Condenser scaling fault diagnosis method based on physical guidance and data driving

Abstract

The invention discloses a condenser scaling fault diagnosis method based on physical guidance and data driving, and belongs to the technical field of condenser scaling fault diagnosis, and the condenser scaling fault diagnosis method comprises the following steps: S1, collecting data of a data target domain, and constructing a fault data set together with data of a source domain; s2, respectively detecting the source domain data and the target domain data to obtain input data; preprocessing the input data to obtain steady-state data; s3, feature extraction; s4, performing feature alignment; s5, constructing and training a classifier; and S6, outputting fault diagnosis. According to the condenser scaling fault diagnosis method based on physical guidance and data driving, intelligent identification of the scaling degrees of the condensers in water chilling units of different structures is achieved by introducing a transfer learning and physical constraint mechanism, and visual display and maintenance suggestions are provided.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of condenser fouling fault diagnosis, and particularly relates to a condenser fouling fault diagnosis method based on physical guidance and data driving. BACKGROUND

[0002] Condenser fouling fault diagnosis is a technology for identifying the degree of condenser fouling, and the core goal is to realize intelligent identification of the degree of condenser fouling in different structure water chillers through technical means, so as to provide support for online fault diagnosis of industrial cooling systems and help predictive maintenance and reduce operation and maintenance costs.

[0003] However, the prior art has the following disadvantages: relying on a single data-driven model (such as LSTM, DANN), when migrating between heterogeneous devices, the feature space is offset due to the difference in sensor configuration, and the diagnostic accuracy decreases by more than 30%; and the feature distribution offset caused by the difference in device parameters is not considered, for example, when the source domain and target domain sensor configuration difference is greater than or equal to 30%, the accuracy of the traditional transfer learning method is reduced to less than 72%; the deep learning method needs to rely on a large amount of labeled data, but in actual industrial scenarios, fault samples of target devices are scarce, which easily causes model overfitting and is difficult to adapt to small sample scenarios; the existing data-driven model is a black box structure, and cannot associate fault features with the physical mechanism of condenser fouling, about 20% of false positive cases violate the laws of thermodynamics, resulting in low trust of operation and maintenance personnel on the diagnostic results and refusal to adopt the suggestions.

[0004] Therefore, a new method is urgently needed. SUMMARY

[0005] The purpose of the present application is to provide a condenser fouling fault diagnosis method based on physical guidance and data driving, which realizes intelligent identification of the degree of condenser fouling in different structure water chillers by introducing transfer learning and physical constraint mechanism, and provides visual display and maintenance suggestions.

[0006] To achieve the above purpose, the present application provides a condenser fouling fault diagnosis method based on physical guidance and data driving, comprising the following steps: S1, using the ASHRAE RP-1043 public data set, taking the condenser fouling fault level data as the source domain data, taking the sensor collected data as the target domain data, and using the source domain data and the target domain data to jointly construct a fault data set; S2, filtering and steady state detection are performed on the source domain data and the target domain data respectively to obtain input data; the input data are preprocessed to obtain preprocessed steady state data; S3, the preprocessed steady state data in S2 are input into an autoencoder, and feature extraction is performed on the source domain and target domain samples respectively to obtain low-dimensional feature representations of the source domain and the target domain; S4, aligning the feature distribution of the source domain and the target domain through the maximum mean difference method to obtain the aligned features; S5, training a classifier using the aligned features in S4, and the loss function of the classifier includes cross-entropy loss, maximum mean difference loss and physical consistency loss; S6, judging and classifying the condenser fouling state according to the trained classifier in S5, and outputting the fault level.

[0007] Preferably, in S2, the filtering processing adopts the Savitzky-Golay filtering algorithm, the sliding window width is set to 25 points, and the fitting polynomial order is 2.

[0008] Preferably, in S2, the steady state detection adopts a sliding window steady state determination method, and the condenser inlet and outlet temperatures, the refrigerant outlet temperature and the like are used as steady state indicators, the time window is set to 180 seconds, and the signal change amplitude is less than ±0.3℃.

[0009] Preferably, S3 further includes minimizing the reconstruction error between the input and the reconstructed output, and the calculation method is as follows: ; Wherein, is the reconstruction loss; is the input data; is the data reconstructed by the decoder.

[0010] Preferably, in S5, the calculation method of the maximum mean difference loss is as follows: ; Wherein, is the maximum mean difference loss; is the square of the maximum mean difference.

[0011] Preferably, in S5, the calculation method of the physical consistency loss is as follows: ; Wherein, is the heat transfer performance index; is the condenser heat exchange capacity; is the effective heat exchange area of the condenser; is the cooling water temperature difference.

[0012] The application also provides a condenser fouling fault diagnosis system based on physical guidance and data driving, comprising: A data acquisition module adopts the ASHRAE RP-1043 public data set, takes the condenser fouling fault level data as the source domain data, takes the sensor acquisition data as the target domain data, and uses the source domain data and the target domain data to jointly construct a fault data set; The data preprocessing module is connected with the data acquisition module, and is used for filtering and steady-state detection on the source domain data and the target domain data respectively to obtain input data; and pre-processing the input data to obtain steady-state data after pre-processing. The feature alignment module is connected with the data preprocessing module, and aligns the feature distribution of the source domain and the target domain by using a maximum mean difference method. The classifier module is connected with the feature alignment module, and is used for constructing a classifier. The fault diagnosis output module is connected with the classifier module, and is used for outputting a condenser fouling fault level according to the trained classifier.

[0013] Therefore, the condenser fouling fault diagnosis method based on physical guidance and data driving has the following beneficial effects compared with the prior art: (1) The automatic encoder is used to extract low-dimensional features of the source domain and the target domain by using the migration learning mechanism, and the cross-domain feature alignment is realized by using the maximum mean difference, so that the feature distribution difference is minimized, the migration diagnosis accuracy of the model between different structure water chillers is significantly improved, and the performance decline caused by equipment difference is reduced. (2) The knowledge migration of the source domain standard data is used to reduce the dependence on a large number of labeled samples of the target domain, so that a high diagnosis accuracy can be maintained in the case of a lack of target domain samples, and model overfitting is avoided. (3) The neural network classifier is constructed by using a physical consistency loss function, the condenser heat transfer performance index is used as a core constraint, the prediction results inconsistent with the physical law are punished, the diagnosis results conform to the basic law of thermodynamics, the false positives are reduced, and the trust degree of the operation and maintenance personnel on the diagnosis results is improved.

[0014] The technical solutions of the present application will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 The flowchart of the condenser fouling fault diagnosis method based on physical guidance and data driving of the present application is shown. DETAILED DESCRIPTION

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used in the present invention should have the ordinary meaning understood by those skilled in the art.

[0017] Example 1 like Figure 1 As shown, the condenser scaling fault diagnosis method based on physical guidance and data drive of the present invention includes the following steps: S1. Select publicly available industry standard datasets, such as ASHRAE RP-1043 centrifugal chiller unit operation data. This dataset contains 66-dimensional time-series features, including data for all operating conditions such as normal condenser operation, light scaling, moderate scaling, and heavy scaling. By filtering samples labeled with "condenser scaling level," source domain data is constructed as the basis for training the model, providing extensive preliminary experience references for target domain diagnosis.

[0018] Sensors are deployed on the actual chiller unit to be diagnosed to collect operating parameters, including cooling water inlet and outlet temperatures and flow rates, condenser pressure and temperature, etc. The data collection process needs to cover the equipment's daily operation and different load conditions to construct the target domain data; the source domain data and the target domain data together construct the fault dataset.

[0019] S2. The Savitzky-Golay (SG) filtering algorithm is used to smooth the signals of each channel of the chiller unit. The sliding window width is set to 25 points and the order of the fitting polynomial is 2. It is suitable for key parameters such as cooling water inlet temperature (TCI), outlet temperature (TCO), refrigerant outlet temperature (TEI), system load, and flow rate. It can effectively remove high-frequency noise, preserve the signal edge trend, and improve data quality.

[0020] A sliding window steady-state determination method is introduced, using three temperature signals—TCI, TCO, and TEI—as steady-state indices. A time window of 180 seconds is set. If the variation amplitude of the three signals is less than ±0.3℃ within this time window, the data segment is determined to be steady-state data. The extracted steady-state segments are uniformly trimmed to a fixed length and standardized to obtain preprocessed steady-state data. By screening stable operating data segments and eliminating interference such as start-stop fluctuations, the impact of non-steady-state data on the model is reduced. S3. Input the preprocessed steady-state data from S2 into the Autoencoder model; extract nonlinear features from the source and target domain samples respectively. The autoencoder contains two parts: an encoder and a decoder: The encoder compresses the input features into a latent low-dimensional representation , implemented by a multi-layer fully connected neural network, with the following computation: ; where is the encoding function; is the activation function; is the weight matrix of the encoder; is the bias vector of the encoder; The decoder reconstructs the compressed representation into the original input dimension, with the following computation: ; where is the decoder function; is the weight matrix of the decoder; is the bias term of the decoder; The training objective is to minimize the reconstruction error between the input and the reconstructed output, with the following computation: ; where is the reconstruction loss; is the input data; is the data reconstructed by the decoder; In the condenser fouling diagnosis scenario, the reconstruction loss indirectly constrains the model to capture features related to heat transfer mechanisms (such as cooling water temperature difference, heat exchange, etc.), ultimately obtaining low-dimensional feature representations for the source domain and the target domain; The constrained encoder extracts key features to ensure that the low-dimensional features reflect the core information of the original data; S4, use the maximum mean difference (MMD) method to align the feature representations of the source domain and the target domain. MMD measures the distribution distance between sample sets by calculating the mean difference in the reproducing kernel Hilbert space (RKHS), with the following computation: ; where is the MMD squared distance between the source domain and the target domain; is the source domain dataset; is the target domain dataset; is the number of source domain samples; is the number of target domain samples; is the low-dimensional feature representation of the source domain sample; is the low-dimensional feature representation of the target domain sample; ​is the L2 norm square in RKHS space; by quantifying and minimizing the cross-domain feature distribution difference, the migration ability of the model between heterogeneous devices is improved.

[0021] The inner product is approximated by using a radial basis function (RBF) kernel, and the calculation method is as follows: ; Wherein, is the kernel function; is the input sample vector; The MMD value is minimized by back propagation to realize the distribution consistency of source / target features in the latent space; S5, using the aligned features in S4 to train a support vector machine classifier with physical constraint ability, which is used for classification and identification of condenser fouling state; The loss function of the classifier consists of the following three parts: Cross-entropy loss, calculation method as follows: ; Wherein, is the classification loss; is the true label of the first sample; is the predicted probability of the first sample; The calculation method of MMD loss is as follows: ; Wherein, is the maximum mean difference loss; is the square of the maximum mean difference; The physical consistency loss takes the heat transfer performance index as the constraint term in the loss calculation, which is used to punish the prediction results that are inconsistent with the physical law, and the calculation method is as follows: ; Wherein, is the heat transfer performance index; is the condenser heat transfer capacity; is the effective heat transfer area of the condenser; is the cooling water temperature difference; the heat transfer equation is converted into a calculable feature quantity , so that the data-driven model is combined with the physical prior, the heat transfer mechanism constraint is introduced into the model by quantifying the heat transfer performance, and the diagnosis result is ensured to meet the thermodynamic law, and the false alarm that violates the physical law is reduced; The physical loss is constructed as the deviation term of the predicted value and the physical index: ; Wherein, is the physical consistency loss; To derive the label based on the physical mechanism; To the Heat transfer performance index of the first The final multi-objective joint loss function is calculated as follows: ; Wherein, , , Respectively, the adjustable weight coefficient; The joint loss function ensures that the final diagnosis result not only matches the sensor data, but also meets the physical indicators such as heat transfer coefficient and heat exchange capacity; S6, the model trained is used for inference stage, the fault identification of target domain condenser operation data, the model output includes classification result including no scaling, mild, moderate and severe scaling.

[0022] Therefore, the present application adopts the above-mentioned condenser scaling fault diagnosis method based on physical guidance and data driving, which realizes the intelligent identification of the scaling degree of the condenser in the cold water unit with different structures by introducing the transfer learning and physical constraint mechanism, and provides visual display and maintenance suggestions.

[0023] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, but not to limit it, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: it can still modify or replace the technical solutions of the present application, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A method for diagnosing condenser scaling faults based on physical guidance and data-driven approaches, characterized in that, Includes the following steps: S1. Using the ASHRAE RP-1043 public dataset, condenser scaling fault level data is used as the source domain data; sensor-collected data is used as the target domain data; the source domain data and target domain data are used together to construct the fault dataset. S2. Filter and perform steady-state detection on the source domain data and target domain data respectively to obtain the input data; The input data is preprocessed to obtain preprocessed steady-state data. S3. Input the preprocessed steady-state data from S2 into the autoencoder and extract features from the source and target domain samples respectively to obtain low-dimensional feature representations of the source and target domains. S4. Align the feature distributions of the source and target domains using the maximum mean difference method to obtain aligned features; S5. Train the classifier using the aligned features from S4. The loss function of the classifier includes cross-entropy loss, maximum mean difference loss and physical consistency loss. S6. Based on the classifier trained in S5, determine and classify the scaling status of the condenser and output the fault level.

2. The condenser scaling fault diagnosis method based on physical guidance and data driving according to claim 1, characterized in that, In S2, the Savitzky-Golay filtering algorithm is used for filtering, and the sliding window width is set to 25 points, with the order of the fitted polynomial being 2.

3. The condenser scaling fault diagnosis method based on physical guidance and data driving according to claim 1, characterized in that, In S2, the steady-state detection adopts the sliding window steady-state judgment method, using the three temperature signals of condenser inlet and outlet temperature and refrigerant outlet temperature as steady-state indicators, setting the time window to 180 seconds, and the signal change amplitude is less than ±0.3℃.

4. The condenser scaling fault diagnosis method based on physical guidance and data driving according to claim 1, characterized in that, S3 also includes minimizing the reconstruction error between the input and the reconstructed output, calculated as follows: ； in, For reconstruction loss; Input data; The data reconstructed by the decoder.

5. The condenser scaling fault diagnosis method based on physical guidance and data driving according to claim 1, characterized in that, The method for calculating the maximum mean difference loss in S5 is as follows: ； in, This represents the loss due to the maximum mean difference. It is the square of the maximum mean difference.

6. The condenser scaling fault diagnosis method based on physical guidance and data driving according to claim 1, characterized in that, The calculation method for physical consistency loss in S5 is as follows: ； in, The heat transfer performance index; Heat exchange for the condenser; The effective heat exchange area of ​​the condenser; This refers to the temperature difference of the cooling water.

7. A condenser scaling fault diagnosis system based on physical guidance and data-driven methods, used to implement the condenser scaling fault diagnosis method based on physical guidance and data-driven methods according to any one of claims 1-6, characterized in that, include: The data acquisition module uses the ASHRAE RP-1043 public dataset and takes the condenser scaling fault level data as the source domain data. Use sensor-collected data as target domain data; A fault dataset is constructed using source domain data and target domain data. A data preprocessing module, connected to the data acquisition module, is used to perform filtering and steady-state detection on the source domain data and the target domain data respectively to obtain input data; The input data is preprocessed to obtain preprocessed steady-state data; The feature alignment module, connected to the data preprocessing module, aligns the feature distributions of the source and target domains using the maximum mean difference method. A classifier module, connected to the feature alignment module, is used to construct a classifier; The fault diagnosis output module is connected to the classifier module and is used to output the condenser scaling fault level based on the trained classifier.

Images