Refrigerator fault diagnosis method and system based on multi-model integration and adaptive learning

Abstract

The application provides a refrigerating machine fault diagnosis method and system based on multi-model integration and adaptive learning, comprising: obtaining historical refrigerating machine operation parameters as reference samples, and obtaining refrigerating machine operation parameters to be judged as judgment samples; constructing a working condition box based on the current reference samples, judging whether the current judgment samples are within the current working condition box range, if not, deleting the current judgment samples; otherwise, retaining; constructing a refrigerating machine fault identifier, using the constructed refrigerating machine fault identifier to perform fault identification based on the judgment samples; constructing a refrigerating machine power physical gray box model, and using the constructed refrigerating machine power physical gray box model to predict the refrigerating machine power; constructing a power prediction range based on the current predicted refrigerating machine power; comparing the refrigerating machine power in the judgment samples with the power prediction range; and judging whether the current refrigerating machine has an abnormality in combination with the identification result of the refrigerating machine fault identifier.

Description

Technical Field

[0001] This invention relates to the field of refrigeration machine fault diagnosis technology, and more specifically, to a refrigeration machine fault diagnosis method and system based on multi-model integration and adaptive learning. Background Technology

[0002] As critical energy equipment in industrial and commercial buildings, the operational stability and reliability of refrigeration units directly impact energy consumption and equipment lifespan. However, due to the complex and variable operating conditions and diverse potential fault types of refrigeration units, traditional fault diagnosis methods face numerous challenges in terms of real-time performance and accuracy. Currently, refrigeration unit fault diagnosis technologies mainly include rule-based expert systems, physical model-based analysis methods, and data-driven machine learning methods.

[0003] Rule-based expert systems rely on predefined fault rules and empirical knowledge. While simple to implement, they often lack adaptability when facing unknown faults or changing operating conditions. Physical model-based diagnostic methods predict operating states by establishing mathematical models (such as thermodynamic models) of the refrigeration system. However, these methods typically require precise equipment parameters and operating conditions, involve complex modeling processes, and are sensitive to parameter changes, making them difficult to apply to dynamic environments in actual operation. In recent years, with the development of data acquisition technologies, data-driven methods based on machine learning have gradually gained attention. These methods train classification models using historical data, enabling the identification of complex fault modes. However, single machine learning methods are prone to overfitting or misjudgment when samples are imbalanced or operating conditions vary significantly, and they lack in-depth utilization of the system's physical characteristics.

[0004] In existing technologies, some studies attempt to combine physical models and data-driven methods, such as modeling equipment behavior using gray-box models and then supplementing it with statistical methods for anomaly detection. However, these methods still have shortcomings in multi-level fault classification, dynamic sample updates, and screening for complex operating conditions. For example, existing gray-box modeling methods are mostly based on static parameter fitting, making it difficult to adapt to changes in operating parameters in real time; at the same time, the labeling of fault samples is usually relatively simple, failing to fully utilize multi-level labeling strategies to improve classification accuracy. In addition, when encountering new fault types, existing methods lack effective sample expansion and model adaptation mechanisms, resulting in limited diagnostic capabilities.

[0005] Patent document CN112766047A (application number: 202011588458.5) discloses a method and device for refrigeration system fault diagnosis, belonging to the field of fault diagnosis and artificial intelligence, including (1) constructing a BN network model; (2) obtaining the prior probability values ​​of the target signal node and the fault signal node; (3) collecting BN network information; (4) performing relaxation operations on the data of the conditional mutual information matrix to construct a TAN classifier model that matches the fault features; (5) recalculating the conditional probability matrix; (6) calculating the posterior probability between the fault signal node and the feature signal node; (7) verifying the posterior probability values; (8) sorting the posterior probability values ​​of each state in descending order, and taking the state corresponding to the largest posterior probability value as the priority diagnosis / prediction classification result of the target signal node, etc.

[0006] Therefore, there is an urgent need for an online fault diagnosis method that can integrate the physical constraints of gray box modeling with the pattern recognition advantages of machine learning, and also has multi-level fault classification capabilities, operating condition adaptability, and model adaptive update functions, in order to improve the real-time performance, accuracy, and robustness of refrigeration fault diagnosis. Summary of the Invention

[0007] In view of the shortcomings of the prior art, the purpose of this invention is to provide a method and system for refrigeration machine fault diagnosis based on multi-model integration and adaptive learning.

[0008] A refrigerator fault diagnosis method based on multi-model ensemble and adaptive learning, provided by the present invention, includes:

[0009] Step S1: Obtain historical chiller operating parameters as a baseline sample, obtain the chiller operating parameters to be judged as a judgment sample; construct a test case based on the current baseline sample, determine whether the current judgment sample is within the range of the current test case, if it is not within the range of the current test case, delete the current judgment sample; otherwise, retain it and trigger the execution of step S2.

[0010] Step S2: Construct a refrigeration machine fault identifier and use the constructed refrigeration machine fault identifier to identify faults based on judgment samples;

[0011] Step S3: Construct a physical gray box model of chiller power, and use the constructed physical gray box model of chiller power to predict the chiller power of the judgment sample; construct a power prediction range based on the currently predicted chiller power; compare the chiller power in the judgment sample with the power prediction range; and at the same time, combine the identification results of the chiller fault identifier to determine whether there is an abnormality in the current chiller.

[0012] The refrigeration unit fault identifier is based on a multi-layer fault sample set and utilizes multiple fault identification models to identify faults.

[0013] Preferably, the operating parameters of the refrigeration unit include: evaporator outlet water temperature, evaporator inlet water temperature, evaporator saturation temperature, evaporator saturation pressure, condenser outlet water temperature, condenser inlet water temperature, condenser saturation temperature, condenser saturation pressure, evaporator chilled water flow rate, current percentage, and power.

[0014] Preferably, step S1 includes:

[0015] Historical operating parameters of the refrigeration unit are obtained and used as a baseline sample; operating parameters of the refrigeration unit to be judged are obtained and used as a judgment sample.

[0016] The baseline sample is binned, with the standard deviation of each operating parameter as the bin size, forming several operating condition bins; a preset number of valid samples are retained in each operating condition bin; it is determined whether each parameter in the current judgment sample is within the corresponding parameter range of the operating condition bin, and if so, it is retained, otherwise it is discarded.

[0017] Preferably, step S2 includes:

[0018] Step S2.1: Obtain various types of refrigeration machine fault mode data, and divide them into hierarchical levels according to the dimensions of refrigeration machine fault status, fault location and fault mode to form a multi-level fault sample set;

[0019] Step S2.2: Preprocess the fault sample data of each level in the multi-level fault sample set, and label the preprocessed fault sample data to form a fault sample set for each level.

[0020] Step S2.3: Construct corresponding fault identification models based on multi-layer fault sample sets, and train the corresponding fault identification models using the fault sample dataset of each layer to obtain the trained fault identification models;

[0021] Step S2.4: Obtain a fault identifyer by integrating the fault identification models of each layer after training.

[0022] Preferably, step S2.3 includes:

[0023] Step S2.3.1: Construct fault identification models based on ensemble learning models, deep neural networks, logistic regression, support vector machines, and K-nearest neighbors, respectively;

[0024] Step S2.3.2: For any fault identification model, a greedy algorithm is adopted, with the multi-objective optimization function Score with category number penalty as the objective. The worst classification category is merged into other categories through iteration, the Score is calculated, and the merging method with the highest current Score is retained. This process is repeated until the Score no longer improves or the number of categories is less than or equal to the preset value, and the current trained fault identification model is obtained.

[0025] Step S2.3.3: For any trained fault identification model, select the trained fault identification model with the highest Score as the final trained fault identification model.

[0026] Preferably, the multi-objective optimization function Score includes:

[0027]

[0028] Where, N classes N represents the number of categories after merging. original α represents the original number of categories, and α is the category number penalty coefficient. Macro-F1 is a classification model evaluation metric suitable for imbalanced data with multiple classes.

[0029] Preferably, the physical gray box model of the chiller power includes:

[0030]

[0031] Among them, T c T represents the condenser outlet water temperature. e The outlet water temperature of the evaporator is c1~c 10 For model parameters, COP ch Q is the energy efficiency ratio of the refrigeration unit. ch For the cooling capacity of the refrigeration system; P ch This refers to the power of the refrigeration unit.

[0032] Preferably, step S3 includes:

[0033] Step S3.1: Use the benchmark samples to perform linear regression parameter fitting on the constructed chiller power physical gray box model to obtain the fitted chiller power physical gray box model;

[0034] Step S3.2: Use the fitted physical gray box model of chiller power to predict the chiller power of the sample; based on the predicted chiller power, obtain the corresponding power prediction range according to the preset requirements;

[0035] Step S3.3: Obtain the percentage of refrigeration units in the sample whose power exceeds the corresponding power prediction range within a preset time period. If the percentage exceeds the preset value and is identified as a fault by the fault identifier, then the current refrigeration unit is judged to be abnormal and the fault type is marked; otherwise, it is judged to be normal.

[0036] Preferably, step S3.1 includes: using benchmark samples to fit the constructed physical gray box model of chiller power to linear regression parameters using the median regression method, thereby obtaining the fitted physical gray box model of chiller power.

[0037] A refrigerator fault diagnosis system based on multi-model ensemble and adaptive learning, provided by the present invention, includes:

[0038] Module M1: Obtain historical chiller operating parameters as a baseline sample, obtain chiller operating parameters to be judged as a judgment sample; construct a test case based on the current baseline sample, determine whether the current judgment sample is within the range of the current test case, if not within the range of the current test case, delete the current judgment sample; otherwise, retain it and trigger the execution of module M2.

[0039] Module M2: Construct a refrigeration machine fault identifier, and use the constructed refrigeration machine fault identifier to identify faults based on judgment samples;

[0040] Module M3: Constructs a physical gray box model of chiller power, uses the constructed physical gray box model to predict chiller power in the judgment sample, constructs a power prediction range based on the currently predicted chiller power, compares the chiller power in the judgment sample with the power prediction range, and combines the identification results of the chiller fault identifier to determine whether there is an abnormality in the current chiller.

[0041] The refrigeration unit fault identifier is based on a multi-layer fault sample set and utilizes multiple fault identification models to identify faults.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. This invention constructs a multi-layer fault sample set through a hierarchical labeling strategy and trains a multi-model integrated fault identifier, thereby achieving high-precision identification of complex fault modes of refrigeration machines.

[0044] 2. This invention addresses the problem of decreased generalization of diagnostic models caused by fluctuations in operating conditions. By filtering and removing data outside the operating conditions from the judgment samples, it reduces the impact of non-operating condition-related interference on fault diagnosis and improves the robustness of the model in real operating environments.

[0045] 3. This invention addresses the false triggering problem caused by noise or operating condition fluctuations in traditional threshold methods or pure data-driven models. It significantly reduces the false alarm rate by using a physical gray box model prediction interval and a multi-model integrated recognizer as the joint criterion.

[0046] 4. This invention addresses the problems of low fault type recognition rate and lagging model iteration in small sample sizes by dynamically expanding the fault sample dataset and retraining the model through SMOTE technology, combined with a closed-loop mechanism of on-site personnel feedback, to achieve continuous optimization of the fault diagnosis system; and overcomes the limitations of traditional methods that rely on fixed historical data by using an adaptive learning framework to quickly adapt to new fault types. Attached Figure Description

[0047] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0048] Figure 1 This is a flowchart of a refrigeration machine fault diagnosis method based on multi-model ensemble and adaptive learning.

[0049] Figure 2 This is a flowchart illustrating an embodiment of a refrigeration machine fault diagnosis method based on multi-model ensemble and adaptive learning. Detailed Implementation

[0050] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0051] Example 1

[0052] According to the present invention, a refrigeration machine fault diagnosis method based on multi-model ensemble and adaptive learning is provided, such as... Figures 1 to 2 As shown, it includes the following steps:

[0053] Step 1: Collect data on typical failure modes of various refrigeration units, and hierarchically divide the data according to three dimensions: failure state, failure location, and failure mode, forming a multi-level failure sample set. For each level of the failure type in this sample set, perform data standardization and label encoding, optimal classification strategy search, model selection, and hyperparameter optimization training steps in sequence, ultimately constructing a multi-model ensemble fault identifyer. This embodiment constructs a multi-level failure sample set and trains a multi-model ensemble fault identifyer through a hierarchical labeling strategy, achieving high-precision identification of complex failure modes of refrigeration units and solving the problem of insufficient classification capability of traditional single models for multiple types of faults.

[0054] Step 2: Real-time acquisition of core operating parameters of the chiller as judgment samples, and historical data as benchmark samples; constructing operating condition bins with the standard deviation of each parameter in the benchmark samples as the bin size; judgment samples must meet the requirement that all operating parameters are within the corresponding operating condition bin range, otherwise they are eliminated; this embodiment reduces the impact of non-operating condition related interference on fault diagnosis by filtering and eliminating data outside the operating conditions in the judgment samples, and improves the robustness of the model in real operating environments.

[0055] Step 3: Based on the benchmark samples, perform linear regression parameter fitting on the physical gray box model of the chiller power, and calculate the 99.5% prediction interval for the judgment samples. Simultaneously, input the judgment samples into the integrated fault identifier for fault identification. If the number of samples exceeding the prediction interval exceeds 10% of the total number of judgment samples and is identified as a fault, it is judged as abnormal and the fault type is marked; otherwise, it is judged as a normal sample. This embodiment significantly reduces the false alarm rate by using the joint criterion of the gray box model prediction interval and the multi-model integrated identifier, avoiding false triggering caused by noise or operating condition fluctuations in traditional threshold methods or pure data-driven models.

[0056] Step 4: Based on the on-site investigation results, label the actual causes of the faults and add them to the fault sample set. If identification fails, use SMOTE technology to expand the fault sample set and retrain the fault identifyr. This embodiment solves the problem of low identification rate for small-sample fault types by dynamically expanding the fault sample dataset and retraining the model using SMOTE technology. This enables the model to have continuous adaptive optimization capabilities, which is superior to existing methods using static training. Simultaneously, through the closed-loop feedback mechanism from on-site personnel and the iterative update mechanism of the model, the fault diagnosis system achieves self-evolution, overcoming the shortcomings of traditional methods that rely on fixed historical data and are difficult to adapt to new fault types.

[0057] In this embodiment, the specific hierarchical labeling strategy is shown in Table 1 below. It should be noted that when performing multi-level classification of fault data, the fault types in each level should be enriched as much as possible, which will help in the search for the optimal classification strategy in subsequent steps.

[0058] Table 1

[0059]

[0060]

[0061] Specifically, the optimal classification strategy search in step 1 is based on a random forest model and employs a greedy algorithm. It uses a multi-objective optimization function, Score, with a penalty for the number of categories as the objective. Through iteration, the worst-performing classification category is merged into other categories. The model is trained to calculate the Score, and the merging method with the highest Score is retained. This process is repeated until the Score no longer improves or the number of categories is less than or equal to 2, thus yielding the optimal classification strategy. The expression for the multi-objective optimization function Score is:

[0062]

[0063] Where, N classes N represents the number of categories after merging. originalα represents the original number of categories, and α is the category number penalty coefficient with a value of 0.7. Macro-F1 is a classification model evaluation metric suitable for imbalanced multi-class data. Here, Macro-F1 is an evaluation metric used to reflect the predictive performance of the model. It is suitable for imbalanced datasets. The larger the Macro-F1, the better the model performance, and vice versa.

[0064]

[0065] Among them, Precision i Recall is the precision for category i. i The recall rate is for category i.

[0066] Specifically, the model selection in step 1 includes at least one of the following models: ensemble learning model (Bagging / Boosting / Stacking), deep neural network, logistic regression, support vector machine, and K-nearest neighbors.

[0067] In this embodiment, the model composition in the multi-model integrated fault identifier is shown in Table 2 below.

[0068] Table 2

[0069]

[0070] Specifically, the core operating parameters of the refrigeration unit mentioned in step 2 include multiple or all of the following parameters: evaporator outlet water temperature, evaporator inlet water temperature, evaporator saturation temperature, evaporator saturation pressure, condenser outlet water temperature, condenser inlet water temperature, condenser saturation temperature, condenser saturation pressure, evaporator chilled water flow rate, current percentage, and power.

[0071] Specifically, the method for screening operating conditions in step 2 is as follows: the operating parameter data of the refrigeration unit is divided into bins, and the standard deviation of each operating parameter is used as the bin size to form several operating condition bins. 30 valid samples are retained under each operating condition bin. The method for removing samples that are outside the operating condition is as follows: it is determined whether each parameter in the sample is within the corresponding parameter range of the above operating condition bin. If so, it is retained; otherwise, it is removed.

[0072] Specifically, the physical gray box model described in step 3 is as follows:

[0073]

[0074] Among them, T c T represents the condenser outlet water temperature. e The outlet water temperature of the evaporator is c1~c 10 For the model parameters to be identified, COP ch Q is the energy efficiency ratio of the refrigeration unit.ch For the cooling capacity of the refrigeration unit, P ch This refers to the power of the refrigeration unit.

[0075] Specifically, the linear regression parameter fitting method described in step 3 adopts the median regression method.

[0076] The present invention also provides a refrigerator fault diagnosis system based on multi-model integration and adaptive learning. The refrigerator fault diagnosis system based on multi-model integration and adaptive learning can be implemented by executing the process steps of the refrigerator fault diagnosis method based on multi-model integration and adaptive learning. That is, those skilled in the art can understand the refrigerator fault diagnosis method based on multi-model integration and adaptive learning as a preferred embodiment of the refrigerator fault diagnosis system based on multi-model integration and adaptive learning.

[0077] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0078] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for refrigeration machine fault diagnosis based on multi-model ensemble and adaptive learning, characterized in that, include: Step S1: Obtain historical chiller operating parameters as a baseline sample, and obtain the chiller operating parameters to be judged as a judgment sample; A test case is constructed based on the current benchmark sample. It is then determined whether the current test sample is within the range of the current test case. If it is not within the range of the current test case, the current test sample is deleted. Otherwise, retain the result and trigger step S2. Step S2: Construct a refrigeration machine fault identifier and use the constructed refrigeration machine fault identifier to identify faults based on judgment samples; Step S3: Construct a physical gray box model of chiller power, and use the constructed physical gray box model of chiller power to predict chiller power based on the current predicted chiller power; construct the power prediction range based on the current predicted chiller power. The power of the chiller in the judgment sample is compared with the power prediction range; at the same time, the identification results of the chiller fault identifier are combined to determine whether there is an abnormality in the current chiller. The refrigeration unit fault identifier is based on a multi-layer fault sample set and uses multiple fault identification models to identify faults. Step S2 includes: Step S2.1: Obtain various types of refrigeration machine fault mode data, and divide them into hierarchical levels according to the dimensions of refrigeration machine fault status, fault location and fault mode to form a multi-level fault sample set; Step S2.2: Preprocess the fault sample data of each level in the multi-level fault sample set, and label the preprocessed fault sample data to form a fault sample set for each level. Step S2.3: Construct corresponding fault identification models based on multi-layer fault sample sets, and train the corresponding fault identification models using the fault sample dataset of each layer to obtain the trained fault identification models; Step S2.4: Obtain a fault identifyer by integrating the trained fault identification models at each layer; Step S2.3 includes: Step S2.3.1: Construct fault identification models based on ensemble learning models, deep neural networks, logistic regression, support vector machines, and K-nearest neighbors, respectively; Step S2.3.2: For any fault identification model, a greedy algorithm is adopted, with the multi-objective optimization function Score with category number penalty as the objective. The worst classification category is merged into other categories through iteration, the Score is calculated, and the merging method with the highest current Score is retained. This process is repeated until the Score no longer improves or the number of categories is less than or equal to the preset value, and the current trained fault identification model is obtained. Step S2.3.3: For any trained fault identification model, select the trained fault identification model with the highest Score as the final trained fault identification model. The multi-objective optimization function Score includes: in, N classes This represents the number of categories after the merger. N original This represents the original number of categories. Macro-F1 is a classification model evaluation metric suitable for imbalanced data with multiple classes, where the number of categories is used as a penalty coefficient.

2. The refrigerator fault diagnosis method based on multi-model ensemble and adaptive learning according to claim 1, characterized in that, The operating parameters of the refrigeration unit include: evaporator outlet water temperature, evaporator inlet water temperature, evaporator saturation temperature, evaporator saturation pressure, condenser outlet water temperature, condenser inlet water temperature, condenser saturation temperature, condenser saturation pressure, evaporator chilled water flow rate, current percentage, and power.

3. The refrigerator fault diagnosis method based on multi-model ensemble and adaptive learning according to claim 1, characterized in that, Step S1 includes: Historical operating parameters of the refrigeration unit are obtained and used as a baseline sample; operating parameters of the refrigeration unit to be judged are obtained and used as a judgment sample. The baseline sample is binned, with the standard deviation of each operating parameter as the bin size, forming several operating condition bins; a preset number of valid samples are retained in each operating condition bin; it is determined whether each parameter in the current judgment sample is within the corresponding parameter range of the operating condition bin, and if so, it is retained, otherwise it is discarded.

4. The refrigerator fault diagnosis method based on multi-model ensemble and adaptive learning according to claim 1, characterized in that, The physical gray box model of the chiller power includes: in, T c This refers to the condenser outlet water temperature. T e The evaporator outlet water temperature. c 1 ~c 10 For model parameters, COP ch This refers to the energy efficiency ratio of the refrigeration unit. Q ch For the cooling capacity of the refrigeration mechanism; P ch This refers to the power of the refrigeration unit.

5. The refrigerator fault diagnosis method based on multi-model ensemble and adaptive learning according to claim 1, characterized in that, Step S3 includes: Step S3.1: Use the benchmark samples to perform linear regression parameter fitting on the constructed chiller power physical gray box model to obtain the fitted chiller power physical gray box model; Step S3.2: Use the fitted physical gray box model of chiller power to predict the chiller power of the sample; based on the predicted chiller power, obtain the corresponding power prediction range according to the preset requirements; Step S3.3: Obtain the percentage of refrigeration units in the sample whose power exceeds the corresponding power prediction range within a preset time period. If the percentage exceeds the preset value and is identified as a fault by the fault identifier, then the current refrigeration unit is judged to be abnormal and the fault type is marked; otherwise, it is judged to be normal.

6. The refrigerator fault diagnosis method based on multi-model ensemble and adaptive learning according to claim 5, characterized in that, Step S3.1 includes: using benchmark samples to fit the constructed physical gray box model of chiller power to linear regression parameters using the median regression method, thereby obtaining the fitted physical gray box model of chiller power.

7. A refrigeration unit fault diagnosis system based on multi-model ensemble and adaptive learning, characterized in that, The refrigerator fault diagnosis method based on multi-model ensemble and adaptive learning as described in claim 1 includes: Module M1: Obtain historical chiller operating parameters as a baseline sample, obtain chiller operating parameters to be judged as a judgment sample; construct a test case based on the current baseline sample, determine whether the current judgment sample is within the range of the current test case, if not within the range of the current test case, delete the current judgment sample; otherwise, retain it and trigger the execution of module M2. Module M2: Construct a refrigeration machine fault identifier, and use the constructed refrigeration machine fault identifier to identify faults based on judgment samples; Module M3: Constructs a physical gray box model of chiller power, uses the constructed physical gray box model to predict chiller power in the judgment sample, constructs a power prediction range based on the currently predicted chiller power, compares the chiller power in the judgment sample with the power prediction range, and combines the identification results of the chiller fault identifier to determine whether there is an abnormality in the current chiller. The refrigeration unit fault identifier is based on a multi-layer fault sample set and utilizes multiple fault identification models to identify faults.

Images