a cross-time period common feature encoder according to an embodiment of the ground source heat pump system refrigerant leakage fault monitoring method provided in this application; Figure 4 This is a schematic diagram of the fault monitoring results based on CUSUM, representing an embodiment of the ground source heat pump system refrigerant leakage fault monitoring method provided in this application. Figure 5 This is a schematic diagram of the fault monitoring results based on EWMA, representing an embodiment of the ground source heat pump system refrigerant leakage fault monitoring method provided in this application. Figure 6 This is a schematic diagram of an embodiment of the refrigerant leakage fault monitoring device for a ground source heat pump system provided in this application; Figure 7 This 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

[0017] 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.

[0018] 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.

[0019] 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.

[0020] This application provides a method for monitoring refrigerant leakage faults in a ground source heat pump system. (See reference...) Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the refrigerant leakage fault monitoring method for a ground source heat pump system provided in this application. The method includes: S10: Acquire sensor data for the current time period and preprocess it to obtain operating condition variable data and performance variable data.

[0021] A ground source heat pump system is a highly efficient and energy-saving air conditioning system that utilizes shallow underground geothermal resources (also known as geothermal energy, including groundwater, soil, or surface water) to provide both heating and cooling. Its basic principle is to transfer heat energy from a low-temperature heat source to a high-temperature heat source by inputting a small amount of electrical energy.

[0022] Ground source heat pump systems typically consist of three main circulation loops: Ground source side circulation loop: also known as underground heat exchange section, usually includes buried pipes or groundwater wells, which exchange heat with soil or groundwater through a circulating medium (usually water or antifreeze).

[0023] Heat pump unit circulation loop: This is the core of the system, including major components such as the compressor, condenser, evaporator, and expansion valve. The refrigerant absorbs and releases heat through a phase change within this loop.

[0024] Load-side circulation loop: also known as the user terminal section, usually includes fan coil units, radiators, etc., which use circulating water to deliver cooling or heating to the indoor space.

[0025] In cooling mode, the ground-source heat pump's circulating loop absorbs cold energy from the ground and transfers it to the load-side circulating loop via the heat pump unit, ultimately providing cooling for the indoor environment. In heating mode, the process is reversed. Because the refrigerant in the heat pump unit circulates under high pressure, and the system has complex piping and numerous connection points, refrigerant leakage may occur during long-term operation due to corrosion, vibration, aging, and other reasons. Refrigerant leakage not only reduces system efficiency and increases energy consumption but may also impact the environment and even cause serious problems such as compressor failure. Therefore, real-time and accurate monitoring of refrigerant leakage in ground-source heat pump systems is of significant practical importance.

[0026] In this application, various sensors are deployed at key nodes of the ground source heat pump system to acquire real-time operational data recorded by the sensors, including but not limited to the following types of sensors: Temperature sensor: Used to monitor temperature changes at key locations in the system, including the return water temperature T on the load side. lr Geological source side outlet water temperature T go Intake temperature T s Evaporation temperature T e Condensation temperature T c , liquid outlet temperature T l wait; Pressure sensor: Used to monitor the operating conditions of the compressor, a core component of the heat pump unit, including the compressor suction pressure P. e and compressor discharge pressure P c wait; Flow sensor: Used to monitor the flow rate of circulating media, including the instantaneous flow rate F on the load side. l wait; Power sensor: Used to monitor the energy consumption of the system, including compressor power (kW), etc.

[0027] All the aforementioned sensors are connected to a data acquisition unit, which samples and records data from each sensor at a preset sampling frequency. The preset sampling frequency can be flexibly set according to the actual application scenario and requirements, such as 1 minute, 1.5 minutes, or 2 minutes. For example, for a typical medium-sized commercial system, a sampling frequency of 1 minute is usually a good balance among the aforementioned factors, capturing abnormal parameter changes caused by leakage without creating excessive data storage and processing pressure.

[0028] In this application, sensor data for the current time period is acquired and preprocessed to obtain operating condition variable data and performance variable data, including: S11: Obtain sensor data for the current time period and calculate derived variable data to obtain multiple sets of variable data.

[0029] In this application, all sensor data collected within a preset time interval is acquired, that is, all sensor data collected within a continuous time period from the current moment back, and this data is recorded as the sensor data for the current time period. This time interval can be adjusted according to actual needs and is not specifically limited; for example, all sensor data from the past 60 minutes can be collected every 60 minutes.

[0030] The sensor data collected above includes, but is not limited to, the host load L. g , Load side return water temperature T lr Geological source side outlet water temperature T go Instantaneous flow rate F on the load side l Intake temperature T s Evaporation temperature T e Condensation temperature T c , liquid outlet temperature T l Compressor suction pressure P e and compressor discharge pressure P c Multiple sets of variable data, such as compressor power (kW) and on / off status (R).

[0031] Simultaneously, the data corresponding to the derived variables are calculated based on the directly collected sensor data. Specifically, based on the inhalation temperature T... s and evaporation temperature T e Calculate the derived variable, intake superheat T suc The calculation formula is: intake superheat T suc =Intake temperature T s - Evaporation temperature T e Based on condensation temperature T c and outlet temperature T l Calculate the derived variable, supercooling T sub The calculation formula is: subcooling T sub =Condensation temperature Tc -Liquid outlet temperature T l The intake temperature, evaporation temperature, condensation temperature, and liquid outlet temperature can all be obtained through corresponding sensors.

[0032] Thus, by combining directly acquired sensor data with calculated derived variable data, the variable data available for the current time period can include, but is not limited to, host load L. g , Load side return water temperature T lr Geological source side outlet water temperature T go Instantaneous flow rate F on the load side l Compressor suction pressure P e Compressor discharge pressure P c Inhalation superheat T suc Supercooling T sub The application uses 10 sets of variable data, such as compressor power (kW) and on / off status (R), as examples for further explanation.

[0033] For the multiple sets of variable data obtained in the current time period, further data cleaning is performed, including removing invalid data and shutdown data.

[0034] Specifically, invalid timestamps and columns containing null values ​​are removed from multiple sets of variable data for the current time period to eliminate invalid data. By removing this incomplete data, a clean and complete dataset can be constructed, thereby avoiding the impact of data quality issues on the accuracy of subsequent leak detection algorithms and ensuring the reliability of the analysis results.

[0035] In the current time period, data sets of multiple variable data that meet the following criteria are considered as shutdown data (i.e., R=0) and whose compressor power (kW) is below a preset percentage of the unit's full load rate (e.g., 0.5%, 1%, or 2%) and are therefore discarded. Only data from the unit's operating period are retained. Furthermore, the remaining variable data sets, excluding those related to the shutdown / on / off status, are pruned to 0.1% to 99.9% quantiles to eliminate the impact of extreme outliers.

[0036] Data records in the "off" state are removed from the current time period's multiple sets of variable data. The criteria for determining the "off" state are: the value of the on / off state variable R is 0, and the compressor power (kW) is below a preset full load rate threshold (e.g., 0.5%, 1%, or 2%). Furthermore, quantile pruning is performed on the remaining numerical variables in the current time period's multiple sets of variable data, excluding the on / off state. Specifically, the value range of each variable can be limited to between its 0.1% and 99.9th percentile to eliminate the influence of extreme outliers.

[0037] For multiple sets of variable data that have undergone data cleaning, normalization is performed to eliminate scale differences between the data, resulting in normalized sets of variable data. Specifically, min-max normalization can be chosen to scale the numerical range of the multiple sets of variable data to the [0,1] interval. The normalization formula is as follows: , Formula 5; in, The value of variable x at time point k is the normalized value; k is a specific time point; K is the total number of time points in each group of variable data. It is the value of the j-th variable in a set of variables; n is the total number of variables, such as 10; i is the traversal index, representing any one of the time points from the 1st to the Kth time point; Let x be the minimum value of the variable. Let x be the maximum value of variable x.

[0038] S12: Based on the sliding window method, multiple sets of variable data are divided into multiple sequence segments.

[0039] For the normalized multi-set variable data obtained in step S11, the time series sliding window method is used for reconstruction, which discretizes the continuous time series into local observation windows, thereby dividing the multi-set variable data into multiple sequence segments.

[0040] In one specific embodiment, taking a window length of L=60 (corresponding to a time dimension of 60 minutes) and a sliding step size of S=10 (corresponding to a time dimension of 10 minutes) as an example, multiple local windows are constructed in chronological order. Each window is a sequence segment, and the sequence segment is used as a sample for subsequent data processing.

[0041] Each window is denoted as Its definition is as follows: , Formula 6; in, Let be the sequence segment consisting of all data points within the i-th window, i.e., from... The data points contained within the initial continuous length L; These are the normalized data points; Let be the starting index of the i-th window; This is the starting data point for the i-th window. Adjacent windows... and The start times differ by S.

[0042] S13: Based on the preset stationarity determination mechanism, calculate the corresponding statistical characteristic values ​​for multiple key variables that are sensitive to operating conditions in each sequence segment.

[0043] According to stochastic process theory, when a system is in steady-state operation, its observed variables should approximately follow a Gaussian distribution, and its first moment (mean) and second central moment (variance) should remain time-invariant. Based on this, this application presupposes a stationarity determination mechanism to determine the stationarity of multiple sequence segments obtained in step S12 above, so as to extract quasi-stationary segments.

[0044] First, a multidimensional feature vector is constructed by selecting several key variables that are most sensitive to fluctuations in operating conditions. Taking the load-side return water temperature T as an example... lr Geological source side outlet water temperature T go Compressor suction pressure P e Compressor discharge pressure P c and load-side instantaneous flow rate F l These five key variables will be used as examples to illustrate this.

[0045] Then, for any sample The statistical characteristic values ​​corresponding to each key variable within the sample are used to determine stationarity. Specifically, the statistical characteristic values ​​can be the sample mean. (First-order raw moment estimation) and sample variance (Second-order central moment estimation), the calculation formula is as follows: , , Formula 7; in, Let be the sample mean of the j-th variable within the i-th window; Let L be the sample variance of the j-th variable within the i-th window; L is the window length. For the j-th variable in the multiple sets of variables within the i-th window One data point; =0, 1, ..., L-1.

[0046] In some embodiments, the statistical characteristic values ​​calculated in the preset stationarity determination mechanism can be statistical characteristic values ​​such as unit roots, in addition to variance. Furthermore, the preset stationarity determination mechanism can also directly utilize unsupervised clustering algorithms to automatically identify stationarity.

[0047] S14: In response to any key variable’s statistical feature value being greater than a preset statistical threshold, filter the corresponding sequence segment.

[0048] The statistical characteristic values ​​corresponding to each key variable calculated in step S13 are compared with the preset variance thresholds corresponding to each key variable to determine whether there are significant non-stationary dynamics (such as start-stop oscillations or load transients) within the window. The preset variance thresholds are used to determine whether significant non-stationary dynamics exist within the window. As a critical value used to determine whether the data of each key variable is stable, it actually defines the statistically permissible fluctuation range of the system's steady-state operation. It can be set based on historical steady-state operation data, for example, it can be set to... No restrictions are imposed here.

[0049] When the sample variance of any key variable within the window If the window contains significant non-stationary dynamics (such as start-stop oscillations or load transients), it is considered to belong to the non-steady-state interval and is truncated and discarded; otherwise, if the sample variance of any key variable within the window exceeds the stationary dynamics (such as start-stop oscillations or load transients), it is considered to belong to the non-steady-state interval and is discarded. If the window (sequence segment) is in a quasi-steady-state region, then the data in that window is retained.

[0050] In one specific embodiment, the preset variance thresholds for each key variable are shown in Table 1 below. These preset variance thresholds actually define the allowable fluctuation range of the system's steady-state operation in a statistical sense.

[0051]

[0052] Table 1. Preset variance thresholds for each key variable The minimum and maximum values ​​of the sample variance for each key variable within the window are shown in Table 2 below.

[0053]

[0054] Table 2. Minimum and maximum sample variances of each key variable within the window. A comparison between Table 1 and Table 2 above shows that the actual variance fluctuations of each key variance are greater than the preset quasi-steady-state threshold. Taking the load-side return water temperature T as an example... lr Taking this key variable as an example, the load-side return water temperature T lr The actual maximum variance is 0.118, which is greater than its quasi-steady-state threshold of 0.045. By limiting most of the load-side return water temperature data within the quasi-steady-state threshold, the interference of severe operating condition fluctuations caused by start-up, shutdown, load changes, etc. can be eliminated, ensuring that subsequent operations are based on quasi-steady-state data.

[0055] Unlike existing technologies where frequent changes in operating conditions such as load and ambient temperature of ground source heat pump units lead to fluctuations in performance variables (such as power and temperature), making traditional fixed threshold methods prone to false alarms or missed alarms, the refrigerant leakage fault monitoring method for ground source heat pump systems provided in this application filters out non-steady-state sequence segments through the aforementioned preset stability judgment mechanism, retaining all quasi-steady-state sequence segments for subsequent data processing. This effectively eliminates interference from non-steady-state data, making subsequent refrigerant leakage monitoring based on quasi-steady-state data more accurate and reliable. Compared to the traditional fixed threshold method, this significantly improves the accuracy and sensitivity of refrigerant leakage monitoring, reduces false alarms and missed alarms, and provides strong protection for the safe and stable operation of ground source heat pump systems.

[0056] S15: Integrate and divide the filtered sequence segments to obtain operating condition variable data and performance variable data.

[0057] For the high-quality sequence segments filtered in step S14, the sequence segments are integrated and variable sets are divided. Specifically, the variable data describing the external environment or unit control settings are divided into operating condition variable data, including: main unit load L. g , Load side return water temperature T lr Geological source side outlet water temperature T go And load-side instantaneous flow rate Fl and other variable data. Variables describing changes in the unit's internal operating performance are classified as performance variables, including: compressor suction pressure P. e Compressor discharge pressure P c Inhalation superheat T suc Supercooling T sub And variable data such as compressor power (kW).

[0058] S20: Based on the operating condition variable data and the residual generator, obtain the residual sequence of the predicted performance variable data and the performance variable data.

[0059] A residual generator is used to construct the mapping relationship between operating condition variables and performance variables, generating residual sequences stripped of operating condition fluctuations under dimensionless conditions. Specifically, the residual generator can be implemented using Deep Residual Shrinkage Networks (DRSN) models, feedforward neural networks, support vector machines, random forests, or more complex time series models (such as Long Short-Term Memory networks), without specific limitations. Among them, the DRSN architecture can utilize its unique soft thresholding mechanism to achieve information bottleneck constraints, predicting performance variables by inputting operating condition variables, effectively filtering out irrelevant noise caused by operating condition fluctuations while retaining key prediction features. This application uses the DRSN model as an example of a residual generator for illustration.

[0060] Reference Figure 2 , Figure 2 This is a schematic flowchart illustrating the calculation of the residual sequence by the residual generator in an embodiment of the refrigerant leakage fault monitoring method for a ground source heat pump system provided in this application. In the residual generator constructed by the DRSN model, the input end receives all quasi-steady-state operating condition variable data, and the output end outputs the residual sequence between the predicted performance variable data and the actual performance variable data obtained by S10. This constitutes a many-to-many prediction mapping relationship.

[0061] The residual generator contains at least one convolutional layer, one fully connected layer, and multiple residual shrinkage blocks (RSBs). The core unit of the DRSN model is the residual shrinkage block, which is composed of multiple stacked residual shrinkage blocks and contains at least one convolutional layer to map the operating condition variable data to the input feature map, and one fully connected layer to map the output features to the predicted performance variable.

[0062] In this application, based on the operating condition variable data and the residual generator, a residual sequence of the predicted performance variable data and the performance variable data is obtained, including: S21: Input the operating condition variable data into the residual generator, and obtain the input feature map through high-dimensional mapping of the convolutional layer.

[0063] The quasi-steady-state operating condition variable data obtained in step S10 are input into the residual generator, referring to... Figure 2 The operating condition variable data are processed in the residual generator through one convolution, batch normalization and ReLU activation to generate input feature maps.

[0064] Specifically, the quasi-steady-state operating condition variable data first undergoes a convolution to perform a high-dimensional mapping to obtain the input feature map, calculated as follows: , formula 8; in, Input feature map; This is the weight matrix; For input operating condition variable data; For offset top; This is the ReLU activation function.

[0065] ReLU activation introduces non-linearity by using a piecewise linear function to activate only positive values ​​and suppress negative values, thereby improving the network's expressive power. The expression for ReLU activation is as follows: , formula 9; in, For input operating condition variable data; The function is for finding the maximum value.

[0066] In this process, batch normalization is used to accelerate and stabilize the training process. Before the ReLU activation function, the input data (i.e., the working condition variable data) of each mini-batch is standardized so that the mean of the input data is close to 0 and the variance is close to 1.

[0067] S22: Input the input feature map into the residual shrinking block, and obtain the output feature map based on the adaptive threshold and soft threshold function.

[0068] Reference Figure 2 Input feature map Input to residual shrinkage block, input feature map go through The execution process of each residual shrinking block is calculated using the following formula: , formula 10; in, For the first Convolution, batch normalization, and activation operations in a residual shrink block; This is a soft thresholding function; This is the output feature map of the nth residual shrinkage block.

[0069] Based on an attention mechanism, an adaptive threshold τ is learned for each channel of the input feature map. This threshold is then used in a soft thresholding function to shrink and denoise the main path feature map (obtained by double convolution and batch normalization of the input feature maps entering each residual shrinking block), resulting in a shrunken feature map. This shrunken feature map is then added element-wise to the input feature map to obtain the final output feature map. The soft thresholding function suppresses weak and unimportant noise by subtracting the absolute value of each feature from its corresponding threshold τ and setting any values ​​less than zero to zero. This effectively filters out unimportant features while retaining crucial information useful for prediction, thus improving the effectiveness of feature representation. The mathematical expression is as follows: , Formula 11; Where x is the main path feature map; y is the shrunk feature map; and τ is the adaptive threshold.

[0070] S23: Input the output feature map into the fully connected layer to obtain the predicted performance variable data, and calculate the residual sequence between the predicted performance variable data and the performance variable data.

[0071] Reference Figure 2 The output feature map is obtained after passing through N residual shrinking blocks. Then, global average pooling is used to aggregate the output feature maps into a one-dimensional feature vector in the spatiotemporal dimension. The calculation formula is as follows: , Formula 12; in, This is a one-dimensional feature vector after global average pooling. The time step is the number of discrete sampling points in the output feature map along the time dimension; C is the number of channels in the output feature map. For the output feature map In the equation, the value of the c-th channel at the t-th time position is given.

[0072] Pooled feature vectors The predicted performance variable data is obtained by mapping to the performance variable space through a fully connected layer. The calculation formula is as follows: , Formula 13; in, For predicting performance variable data; This is the weight matrix of the fully connected layer; This is the offset top of the fully connected layer.

[0073] After obtaining the predicted performance variable data, residual calculation is performed between it and the actual performance variable data obtained in step S10. The residual sequence is obtained by the difference between the actual performance variable data and the predicted performance variable data. The calculation formula is as follows: , Formula 14; in, The residual sequence is y; the actual performance variable data is y. For predicting performance variable data.

[0074] For training the residual generator constructed using the DRSN model, historical operating condition and performance variable data can be used as the training set, with mean squared error as the loss function. The hyperparameters can be set as follows: 100 training epochs, a learning rate of 0.001, a batch size of 512, and the optimizer Adam. It should be noted that these hyperparameters are merely illustrative settings and can be adjusted according to specific needs in actual applications.

[0075] During training, the network parameters of the residual generator are optimized using a loss function to achieve accurate learning of the performance variables under conditions free from interference, ensuring that the output residual sequence can eliminate interference from operating condition fluctuations. The loss function of the residual generator is as follows: , Formula 15; in, This represents the loss value of the residual generator; This represents the actual performance variable data for the i-th sample; represents the prediction performance variable data for the i-th sample; N is the number of samples in a training batch.

[0076] S30: Input the residual sequence into the cross-time common feature encoder to extract potential decay features from the residual sequence.

[0077] In this application, physical knowledge-guided feature extraction is performed on the residual sequence. This feature extraction process is implemented through a trained cross-time period common feature encoder. The training process of the cross-time period common feature encoder is as follows: S31: Segment the historical residual sequence obtained from sensor data based on historical time periods to obtain multiple historical residual segments.

[0078] To ensure the data is collected during refrigerant leakage, sensor data from the ground source heat pump unit during its operation period can be collected based on historical unit maintenance records. This yields sensor data from multiple historical time periods, serving as both training and test sets. Data preprocessing generates operating condition and performance variable data for multiple historical time periods. A residual generator then generates multiple historical residual sequences, which serve as training data for a cross-time period common feature encoder. The specific process for obtaining the historical residual sequences from the sensor data from historical time periods can be found in steps S10-S20 above, and will not be repeated here.

[0079] Next, the historical residual sequences are segmented using the sliding window method. Specifically, the window length can be set to L=180 (180 minutes) and the sliding step size to S=60 (60 minutes). Each historical residual sequence can yield multiple historical residual segments.

[0080] S32: Based on the empirical trend function of leakage rate, set a target weak label for each historical residual segment.

[0081] Refrigerant leakage in ground source heat pump units exhibits a non-linear deceleration trend, characterized by an initial rapid leak followed by a slower leak. Specifically, in the initial stage of leakage, the system's internal pressure is significantly higher than the external ambient pressure, resulting in a larger pressure difference and a faster leakage velocity. However, as the leakage progresses, the internal system pressure gradually decreases, reducing the pressure difference and slowing the leakage velocity. Based on this, this application constructs an empirical trend function for the leakage rate to describe the refrigerant leakage rate within the unit as a function of cumulative operating time. The general evolution relationship is determined, and the refrigerant margin is used as a weak label for the historical residual range.

[0082] In this application, the empirical trend function for leakage rate is: , formula 1; in, The refrigerant margin of the i-th historical residual segment is the target weak label of the i-th historical residual segment. This represents the initial refrigerant balance of the unit. The refrigerant leakage rate can be calculated based on the duration of the historical residual period and the amount of refrigerant leakage. This is the runtime of the unit from its initial state to the i-th historical residual segment.

[0083] In some embodiments, the above-mentioned empirical trend function of leakage rate is not limited to the form of square root (as in Formula 1), but can also be linear, exponential or logarithmic to fit different types of refrigerant leakage processes. It can be flexibly adjusted according to the actual refrigerant leakage type. Here, only Formula 1 is used as an example for illustration, and the specific form of the empirical trend function of leakage rate is not limited.

[0084] S33: Input each historical residual segment into the cross-time period common feature encoder to extract historical potential decay features from each historical residual segment.

[0085] The cross-time common feature encoder is designed to extract common temporal patterns of refrigerant leakage. Its architecture consists of a temporal convolutional network composed of multiple layers of dilated convolutional blocks and a fully connected network.

[0086] Reference Figure 3 , Figure 3 This is a schematic diagram illustrating the process of extracting historical potential attenuation features using a cross-time period common feature encoder, according to an embodiment of the ground source heat pump system refrigerant leakage fault monitoring method provided in this application. The workflow of the cross-time period common feature encoder for extracting historical potential attenuation features is as follows: First, the historical residual segments with target weak labels are input into the cross-time common feature encoder. These historical residual segments are passed through a convolutional layer to perform preliminary feature coupling and abstraction of all residual variables in the time dimension, thereby obtaining an initial feature map.

[0087] Subsequently, the initial feature maps are fed into stacked deep dilated convolutional blocks to extract the final feature maps. This deep network, by employing an exponentially growing dilation coefficient, achieves a large receptive field with its deep network structure while maintaining high computational efficiency, effectively capturing long-range dependencies and leakage trends across multiple unit operation periods in the residual data. Because the cross-period common feature encoder shares all weights and updates synchronously when processing data from all unit operation periods, this ensures that the extracted final feature maps represent the common, essential leakage signals across different time periods.

[0088] Finally, the extracted final feature map is flattened and mapped to a low-dimensional latent space through a fully connected layer to generate the final historical latent decay feature Si.

[0089] Historical potential decay characteristics This reflects the common timing pattern of refrigerant leakage in the unit, and its calculation formula is as follows: , Formula 16; in, The historical potential decay characteristics of the i-th historical residual segment; For weight parameters Parameterized cross-time common encoder; This is the i-th historical residual segment.

[0090] In some embodiments, the cross-temporal common feature encoder is not limited to the form of a temporal convolutional network and a fully connected network consisting of multiple layers of dilated convolutional blocks, as described above. It can also employ architectures such as recurrent neural networks or Transformers to capture dependencies over longer time periods. The cross-temporal common feature encoder aims to extract potential decay features that accurately reflect the degree of leakage and strictly follow physical laws. The form of the cross-temporal common feature encoder is not specifically limited.

[0091] S34: Map the historical potential decay features to the corresponding single-time decoder to generate predicted weak labels for each residual segment.

[0092] After each overall maintenance, the performance of a ground source heat pump unit can be restored to its initial state. However, due to minor changes in assembly precision and environmental factors, slight differences may still exist in its initial state. To improve the versatility of the cross-time common feature encoder and enhance its robustness to these subtle differences, a single-time-specific decoder is introduced to construct a network structure of "cross-time common feature encoder + single-time-specific decoder". Each single-time-specific decoder consists of a fully connected network, and its core task is to map the potential attenuation features Si output by the cross-time common feature encoder back to the corresponding unit operating period, generating predicted weak labels for each residual segment. This combined architecture effectively balances the model's versatility and specificity through the synergy of common feature extraction and specific decoding, ensuring that it can accurately capture leaked signals under different initial states.

[0093] Predicting weak labels The calculation method is as follows: , Formula 17; in, The predicted weak label for the i-th sample; These are the decoder parameters corresponding to the operating period of the kth unit; This is the decoder corresponding to the operating period of the k-th unit; The historical potential decay characteristics of the i-th historical residual segment; This indicates determining which unit's operating period the i-th historical residual segment belongs to.

[0094] By employing a network structure of "cross-time period common feature encoder + single-time period specific decoder," physical knowledge-guided feature extraction and decoupling for adaptation to differences across unit operating periods are achieved. The cross-time period common feature encoder is responsible for extracting historical potential attenuation features with clear physical significance across unit operating periods. The single-time-specific decoder is responsible for fine-tuning by incorporating historical potential decay features. The applicability of the cross-period common feature encoder is monitored by mapping it to the operation period of each unit, thereby enhancing its versatility and adaptability. Ultimately, the cross-period common feature encoder can remove noise from the input historical residual segments and extract historical potential decay features with real-world physical significance and cross-period applicability. .

[0095] S35: Calculate the loss value based on the target weak label and the predicted weak label, and backpropagate the loss value to optimize the cross-time common feature encoder and each single-time decoder to obtain the trained cross-time common feature encoder.

[0096] During training, the loss value is backpropagated starting from the output layer. This is done by adjusting each weight parameter (including the encoder weights) in the "cross-time common feature encoder + single-time specific decoder" network. and all decoder weights This minimizes the loss value, ultimately maximizing the extracted historical potential decay features. It can accurately reflect the degree of leakage and strictly follow the laws of physics.

[0097] To deeply integrate physical knowledge into the "cross-time common feature encoder + single-time specific decoder" network, a physically guided loss function was designed to calculate the loss value. This loss function includes at least one or more of the following: prediction loss function, physical consistency loss function, and monotonic regularization function, which together ensure the extraction of historical potential decay features. It accurately predicts the refrigerant leakage stage while strictly conforming to the laws of physics.

[0098] In this application, the loss value includes at least the predicted loss. Physical consistency loss and monotonicity regularization One or more of these; Predicting losses Predicting weak labels for a single time-specific decoder And the target weak label calculated based on the empirical trend function of refrigerant leakage rate The mean square error between them. Predicted loss. The calculation function is: , formula 2; in, To predict losses; The target weak label for the i-th historical residual segment; The predicted weak label for the i-th historical residual segment; Let be the prediction loss of the i-th historical residual segment, which is the mean square error of the i-th historical residual segment.

[0099] Physical consistency loss These are the residual feature values ​​derived internally by the cross-time common feature encoder. The theoretical value calculated using the empirical trend function of refrigerant leakage rate The differences between them. Physical consistency loss. The calculation function is: , formula 3; in, This represents a loss of physical consistency. The historical potential decay characteristics of the i-th historical residual segment; For physical mapping network functions; is the margin characteristic value obtained based on historical potential decay characteristics, that is, the health characteristic value used to characterize the refrigerant margin; N is the number of historical residual segments corresponding to the historical potential decay characteristics.

[0100] Monotonicity regularization term Based on the prior knowledge that "refrigerant leakage is irreversible," this is achieved by penalizing historical potential degradation characteristics. The inverse transformation is used to extract the historical potential decay feature sequence. It exhibits a monotonic trend related to time. (Monotonicity regularization term) The calculation function is: , formula 4; in, It is a monotonicity regularization term; To correct the linear unit function; For hyperparameters; The historical potential decay characteristics of the i-th historical residual segment; This represents the historical potential decay characteristics of the (i-1)th historical residual segment; This is the loss due to the one-way regularization term.

[0101] In one specific embodiment, the total loss function of the "cross-time common feature encoder + single-time specific decoder" network includes a prediction loss function, a physical consistency loss function, and a monotonic regularization term function, calculated as follows: , Formula 18; in, Total loss; To predict losses; This represents a loss of physical consistency. It is a monotonicity regularization term.

[0102] Furthermore, the various losses in Formula 18 above are also assigned corresponding weights according to actual needs, and the total loss is obtained by weighted summation. No specific limitations are specified here.

[0103] In the process of training a cross-time period common feature encoder using historical residual segments, the cross-time period common feature encoder and each single-time period decoder are optimized through backpropagation loss values ​​to obtain a well-trained cross-time period common feature encoder. The hyperparameter settings for the cross-time period common feature encoder can be as follows: 100 training epochs, a learning rate of 0.001, a batch size of 512, and the optimizer Adam. It should be noted that the above hyperparameters are only exemplary settings and can be adjusted according to specific needs in actual applications.

[0104] By introducing a physically guided loss function (specifically, a physical consistency loss and a monotonicity regularization term), the extracted potential decay features are ensured to be not only valid in terms of data, but also physically consistent with the refrigerant leakage process, thereby enhancing the credibility and interpretability of the model's decisions.

[0105] In this application, for the residual sequence obtained in step S20 above, before inputting the residual sequence into the cross-time common feature encoder to extract potential attenuation features from the residual sequence, the following steps are also included: Determine whether the residual sequence of the current time period has reached the preset duration threshold; In response to the failure to reach the preset duration threshold, the residual sequence of the current time period is added to the buffer; The residual sequence in the buffer is input into the cross-time common feature encoder until the residual sequence in the buffer reaches the preset duration threshold.

[0106] Since some data (such as invalid data, shutdown data and non-stationary data) in the data obtained in the current time period were removed in the aforementioned step S10, the length of the residual sequence obtained in step S20 in the time dimension may not match the preset duration threshold that can be used as a window to input to the cross-time period common feature encoder.

[0107] Therefore, after obtaining the residual sequence output by the residual generator, it is first determined whether the residual sequence of the current time period reaches the preset duration threshold. If so, the window of the preset duration threshold can be directly extracted and input into the cross-time period common feature encoder. If not, the residual sequence of the current time period is added to the buffer, waiting to be combined with the residual sequence of the next time period, or with the residual sequence of the previous time period in the buffer, until a residual sequence that reaches the preset duration threshold can be obtained, and then it is input into the cross-time period common feature encoder. The value of the preset duration threshold is set according to the historical residual segment. If the duration of the historical residual segment is 180 minutes, the preset duration threshold can be set to 180 minutes.

[0108] The residual sequence is input into a pre-trained cross-time period common feature encoder to extract the potential decay features of the current time period from the residual sequence. .

[0109] Unlike existing technologies that rely heavily on manual experience to set features and fail to automatically extract performance degradation features weakly correlated with operating conditions, thus hindering early warning, this application provides a ground source heat pump system refrigerant leakage fault monitoring method. This method inputs a residual sequence, effectively stripped of interference from complex operating condition fluctuations and unsteady-state processes, into a trained cross-time period common feature encoder. Through physical knowledge-guided feature extraction, it automatically extracts potential degradation features for the current time period. Without requiring manual experience to set these features, which are weakly correlated with operating conditions, it more accurately reflects the performance degradation of the ground source heat pump system. This allows for timely warnings in the early stages of refrigerant leakage, effectively preventing system performance degradation and energy waste caused by leaks, and improving the operating efficiency and safety of the ground source heat pump system.

[0110] S40: Based on potential decay characteristics, calculate health indicators and determine the monitoring results for the current period based on a preset dynamic monitoring method.

[0111] In this application, health indicators are calculated based on potential decay characteristics, and monitoring results for the current time period are determined based on a preset dynamic monitoring method, including: S41: Based on principal component analysis, the potential attenuation features are reduced in dimensionality to obtain the reduced projection coordinates.

[0112] For high-dimensional latent decay features extracted from cross-time common feature encoders This contains rich information about leakage trends, but inevitably also includes noise. Compressing high-dimensional features into low-dimensional representations can reduce redundant information and highlight key information. Therefore, it can address potentially decaying features. The calculation formula is as follows: (Decentralized processing is performed.) , Formula 19; in, Potential attenuation characteristics Decentralized feature vector; N is the potential decay feature. The dimension; Potential attenuation characteristics The value corresponding to the k-th dimension.

[0113] Next, Principal Component Analysis (PCA) was used to linearly reduce and fuse the high-dimensional potential decay features after decentralization.

[0114] PCA first processes the decentralized feature vectors... Eigenvalue decomposition yields a set of eigenvalues ​​and corresponding eigenvectors, which are the principal components sorted by their variance contribution. Next, the top K principal components are selected based on their cumulative explained variance ratio (usually set to 95%), and the matrix formed by these principal components is the projection matrix. It is used to map high-dimensional data to a low-dimensional space. The decentralized feature vectors... Projecting onto this component space yields the dimension-reduced projected coordinates. The calculation formula is as follows: , formula 20; in, Let these be the coordinates of the current residual sequence in the K-dimensional principal component space; It is the projection matrix formed by the first K principal components; Potential attenuation characteristics The decentralized feature vector.

[0115] S42: Calculate health indicators by weighting and combining the dimensions of the projected coordinates.

[0116] Using the ratio of the explained variance of each principal component as the weight, the projected coordinates... The various dimensions are weighted and combined to calculate the Health Indicator (HI). Specifically, the weighted combination can be achieved through weighted summation, calculated using the following formula: , Formula 21; in, Health indicators for the current period; express The projection value onto the j-th principal component; This represents the proportion of the principal component eigenvalue to the sum of the eigenvalues ​​of the selected K principal components.

[0117] The current health indicators comprehensively reflect the operating status of the ground source heat pump system and the potential trend of refrigerant leakage. This facilitates the timely detection of potential refrigerant leaks and provides a scientific basis for system maintenance and management. Continuous monitoring of changes in health indicators can provide early warnings of potential faults, reducing system performance degradation and energy waste caused by refrigerant leaks, thereby ensuring the efficient and stable operation of the ground source heat pump system.

[0118] S43: Calculate the leakage statistics for the current time period based on health indicators and the preset dynamic monitoring method.

[0119] In this application, after extracting the historical potential decay features from each historical residual segment, the following is also included: Based on the historical potential decay characteristics, the historical health index of each historical residual segment is calculated to obtain the historical health index sequence. Based on the preset dynamic monitoring method, the historical health indicator sequence is dynamically monitored, and the historical leakage statistics are calculated. Based on historical leakage statistics, a leakage warning threshold and a maintenance warning threshold are determined, wherein the leakage warning threshold is less than the maintenance warning threshold.

[0120] In this application, the historical potential decay features extracted from the training set used to train the cross-time common feature encoder in steps S31 to S35 are also calculated using formulas 19 to 21 to obtain the historical health indicator sequence for each historical residual segment. During the operation period of a unit, the historical health indicator sequence shows a monotonically decreasing trend as the unit operates, which is consistent with the performance degradation process caused by the gradual leakage of refrigerant.

[0121] To quantify the refrigerant leakage degradation trend implied in health indicators, a pre-defined dynamic monitoring method is used to dynamically monitor historical health indicator sequences. The pre-defined dynamic monitoring method can be selected according to specific needs. Here, we illustrate this with two dynamic monitoring methods with different characteristics: the Cumulative Sum Control Chart (CUSUM) and the Exponential Weighted Moving Average (EWMA), which are used to statistically analyze historical health indicator sequences. Both methods can effectively amplify small and persistent process deviations, making them suitable for sensitive detection of early leaks, but their focus and calculation logic differ.

[0122] The process of dynamically monitoring historical health indicator sequences using the cumulative sum and control chart method is as follows: The cumulative sum control chart method can detect persistent, minute shifts in process mean by accumulating small deviations between historical samples and target values. This makes it highly suitable for capturing the slow decline trend of health indicators caused by the initial stage of refrigerant leakage. The cumulative sum control chart method calculates historical leakage statistics. The calculation formula is as follows: , Formula 22; in, This represents the historical leakage statistics for time period t. This represents the health indicator value for time period t; The baseline mean of the historical health indicator series under healthy conditions in the training set; It is a reference value used to suppress misjudgments caused by random fluctuations; This represents the historical leakage statistics for the (t-1)th time period; express It is initialized to 0.

[0123] The process of dynamically monitoring historical health indicator sequences using the exponentially weighted moving average control chart method is as follows: The exponentially weighted moving average control chart method, by assigning higher weights to recent data while accumulating historical information, can smooth out random fluctuations and sensitively detect persistent small shifts in the process mean, making it very suitable for monitoring early leakage trends.

[0124] First, calculate the historical health indicator series relative to its health baseline mean. positive deviation This positive bias only considers the decline of historical health indicators below the health baseline, and the calculation formula is as follows: , formula 23; in, The positive deviation at time t; The baseline mean of the historical health indicator series under healthy conditions in the training set; This represents the health indicator value for time period t.

[0125] Based on Formula 23 above, a positive deviation sequence is obtained, and historical leakage statistics are calculated based on the positive deviation sequence. The calculation formula is as follows: , Formula 24; in, This represents the historical leakage statistics for time period t. This is a smoothing coefficient, and its value can be 0.15, 0.2, or 0.25. This represents the historical leakage statistics for the (t-1)th time period; express It is initialized to 0.

[0126] After calculating the historical leakage statistics, leakage warning thresholds and maintenance warning thresholds are determined based on these historical leakage statistics.

[0127] The early warning threshold for the cumulative control chart method can be set using the empirical quantile method. Specifically, based on training set data representing the health status of the units, historical leakage statistics for all units in the training set during their operating periods are calculated. Historical leakage statistics The average of the first preset quantiles is set as the leakage warning threshold, and historical leakage statistics are used as the basis. The average of the second preset quantile is set as the maintenance warning threshold, where the first preset quantile must be less than the second preset quantile; the specific values ​​of both are not limited. For example, historical leakage statistics can be used. The average of the 84th percentile was set as the leak warning threshold, and historical leak statistics were used as the basis. The average of the 98th percentile is set as the maintenance warning threshold.

[0128] The warning threshold for the exponentially weighted moving average control chart method can be set using the Sigma control limit method based on statistical process control. Specifically, historical leakage statistics are calculated based on training set data representing the unit's health status. The asymptotic standard deviation of the training set is used to determine the leakage warning threshold. Then, the asymptotic standard deviation of the training set mean minus a first preset multiple (i.e., the training set mean) under healthy conditions is set as the leakage warning threshold. The asymptotic standard deviation of the training set mean minus a second preset multiple is set as the maintenance warning threshold. The first preset multiple must be less than the second preset multiple, and their specific values ​​are not limited. For example, the asymptotic standard deviation of the training set mean minus 1.8 times can be set as the leakage warning threshold, and the asymptotic standard deviation of the training set mean minus 7 times can be set as the maintenance warning threshold.

[0129] By setting warning thresholds using the two optional methods described above, tiered warning systems for refrigerant leaks in ground source heat pump systems can be implemented. This facilitates the timely detection of early signs of refrigerant leaks, preventing further deterioration and effectively reducing maintenance costs and system downtime. Furthermore, the tiered warning mechanism provides maintenance personnel with clearer and more specific handling guidelines, enabling them to quickly take appropriate measures based on different warning levels, ensuring the safe and stable operation of the ground source heat pump system.

[0130] For the health indicators for the current time period calculated in step S42, the leakage statistics for the current time period are calculated using formula 22 above. Alternatively, use formulas 23-24 to calculate the leakage statistics for the current period. The current leakage statistics are compared with the leakage warning threshold and the maintenance warning threshold to determine the monitoring results.

[0131] S44: In response to the leakage statistics value of the current period being greater than the leakage warning threshold, the monitoring result is a level 1 leakage, triggering a leakage warning.

[0132] If the current leakage statistics value is greater than the leakage warning threshold, it means that the health indicator has shown a statistically significant downward drift. At this time, the monitoring result is a level 1 leakage with relatively mild severity, which can trigger a refrigerant leakage warning.

[0133] S45: In response to the leakage statistics value of the current period being greater than the maintenance warning threshold, the monitoring result is a level 2 leakage, triggering a maintenance warning.

[0134] If the leakage statistics for the current period exceed the maintenance warning threshold, the monitoring result is a level 2 leakage with a high degree of severity. It is determined that the refrigerant leakage of the unit is serious and a unit maintenance warning needs to be triggered.

[0135] Unlike existing technologies that only issue warnings after a significant performance degradation, making early fault detection difficult, the refrigerant leakage fault monitoring method for ground source heat pump systems provided in this application, through physical knowledge-guided feature extraction and highly sensitive statistical process control methods, can capture the weak attenuation signal at the initial stage of refrigerant leakage. This significantly advances the warning time, gaining valuable time for implementing predictive maintenance and avoiding energy waste and major losses.

[0136] Upon triggering a leak or maintenance alert, relevant data can be automatically recorded and an alert report generated. The alert report will detail the time the alert was triggered, the current health indicators and leak statistics for that period, and a comparison with the alert threshold, providing strong data support for subsequent troubleshooting and unit maintenance. Simultaneously, the alert report can be sent to the mobile devices or email addresses of relevant personnel, ensuring they can promptly receive the alert information and take appropriate action.

[0137] The aforementioned early warning mechanism enables real-time and accurate monitoring of refrigerant leakage in ground source heat pump systems. A Level 1 leak triggers a timely leak warning, alerting relevant personnel to monitor the system status and conduct preliminary investigations to prevent further deterioration. A Level 2 leak triggers a maintenance warning, indicating a more severe refrigerant leak necessitates immediate professional inspection and maintenance of the unit to prevent system malfunction due to excessive refrigerant leakage, which could impact the overall performance and lifespan of the ground source heat pump system. Furthermore, the automatic data recording and early warning report generation function not only provides detailed data for troubleshooting but also allows for analysis of historical warning data to summarize system operating patterns, identify potential safety hazards in advance, and implement preventative maintenance of the ground source heat pump system, thereby improving its reliability and stability.

[0138] In one specific embodiment, taking the ground source heat pump monitoring system of the chiller room in a commercial complex in a southern city as an example, the chiller room in this commercial complex is equipped with one ground source heat pump unit (rated cooling power 171kW, rated heating power 212kW, rated cooling capacity 1000kW, rated heating capacity 1109kW, rated energy efficiency ratio 5.84), two centrifugal chiller units (rated cooling power 511kW, rated cooling capacity 2813kW, rated energy efficiency ratio 5.5), four load-side circulation pumps (three circulation pumps with rated power of 75kW, one circulation pump with rated power of 30kW) and four ground source-side circulation pumps (three circulation pumps with rated power of 75kW, one circulation pump with rated power of 30kW), and is equipped with a temperature monitoring system (load-side return water temperature T). lr Geological source side outlet water temperature T go Intake temperature T s Evaporation temperature T e Condensation temperature T c , liquid outlet temperature T l (etc.), pressure (compressor suction pressure P) e and compressor discharge pressure P c It uses multiple sensors such as flow rate (instantaneous flow rate on the load side, etc.) and power (compressor power, kW, etc.), with a data collection interval of 1 minute.

[0139] According to historical maintenance records, this ground source heat pump unit underwent two comprehensive maintenance procedures on June 7, 2023, and August 5, 2025. Based on this, the historical operating data was divided into three operating periods, each corresponding to a complete degradation process following a full maintenance, showing the unit gradually evolving from an initial healthy state to one with increased refrigerant leakage. Operating period 1 is from 00:00 on April 1, 2023 to 00:00 on June 6, 2023; operating period 2 is from 00:00 on April 1, 2024 to 00:00 on October 31, 2024; and operating period 3 is from 00:00 on August 6, 2025 to 00:00 on October 31, 2025. Data from operating periods 1 and 2 were used as the training set, and data from operating period 3 was used as the test set.

[0140] A cross-time period common feature encoder was trained using training set data from unit operation periods 1 and 2, and potential decay features were extracted to calculate a health indicator sequence. Leakage statistics were calculated from the health indicator sequence of the training set. and Record and draw. and The trends of both were analyzed. Results showed that, over approximately 2040 hours of operation, CUSUM reached the preset leak warning threshold at around 1700 hours and the preset maintenance warning threshold at around 2000 hours; EWMA reached the preset leak warning threshold at around 1700 hours and the preset maintenance warning threshold at around 1900 hours. Timely maintenance of the ground source heat pump unit upon triggering a maintenance warning can effectively prevent greater energy losses and failure risks that may be caused by refrigerant leaks.

[0141] Unlike existing technologies that rely on threshold-based early warning methods based on a single performance measurement variable (such as power or temperature), which suffer from lag and inability to provide timely warnings in the early stages of unit performance degradation, leading to increased energy consumption and maintenance costs, this application provides a ground-source heat pump system refrigerant leakage fault monitoring method. By constructing and training a cross-time period common feature encoder to extract potential degradation features, it can more comprehensively and accurately capture early signs of refrigerant leakage. Simultaneously, by calculating a health indicator sequence and analyzing the changing trends of leakage statistics, it achieves dynamic monitoring and early warning of refrigerant leakage, thereby significantly improving the accuracy and timeliness of early warnings.

[0142] In one embodiment, to verify the effectiveness of the refrigerant leakage fault monitoring method for ground source heat pump systems provided in this application, the following comparative experiment was conducted: The ground source heat pump system refrigerant leakage fault monitoring method described in steps S10 to S40 above (which uses two statistical decision strategies, CUSUM and EWMA, respectively) is compared with the fault monitoring method based on the PCA model and the fault monitoring method based on the linear regression model.

[0143] To ensure that the PCA and linear regression models can make predictions based on unit health status data under multiple operating conditions, this study selects the top 40% of data points from each of the unit operation periods 1 and 2 to form a training set, and uses the data from the unit operation period 3 as the corresponding test set.

[0144] For the method of this application, the training and testing process is described in detail in steps S10 to S40 above, and will not be repeated here. Finally, the two fault monitoring results based on CUSUM and EWMA on the test set are obtained.

[0145] See Figure 4 , Figure 4 This is a schematic diagram of the CUSUM-based fault monitoring results of an embodiment of the refrigerant leakage fault monitoring method for ground source heat pump systems provided in this application. The figure shows the results of monitoring the test set data using the Cumulative Sum (CUSUM) control chart method. The horizontal axis represents operating time (hours), and the vertical axis represents the calculated CUSUM statistical value. The light-colored dashed line represents the preset leakage warning threshold (corresponding to Level 1 leakage); the dark solid line represents the preset maintenance warning threshold (corresponding to Level 2 leakage). As can be seen from the figure, at approximately 1700 hours, the curve (CUSUM statistical value) breaks through the light-colored dashed line (leakage warning threshold), indicating that the system has shown early signs of refrigerant leakage at this time, triggering a Level 1 leakage warning. At this point, the fault signal is still weak, and maintenance is not required temporarily. Subsequently, at approximately 2000 hours, the curve continues to rise and breaks through the red solid line (maintenance warning threshold), indicating that the leakage has worsened, triggering a Level 2 maintenance warning.

[0146] See Figure 5 , Figure 5This is a schematic diagram of the fault monitoring results based on EWMA, an embodiment of the ground source heat pump system refrigerant leakage fault monitoring method provided in this application. The figure shows the results of monitoring the test set data using the Exponentially Weighted Moving Average (EWMA) control chart method. The horizontal axis represents the operating time (hours), and the vertical axis represents the calculated EWMA statistical value. To define the "normal" and "abnormal" states of refrigerant charge, the EWMA control chart adopts a two-sided control limit design: the light-colored dashed lines represent the upper limit (UCL) and lower limit (LCL) of the leakage warning threshold, typically determined based on the mean of healthy data plus or minus k times the standard deviation. The dark-colored solid lines represent the upper limit (UCL) and lower limit (LCL) of the maintenance warning threshold, representing the system's allowable extreme fluctuation range. As long as the EWMA statistical value fluctuates within the range defined by the above upper and lower limits, it is considered normal background noise. As can be seen from the graph, at approximately hour 1700, the EWMA statistic broke below the light-colored dashed line (the lower limit of the leak warning), indicating a significant shift in the data. At this point, the system showed early signs of refrigerant leakage, triggering a Level 1 warning.

[0147] For the fault monitoring method based on the PCA model, the same variable data (host load L) as the method in this application will be used. g , Load side return water temperature T lr Geological source side outlet water temperature T go Instantaneous flow rate F on the load side l Compressor suction pressure P e Compressor discharge pressure P c Inhalation superheat T suc Supercooling T sub The compressor power (kW) is input into the PCA model after conventional data preprocessing. The cumulative variance contribution rate threshold of the PCA model is set to 95%, and the squared prediction error at time i is calculated. .

[0148] For fault monitoring methods based on linear regression models, the variable data that are the same as those in this application are first subjected to conventional data preprocessing, and divided into independent variables (host load L). g , Load side return water temperature T lr Geological source side outlet water temperature T go ) and dependent variable (compressor suction pressure P) e Compressor discharge pressure P c Inhalation superheat T suc Supercooling T sub After inputting the compressor power (kW) into the model, multiple independent linear regression equations are trained. The output variable is predicted through the multiple linear regression equations, and the residuals of the output variable are calculated. .

[0149] In the two comparison methods, respectively, As a monitoring indicator, when A value exceeding the leakage warning threshold is considered a fault, thus yielding fault monitoring results for both the PCA-based and linear regression-based methods. Both comparison methods use... The 95th percentile is used as the corresponding maintenance warning threshold.

[0150] The 72 operating hours before the end of each unit's operating period are designated as the failure period. Fault monitoring results from various methods are evaluated using multi-dimensional evaluation metrics to determine the corresponding fault diagnosis performance. Evaluation metrics include: fault identification accuracy (the ratio of correctly predicted samples during the failure period to the total number of samples during the failure period), overall accuracy (the ratio of correctly identified samples to the total number of samples), false alarm rate (the ratio of falsely reported fault samples during the normal period to the total number of samples during the normal period), overall index (a weighted combination of fault identification accuracy, overall accuracy, and false alarm rate), and fault warning lead time (to reduce noise interference, a fault warning is defined as the first warning when five consecutive samples detect a fault; the fault warning lead time refers to the time interval between the first fault warning and the actual equipment maintenance time).

[0151] The comparison results of various fault monitoring methods are shown in Table 3 below. It is evident that the method proposed in this application exhibits significant advantages across all evaluation indicators. The fault identification accuracy of this method far surpasses that of the fault monitoring method based on the PCA model (37.06%) and the fault monitoring method based on the linear regression model (44.41%), with the total indicator exceeding the other two methods by at least 0.2419. The fault monitoring methods based on the PCA model and the fault monitoring method based on the linear regression model only provide a refrigerant leakage fault lead time of 169 hours, indicating their insensitivity to minor faults; while the method proposed in this application can provide lead times of 325 hours and 299 hours, respectively, making it more valuable for engineering applications.

[0152]

[0153] Table 3 Comparison of various fault monitoring methods This application provides a refrigerant leakage fault monitoring device for a ground source heat pump system, see reference. Figure 6 , Figure 6 This is a schematic diagram of an embodiment of the refrigerant leakage fault monitoring device for a ground source heat pump system provided in this application. The ground source heat pump system refrigerant leakage fault monitoring device includes: Data processing module 10 is used to acquire and preprocess sensor data for the current time period to obtain operating condition variable data and performance variable data; The residual generation module 20 is used to obtain the residual sequence of the predicted performance variable data and the performance variable data based on the operating condition variable data and the residual generator. Feature extraction module 30 is used to input the residual sequence into the cross-time common feature encoder to extract potential decay features from the residual sequence; The indicator calculation module 40 is used to calculate health indicators based on potential decay characteristics and obtain the monitoring results for the current period based on a preset dynamic monitoring method.

[0154] The data processing module 10, residual generation module 20, feature extraction module 30 and index calculation module 40 interact to realize the process of refrigerant leakage fault monitoring of ground source heat pump system. You can refer to the specific description of steps S10 to S40 above. Where there are repetitions, they will not be repeated here.

[0155] See Figure 7 , Figure 7 This is a schematic diagram of an embodiment of the storage medium provided in this application.

[0156] The storage medium 700 stores program data 710, which, when executed by the processor, implements, as follows: Figure 1 The steps of the described method for monitoring refrigerant leakage faults in a ground source heat pump system.

[0157] 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.

[0158] 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.

[0159] See Figure 8 , Figure 8 This is a schematic diagram of the structure of an embodiment of the computer device provided in this application.

[0160] 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 above-described ground source heat pump system refrigerant leakage fault monitoring method.

[0161] Unlike existing technologies, this application discloses a method, apparatus, and equipment for monitoring refrigerant leakage faults in ground source heat pump systems. This method divides sensor data into operating condition variable data and performance variable data. Using a residual generator, it constructs a residual sequence between the predicted performance variable data obtained from the operating condition variable data and the actual performance variable data. This effectively eliminates interference from complex operating condition fluctuations, improves the accuracy of subsequent analysis, and fundamentally reduces the false alarm and false negative rates of leakage. By constructing a cross-time period common feature encoder, it extracts physically meaningful and cross-time period universal potential attenuation features from the residual sequence, enabling sensitive capture of weak attenuation signals in the early stages of refrigerant leakage, improving the timeliness of leakage fault early warning, and achieving early warning of leakage faults.

[0162] 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.

[0163] 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.

[0164] 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.

[0165] 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.

[0166] 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.

[0167] 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 ground source heat pump system refrigerant leakage fault monitoring method characterized by, include: Acquire and preprocess sensor data for the current time period to obtain operating condition variable data and performance variable data; Based on the operating condition variable data and the residual generator, a residual sequence between the predicted performance variable data and the performance variable data is obtained; The residual sequence is input into a cross-time period common feature encoder to extract potential decay features from the residual sequence; Based on the potential decay characteristics, health indicators are calculated, and the monitoring results for the current period are determined based on a preset dynamic monitoring method.

2. The ground source heat pump system refrigerant leakage fault monitoring method according to claim 1, characterized by, The process of acquiring and preprocessing sensor data for the current time period to obtain operating condition variable data and performance variable data includes: Acquire sensor data for the current time period and calculate derived variable data to obtain multiple sets of variable data; Based on the sliding window method, the multiple sets of variable data are divided into multiple sequence segments; Based on the preset stationarity determination mechanism, the corresponding statistical characteristic values ​​are calculated for each key variable that is sensitive to the operating conditions in each sequence segment. In response to any of the key variables having a statistical feature value greater than a preset statistical threshold, the corresponding sequence segment is filtered out. The filtered sequence segments are integrated and divided to obtain the operating condition variable data and the performance variable data.

3. The ground source heat pump system refrigerant leakage fault monitoring method according to claim 1, characterized by, The residual generator comprises at least one convolutional layer, one fully connected layer, and multiple residual shrink blocks; The step of obtaining the residual sequence between the predicted performance variable data and the performance variable data based on the operating condition variable data and the residual generator includes: The operating condition variable data is input into the residual generator, and the input feature map is obtained by high-dimensional mapping through the convolutional layer. The input feature map is input into the residual shrinkage block, and the output feature map is obtained based on the adaptive threshold and soft threshold functions. The output feature map is input into the fully connected layer to obtain the predicted performance variable data, and the residual sequence between the predicted performance variable data and the performance variable data is calculated.

4. The ground source heat pump system refrigerant leakage fault monitoring method according to claim 1, characterized by, The training process of the cross-time common feature encoder is as follows: The historical residual sequence obtained from sensor data based on historical time periods is segmented to obtain multiple historical residual segments; Based on the empirical trend function of leakage rate, a target weak label is set for each of the historical residual segments. Each of the historical residual segments is input into the cross-time period common feature encoder to extract historical potential decay features from each of the historical residual segments; The historical potential decay features are mapped to the corresponding single-time decoder to generate predicted weak labels for each residual segment. The loss value is calculated based on the target weak label and the predicted weak label. The loss value is backpropagated to optimize the cross-time common feature encoder and each of the single-time decoders to obtain the trained cross-time common feature encoder.

5. The ground source heat pump system refrigerant leakage fault monitoring method according to claim 4, characterized by, The empirical trend function for the leakage rate is: , Equation 1 ; in, The target weak label of the i-th historical residual segment; This represents the initial refrigerant balance of the unit. This refers to the refrigerant leakage rate; This is the runtime of the unit from its initial state to the i-th historical residual segment.

6. The ground source heat pump system refrigerant leakage fault monitoring method according to claim 4, characterized by, The loss value includes at least one or more of the following: prediction loss, physical consistency loss, and monotonicity regularization term; The function for calculating the prediction loss is: , Equation 2; wherein, is the predicted loss; is the target weak label for the i-th historical residual segment; is the predicted weak label for the i-th historical residual segment; The calculation function for the physical consistency loss is: , Equation 3; wherein, is the physical consistency loss; is the historical latent decay feature of the i-th historical residual segment; is the residual feature value based on the historical latent decay feature; N is the corresponding number of historical residual segments of the historical latent decay feature. The function for calculating the monotonicity regularization term is: , Equation 4; in, This refers to the monotonicity regularization term; To correct the linear unit function; This is a hyperparameter.

7. The ground source heat pump system refrigerant leakage fault monitoring method according to claim 4, characterized by, After extracting the historical potential decay features from each of the historical residual segments, the method further includes: Based on the historical potential decay characteristics, the historical health index of each historical residual segment is calculated to obtain the historical health index sequence. The historical health indicator sequence is dynamically monitored based on a preset dynamic monitoring method, and the historical leakage statistics are calculated. Based on the historical leakage statistics, a leakage warning threshold and a maintenance warning threshold are determined, wherein the leakage warning threshold is less than the maintenance warning threshold.

8. The ground source heat pump system refrigerant leakage fault monitoring method according to claim 7, characterized by, Based on the potential attenuation characteristics Calculate health indicators and determine the monitoring results for the current period based on a preset dynamic monitoring method, including: Based on principal component analysis, the potential attenuation features are dimensionality reduced to obtain the dimensionality-reduced projected coordinates. The health index is calculated by weighting and combining the dimensions of the projected coordinates. Based on the health indicators and the preset dynamic monitoring method, calculate the leakage statistics for the current time period; If the leakage statistics value for the current time period is greater than the leakage warning threshold, the monitoring result is a Level 1 leakage, triggering a leakage warning. If the leakage statistics value for the current period exceeds the maintenance warning threshold, the monitoring result is a level 2 leakage, triggering a maintenance warning.

9. A ground source heat pump system refrigerant leakage fault monitoring apparatus characterized by, include: The data processing module is used to acquire and preprocess the sensor data for the current time period to obtain operating condition variable data and performance variable data; The residual generation module is used to obtain the residual sequence between the predicted performance variable data and the performance variable data based on the operating condition variable data and the residual generator. The feature extraction module is used to input the residual sequence into the cross-time common feature encoder and extract potential decay features from the residual sequence; The indicator calculation module is used to calculate health indicators based on the potential decay characteristics and determine the monitoring results for the current period based on a preset dynamic monitoring method.

10. A computer device comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program comprises instructions that, when executed by the processor, cause the processor to perform the method of any one of claims 1-9. The processor executes the computer program to implement the steps of the ground source heat pump system refrigerant leakage fault monitoring method according to any one of claims 1-8.

Images