Water chilling unit fault prediction and diagnosis method based on GRU-SA model

Abstract

The invention relates to the technical field of fault prediction, in particular to a water chilling unit fault prediction and diagnosis method based on a GRU-SA model. The method comprises the following steps: acquiring a multi-mode water chilling unit data set; performing data preprocessing on the multi-modal water chilling unit data set to generate a multi-modal water chilling unit preprocessing data set; physical knowledge virtual feature enhancement is carried out on the multi-modal water chilling unit preprocessing data set, and a multi-modal water chilling unit virtual feature set is generated; performing dynamic confrontation data sample construction on the virtual feature set of the multi-modal water chilling unit to generate multi-modal water chilling unit model input data; therefore, through the technical means of multi-modal data processing, physical knowledge enhancement, dynamic adversarial data sample generation, GRU-SA model optimization, cloud-edge collaborative reasoning and the like, the problems of insufficient accuracy, poor robustness and the like of a traditional water chilling unit fault diagnosis model are solved, and the accuracy and the real-time performance of fault prediction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault prediction, and in particular to a chiller fault prediction and diagnosis method based on a GRU-SA model. Background Art

[0002] Traditional chiller fault diagnosis methods rely heavily on static data analysis and are unable to effectively address the dynamic changes in chiller operating conditions. The Gated Recurrent Unit (GRU), a structure based on recurrent neural networks (RNNs), aims to address the vanishing and exploding gradient problems encountered by traditional RNNs when processing long sequences of data. Compared to long short-term memory (LSTM) networks, the GRU structure is simpler and contains two main gating mechanisms: a reset gate and an update gate. The reset gate controls how past information is combined with current input, while the update gate determines how the current state influences future information. This enables the GRU to effectively capture long-term dependencies when processing time series data, making it suitable for fault diagnosis problems with time series characteristics, such as chiller faults. Self-Attention (SA), a widely used mechanism in deep learning, captures global data dependencies by calculating the relative importance of different parts of the input data. In traditional recurrent neural networks, the transmission of information mainly depends on the order of time steps, while the self-attention mechanism can more flexibly assign weights to different parts of the data, so that when processing sequence data, it can better focus on the parts that are important to the current task. The GRU-SA model combines GRU with the self-attention mechanism. It can not only use GRU to capture temporal dependencies, but also capture the complex associations between data in a wider time dimension through the self-attention layer, improving the model's ability to handle dynamic changes and complex relationships. Although the GRU-SA model has shown good performance in many fields, in the existing technology, fault diagnosis systems are usually only analyzed through traditional statistical methods or based on single-modal data. These methods cannot handle the complex associations of multimodal data and cannot effectively utilize the advantages of the GRU-SA model. In addition, the lack of dynamic adjustment and adaptive mechanisms makes the existing methods limited in their processing effect on high-dimensional, nonlinear time series data such as chillers, resulting in insufficient accuracy in fault prediction. Summary of the Invention

[0003] Based on this, it is necessary to provide a chiller fault prediction and diagnosis method based on the GRU-SA model to solve at least one of the above technical problems.

[0004] To achieve the above object, a chiller fault prediction and diagnosis method based on the GRU-SA model is provided, the method comprising the following steps:

[0005] Step S1: obtaining a multimodal chiller dataset; performing data preprocessing on the multimodal chiller dataset to generate a multimodal chiller preprocessed dataset;

[0006] Step S2: Performing physical knowledge virtual feature enhancement on the multimodal chiller preprocessed dataset to generate a multimodal chiller virtual feature set; performing dynamic adversarial data sample construction on the multimodal chiller virtual feature set to generate multimodal chiller model input data;

[0007] Step S3: The multimodal chiller model input data is input to the preset GRU convolutional model, and the self-attention layer is used to calculate the feature correlation matrix to generate a GRU-SA chiller preliminary model; the GRU-SA chiller preliminary model is adaptively adjusted to generate a GRU-SA chiller fault judgment model;

[0008] Step S4: Deploy the GRU-SA chiller fault judgment model on the cloud, perform edge-cloud collaborative reasoning, and generate GRU-SA chiller fault judgment data; draw a contribution heat map of the GRU-SA chiller model output data, thereby completing the chiller fault prediction and diagnosis based on the GRU-SA model.

[0009] The present invention achieves the beneficial effect of ensuring data consistency and accuracy by acquiring and preprocessing a multimodal chiller dataset, enabling subsequent analysis to be based on high-quality data. Data preprocessing not only eliminates noise but also converts the data into a standardized format, providing reliable input for feature enhancement and analysis. Next, a physical knowledge-based virtual feature enhancement technique is employed to generate a multimodal chiller virtual feature set by integrating the physical characteristics of chiller operation. This process provides the model with richer feature information, enabling it to capture deeper operational patterns, particularly in complex environments, effectively extracting potential fault-related patterns. A dynamic adversarial data sample construction method is employed to expand and optimize the multimodal chiller virtual feature set. The generation of adversarial samples enables the model to better cope with various disturbances and improves its ability to identify abnormal conditions and potential faults. This approach enhances the model's robustness, ensuring its stability and accuracy in actual operation. A preliminary GRU-SA chiller model is generated by inputting the constructed input data into a pre-set GRU convolutional model and combining it with a self-attention (SA) mechanism to calculate the feature correlation matrix. The model can further optimize the feature weights through adaptive parameter adjustment, thereby forming a fault diagnosis model with higher accuracy. The trained GRU-SA chiller fault diagnosis model is deployed on the cloud and combined with edge computing for collaborative reasoning, which further improves the data processing efficiency and real-time performance. The collaborative reasoning between the cloud and the edge enables the model to achieve efficient data transmission and processing between different computing nodes, thereby quickly responding to the fault prediction needs in actual operation. In addition, by drawing a contribution heat map, the output results of the model are visualized, and the contribution of each feature to fault diagnosis is intuitively displayed, which helps decision makers to quickly locate the source of the problem. Therefore, the present invention solves the problems of insufficient accuracy and poor robustness of traditional chiller fault diagnosis models through technical means such as multimodal data processing, physical knowledge enhancement, dynamic adversarial data sample generation, GRU-SA model optimization, and cloud-edge collaborative reasoning, thereby improving the accuracy and real-time performance of fault prediction.

[0010] Preferably, step S1 includes the following steps:

[0011] Step S11: Acquire a multimodal chiller dataset, wherein the multimodal chiller dataset includes a chiller vibration signal, a chiller temperature signal, and a chiller pressure data;

[0012] Step S12: performing wavelet packet transform denoising on the unit vibration signal to generate unit wavelet transform denoised data; performing data standardization processing on the unit wavelet transform denoised data to generate unit vibration standardised data;

[0013] Step S13: performing time series drift processing on the unit temperature signal and the unit pressure data to generate unit temperature normalized data and unit pressure normalized data;

[0014] Step S14: synchronize the unit vibration normalized data, the unit temperature normalized data, and the unit pressure normalized data in a time window to generate a multi-modal chiller preprocessing data set.

[0015] The present invention acquires a multimodal chiller dataset containing vibration signals, temperature signals, and pressure data. These data reflect the chiller's operating status and performance changes and serve as an important foundation for fault diagnosis and performance evaluation. Because various signals are susceptible to noise interference during sensor acquisition, the chiller's vibration signals are subjected to wavelet packet transform denoising. Wavelet packet transform effectively removes high-frequency noise from the signal while retaining its effective information, generating clear and representative chiller vibration signal data. This provides a more accurate data foundation for subsequent analysis. Furthermore, the denoised vibration signal data is standardized to effectively eliminate differences between different dimensions and amplitudes, bringing the data into a unified scale and improving the comparability of different features during model training. During the processing of the chiller's temperature and pressure data, drift caused by temperature and pressure measurement errors or environmental factors is eliminated through time-series drift processing, ensuring the stability and consistency of these data. Standardizing these two types of data further eliminates dimensional differences between different sensors, enabling effective comparison and fusion of temperature and pressure data with vibration signal data, thereby ensuring the uniformity and reliability of the data processing process. Finally, time window synchronization was used to align the standardized unit vibration data, unit temperature data, and unit pressure data to form a multimodal chiller preprocessing dataset. This time window synchronization ensures temporal consistency among the different signals, enabling subsequent time series-based analysis methods to effectively extract features and train models. Furthermore, this synchronization avoids data inconsistencies caused by time offsets, improving the model's prediction accuracy and reliability.

[0016] Preferably, step S13 includes the following steps:

[0017] Step S131: Performing a two-dimensional linear projection on the unit temperature signal and the unit pressure data to generate a two-dimensional projection of the unit temperature and pressure; performing linear interpolation outlier correction on the two-dimensional projection of the unit temperature and pressure to generate unit temperature and pressure outlier processed data;

[0018] Step S132: performing white correlation trend analysis on the unit temperature-pressure abnormal value processing data to generate unit temperature-pressure white correlation trend data;

[0019] Step S133: performing time series drift processing on the unit temperature-pressure white correlation trend data to generate unit temperature-pressure time series drift data; performing second-order difference normalization processing on the unit temperature-pressure time series drift data to generate unit temperature normalized data and unit pressure normalized data.

[0020] The present invention converts the unit temperature signal and unit pressure data into a two-dimensional unit temperature-pressure projection by performing a two-dimensional linear projection. This process effectively maps the raw data from a high-dimensional space to a low-dimensional space, simplifying the data complexity while preserving the potential correlation between temperature and pressure. The data after the two-dimensional linear projection can more intuitively reveal the relationship between temperature and pressure, providing a foundation for subsequent data analysis. Furthermore, linear interpolation outlier correction is performed on the two-dimensional unit temperature-pressure projection to eliminate outliers caused by sensor errors or external disturbances, thereby generating more stable and reliable data. This step significantly reduces data bias and improves data quality. White correlation trend analysis is performed on the processed unit temperature-pressure outlier data. This method effectively removes random fluctuations in temperature and pressure data by extracting the trend component of the data, allowing the long-term trends in the data to be clearly presented. White correlation trend analysis can eliminate periodicity and non-stationarity in the data, making subsequent data processing more accurate and providing more stable input for model prediction. Furthermore, white correlation trend analysis helps reveal the changing patterns of unit temperature and pressure over long time series, thereby identifying key factors related to equipment status. The time drift of the temperature data is further eliminated through time series drift processing, thereby ensuring the consistency and stability of the data throughout the time series. This processing step effectively removes the drift of the temperature signal caused by changes in the external environment or equipment failures, so that the data can accurately reflect the performance of the unit under normal operating conditions. Afterwards, the temperature data is subjected to second-order difference standardization processing, which can effectively eliminate trend changes and periodic fluctuations in the temperature signal and convert it into standardized differential data, so that the data dimension can be unified and the model can better learn subtle changes in the temperature signal. In addition, this differential standardization processing also improves the stability of the data, providing more accurate prediction input for the model.

[0021] Preferably, the physical knowledge virtual feature enhancement of the multimodal chiller preprocessing data set comprises the following steps:

[0022] Conduct correlation analysis of the refrigeration cycle physical process based on the unit vibration standardization data to generate the unit cold cycle vibration analysis data; conduct indirect refrigeration cycle efficiency impact analysis on the unit cold cycle vibration analysis data to generate the unit vibration theoretical energy efficiency ratio;

[0023] Based on the unit temperature standardization data, the temperature fluctuation pattern is used to identify potential equipment failures and generate the unit temperature fluctuation pattern data; the unit temperature fluctuation pattern data is subjected to least squares regression physical constraint dimensionality reduction to generate the unit temperature subcooling deviation;

[0024] Perform thermodynamic state analysis on the standardized unit pressure data to generate the unit pressure thermodynamic state curve; perform outlier peak analysis on the unit pressure thermodynamic state curve to generate unit pressure thermodynamic abnormal peak data; perform temperature-pressure correlation analysis on the unit pressure thermodynamic abnormal peak data and the unit temperature subcooling deviation to generate the unit temperature-pressure subcooling deviation.

[0025] One-hot encoding is performed based on the temperature-pressure subcooling deviation of the unit and the theoretical energy efficiency ratio of the unit vibration, and physical prior processing is performed to generate a multi-modal chiller virtual feature set.

[0026] By performing a correlation analysis of the refrigeration cycle physical process on the standardized vibration data of the unit, the present invention can effectively extract vibration characteristics closely related to the refrigeration cycle performance, thereby generating the unit cold cycle vibration analysis data. This analysis process reveals the intrinsic relationship between the unit vibration and its cold cycle process, providing a key basis for further performance evaluation and fault prediction. Next, by performing an indirect refrigeration cycle efficiency impact analysis on the unit cold cycle vibration analysis data, the unit vibration theoretical energy efficiency ratio is generated. This data reflects the energy efficiency performance of the unit during actual operation and provides a quantitative indicator for judging the unit's operating efficiency. During the processing of temperature and pressure data, temperature fluctuation pattern recognition is first performed on the standardized temperature data of the unit to accurately capture potential equipment failure signals and generate unit temperature fluctuation pattern data. Through this pattern recognition, the system can detect abnormal trends in temperature changes in advance and thus identify the occurrence of faults. In addition, after the unit temperature fluctuation pattern data is processed through least squares regression physical constraint dimensionality reduction, the unit temperature subcooling deviation is generated. This step effectively reduces data dimensionality and eliminates noise interference, allowing temperature data to more accurately reflect temperature subcooling deviations, thus providing more precise data input for unit fault warnings. For the standardized unit pressure data, a thermodynamic state analysis is first performed to generate the unit pressure thermodynamic state curve. This analysis reveals the thermodynamic characteristics of the unit under different operating conditions and provides a basis for understanding the unit's operating efficiency. An outlier peak analysis of the pressure thermodynamic state curve generates unit pressure thermodynamic anomaly peak data. This data highlights extreme values ​​or abnormal variations in the pressure data, providing effective clues for fault detection. Furthermore, a temperature-pressure correlation analysis is performed between the unit pressure thermodynamic anomaly peak data and the unit temperature subcooling deviation to generate the unit temperature-pressure subcooling deviation. This analysis helps reveal the mutual influence between unit temperature and pressure, particularly the dynamic variation patterns under abnormal conditions. Finally, by one-hot encoding the unit temperature-pressure subcooling deviation and the unit vibration theoretical energy efficiency ratio and combining it with physical priors, a multimodal chiller virtual feature set is generated. This feature set integrates key information from multiple sensor data, providing rich and efficient feature input for subsequent model training and fault prediction. By encoding and physically processing various features, the model can better learn the underlying patterns in unit operation, further improving the system's fault identification accuracy.

[0027] Preferably, step S2 includes the following steps:

[0028] Step S21: performing physical knowledge virtual feature enhancement on the multimodal chiller preprocessing data set to generate a multimodal chiller virtual feature set;

[0029] Step S22: performing random noise perturbation on the multimodal chiller virtual feature set to generate multimodal random noise perturbation data; performing gradient descent on the multimodal random noise perturbation data to generate multimodal gradient descent data;

[0030] Step S23: scramble the multimodal gradient descent data to generate multimodal gradient descent scrambled data; perform time series perturbation analysis on the multimodal gradient descent scrambled data, thereby completing the construction of dynamic adversarial data samples and generating multimodal chiller model input data.

[0031] This invention enhances the physical features of a multimodal chiller preprocessed dataset using virtual features informed by physical knowledge. Leveraging prior knowledge from the physical field, it systematically enhances the physical feature representation of the chiller data, generating a multimodal chiller virtual feature set. This process effectively combines the physical characteristics of chiller operation with sensor data, providing richer feature information for subsequent model learning. These virtual feature sets more accurately reflect the actual operating status of the chiller, improving the accuracy of subsequent analysis. First, the multimodal chiller virtual feature set is perturbed with random noise to generate multimodal random noise perturbation data. This step artificially introduces noise, simulating interference factors encountered in actual operation and improving the dataset's adaptability to noise and abnormal changes. Noise perturbation not only enhances data diversity but also enables the model to better cope with real-world uncertainties during training. Next, gradient descent is applied to this perturbed data to generate multimodal gradient descent data. Gradient descent optimizes model parameters, allowing the dataset to continuously approach the optimal solution during training, thereby improving the data's feature representation and further enhancing the model's performance and efficiency. The multimodal gradient descent data was scrambled to generate multimodal gradient descent scrambled data. This data scrambling operation further increased the complexity of the dataset, enabling the model to handle more variations when processing this data, avoiding overfitting and improving generalization. Subsequently, time series perturbation analysis was performed to complete the construction of dynamic adversarial data samples. This time series perturbation analysis not only considers temporal variations but also simulates abnormal fluctuations and trend changes in the data over time. Through these steps, multimodal chiller model input data was generated, fully accounting for various real-world interference and abnormal scenarios, further enhancing the diversity and robustness of the dataset.

[0032] Preferably, step S3 includes the following steps:

[0033] Step S31: Input the multimodal chiller model input data into the preset GRU convolutional model, which contains a GRU layer with 64 hidden units, and calculates the feature correlation matrix of the self-attention layer, outputs the hidden state of the GRU layer and the self-attention layer, and constructs a preliminary model of the GRU-SA chiller. The calculation formula for the feature correlation matrix includes:

[0034] ;

[0035] in, , , are query, key, and value matrices respectively, is the feature dimension;

[0036] Step S32: performing cross entropy loss function expansion on the GRU-SA chiller preliminary model to generate a GRU-SA energy regularization loss function;

[0037] Step S33: Adaptively adjust the parameters of the GRU-SA chiller preliminary model and the GRU-SA energy regularization loss function to generate a GRU-SA chiller fault judgment model, wherein the weight of the GRU-SA model is adjusted by the Adam optimizer, and the learning rate is set to 0.001.

[0038] The present invention generates a preliminary model of the GRU-SA chiller by taking the input data of the multimodal chiller model as input to a preset GRU convolutional model and combining it with the self-attention layer to calculate the feature correlation matrix. In this process, the introduction of the GRU convolutional model enables the long-term and short-term dependencies of time series data to be effectively captured, thereby enhancing the dynamic features hidden in the data. At the same time, the self-attention layer can automatically capture key features in the input of multimodal data by calculating the correlation matrix between features, reducing the interference of redundant information. In this way, the model can effectively extract the most representative information from complex multimodal data, further enhancing the accuracy and efficiency of chiller fault prediction. The cross-entropy loss function of the GRU-SA chiller preliminary model is expanded, and the GRU-SA energy regularization loss function is generated. The cross-entropy loss function is usually used for classification problems and can quantify the gap between the predicted value and the true label. The expansion of the cross-entropy loss function enables it to adapt to more types of data features, further optimizing the loss function design of the model. Furthermore, the introduction of an energy regularization loss function effectively controls model complexity and prevents overfitting during training. Energy regularization, as a constraint, effectively limits model parameters, ensuring the model's good generalization capabilities when processing complex data. The GRU-SA chiller preliminary model and the GRU-SA energy regularization loss function were adaptively adjusted to generate a GRU-SA chiller fault diagnosis model. Adaptive parameter adjustment enables the model to automatically optimize model parameters based on the actual data performance during training, improving training efficiency and accuracy. This step further enhances the model's self-regulation capabilities, enabling it to maintain high prediction accuracy under varying operating conditions and data distributions. Through this optimization process, the resulting GRU-SA chiller fault diagnosis model is able to efficiently and accurately diagnose faults in complex and changing data environments, demonstrating strong robustness and adaptability.

[0039] Preferably, the self-attention layer performs feature correlation matrix calculation including the following steps:

[0040] Obtain chiller refrigeration drawings and real-time refrigeration operating parameters;

[0041] Perform 3D modeling on the chiller refrigeration drawings to generate the chiller refrigeration cycle physical structure; construct the refrigeration system topology nodes on the chiller refrigeration cycle physical structure to generate the chiller topology nodes;

[0042] Perform real-time working condition encoding based on the chiller topology nodes and real-time refrigeration working condition parameters to generate dynamic unit working condition drive encoding data; perform sparse matrix correlation matrix synthesis on the dynamic unit working condition drive encoding data to generate a chiller sparse correlation matrix;

[0043] The sparse correlation matrix of the chiller is calculated using the physical constraint self-attention process and irrelevant noise masking is performed to complete the feature correlation matrix calculation of the attention layer.

[0044] By acquiring chiller refrigeration drawings and real-time refrigeration operating parameters, the present invention provides sufficient physical and operational context for subsequent data processing. By performing three-dimensional modeling on the chiller refrigeration drawings, the physical structure of the chiller's refrigeration cycle can be accurately constructed. This process provides a basic framework for subsequent topological node construction. By constructing the refrigeration system topology nodes based on the chiller's refrigeration cycle physical structure, the connections and functions between various devices in the system can be clearly displayed, providing important structural information for further data analysis and fault diagnosis. Real-time operating condition encoding is performed based on the chiller topology nodes and real-time refrigeration operating condition parameters to generate dynamic unit operating condition drive encoding data. In this process, the introduction of real-time operating condition parameters enables the model to dynamically adjust according to actual operating conditions and achieve real-time response. The generated dynamic unit operating condition drive encoding data not only contains the current operating status of the equipment, but also reflects the workload and energy efficiency performance of each component under different operating conditions, providing accurate input data for further fault prediction and performance optimization. A sparse matrix correlation matrix is ​​generated by synthesizing the dynamic unit operating condition drive encoding data. The introduction of sparse matrices can effectively reduce the complexity of data storage and calculation, while retaining key feature information and improving the efficiency of the model when processing large-scale data. The sparse association matrix of the chiller quantifies the correlation between each data point, providing an effective data structure for subsequent feature calculations. The sparse association matrix of the chiller is calculated using a physical constraint self-attention process and subjected to irrelevant noise masking to complete the feature association matrix calculation of the attention layer. The physical constraint self-attention process calculation ensures that the model can identify features closely related to unit failures and performance from the data by considering the physical characteristics of the chiller. Through irrelevant noise masking, redundant information and noise interference can be effectively filtered out, so that the final generated feature association matrix can be more focused on key features, improving the accuracy and stability of the model under complex working conditions.

[0045] Preferably, step S33 includes the following steps:

[0046] Step S331: performing time dimension association on the GRU-SA chiller preliminary model and the GRU-SA energy regularization loss function to generate GRU-SA chiller time dimension data; performing feature dimension association on the GRU-SA chiller preliminary model and the GRU-SA energy regularization loss function to generate GRU-SA chiller feature dimension data;

[0047] Step S332: performing multi-level association aggregation on the GRU-SA chiller time dimension data and the GRU-SA chiller feature dimension data to generate GRU-SA chiller time-feature dimension data;

[0048] Step S333: inject the fault type prior into the GRU-SA chiller time-feature dimension data to generate chiller fault type data; perform temporal fault causality protection on the chiller fault type data, and perform adaptive parameter adjustment on the refrigerant evaporator and compressor suction pipe to generate a GRU-SA chiller fault judgment model.

[0049] The present invention generates GRU-SA chiller time dimension data by associating the GRU-SA chiller preliminary model and the GRU-SA energy regularization loss function in the time dimension. The association processing of the time dimension enables the model to capture the dynamic changes of the chiller at different time nodes and identify potential trends and abnormal changes in the time series. This processing provides a key information basis for subsequent fault diagnosis, ensuring that the model can analyze the changes in the equipment operating status from the perspective of the time series. Then, the GRU-SA chiller preliminary model and the GRU-SA energy regularization loss function are associated in the feature dimension to generate GRU-SA chiller feature dimension data. The association of the feature dimension enables the model to comprehensively consider the mutual influence between multiple sensor data and physical characteristics, and enhances the model's ability to comprehensively analyze the chiller operating status from multiple dimensions. The GRU-SA chiller time dimension data and the GRU-SA chiller feature dimension data are multi-level associated and aggregated to generate GRU-SA chiller time-feature dimension data. Through multi-level association and aggregation, the model not only captures individual information in the time and feature dimensions but also identifies the interactions and inherent connections between them, thereby enhancing its ability to process complex multidimensional data. This step effectively integrates time series features with various physical information about the equipment's operating status, enabling the resulting time-feature dimension data to comprehensively reflect the chiller's operating status. Chiller fault type data was generated by injecting prior knowledge of fault types into the GRU-SA chiller's time-feature dimension data. This process, incorporating prior knowledge of fault types, provides the model with clear fault category information, guiding it to more accurately identify different fault types during fault diagnosis. Next, temporal fault causality assurance was performed, and adaptive parameter adjustments were made to the refrigerant evaporator and compressor suction pipe, further optimizing the fault diagnosis model. Temporal fault causality assurance ensures that the model adheres to physical causal relationships during fault prediction, effectively eliminating unreasonable fault assumptions. Furthermore, adaptive parameter adjustments enable the model to flexibly adapt to the actual operating environment, improving its robustness and accuracy under variable operating conditions.

[0050] Preferably, step S4 includes the following steps:

[0051] Step S41: Deploy the GRU-SA chiller fault judgment model on the cloud to generate a GRU-SA chiller fault cloud judgment model; perform edge-cloud collaborative reasoning on the GRU-SA chiller fault cloud judgment model to generate GRU-SA chiller fault judgment data;

[0052] Step S42: Draw a contribution heat map for the output data of the GRU-SA chiller model to generate a GRU-SA chiller fault heat map;

[0053] Step S43: performing chiller fault prediction and diagnosis evaluation based on the GRU-SA chiller fault heat map, thereby completing chiller fault prediction and diagnosis based on the GRU-SA model.

[0054] The present invention generates a GRU-SA chiller fault cloud-based judgment model by deploying the GRU-SA chiller fault judgment model in the cloud. Cloud deployment provides the model with powerful computing resources and storage capabilities, capable of processing large amounts of real-time data and performing deep learning analysis. This process ensures that the model can quickly respond to operating condition data of different devices and different regions in a distributed environment, and make timely fault judgments on the system. Subsequently, GRU-SA chiller fault cloud-based judgment model is subjected to edge-cloud collaborative reasoning to generate GRU-SA chiller fault judgment data. Edge-cloud collaborative reasoning can allocate some computing tasks to edge devices for rapid processing, reducing the cloud computing burden and enabling faster response to real-time data on the device side. At the same time, the cloud performs in-depth analysis and decision-making on the data, ensuring the accuracy and reliability of the overall reasoning results of the system. This collaborative reasoning architecture greatly improves the real-time performance and response speed of the fault diagnosis system in complex environments, ensuring that the chiller can obtain accurate fault judgments in a timely manner. By plotting a contribution heat map of the GRU-SA chiller model's output data, a GRU-SA chiller fault heat map was generated. Heat map generation provides a visual means of fault analysis, helping engineers quickly locate problem areas by highlighting the fault contributions of key locations. Contribution heat maps not only reflect the health status of each chiller component but also intuitively display the contribution of different components to the overall system failure, making fault diagnosis more accurate and actionable. Finally, a chiller fault predictive diagnosis evaluation was conducted based on the GRU-SA chiller fault heat map, completing the chiller fault predictive diagnosis based on the GRU-SA model. By combining fault heat maps with predictive diagnostic evaluation, the system can predict chiller failures in advance, providing important decision-making support for maintenance personnel, allowing them to intervene early and reduce the probability of failures.

[0055] Preferably, step S43 includes the following steps:

[0056] Step S431: Obtain historical chiller operation data;

[0057] Step S432: extracting key faults from the GRU-SA chiller fault heat map to generate GRU-SA chiller fault key data;

[0058] Step S433: performing fault diagnosis comparison on the GRU-SA chiller fault key data and the historical chiller operation data to generate GRU-SA chiller fault diagnosis comparison data; performing fault judgment evaluation on the GRU-SA chiller fault diagnosis comparison data to generate GRU-SA chiller evaluation data;

[0059] Step S434: constructing a fault judgment report based on the GRU-SA chiller evaluation data, generating a GRU-SA chiller evaluation report, thereby completing the chiller fault prediction and diagnosis based on the GRU-SA model.

[0060] By acquiring historical chiller operating data, the present invention provides rich background information for subsequent fault analysis and assessment. Historical data not only reflects the operating status of the equipment in different time periods, but also provides long-term trends and changing patterns for the fault diagnosis model, providing benchmark data for the system to determine whether the current status is abnormal. Next, the GRU-SA chiller fault heat map is used to extract key faults, generating GRU-SA chiller fault key data. In this process, the fault heat map helps identify and demarcate the most critical fault areas in the system, focusing on extracting key data related to unit performance loss, abnormal fluctuations, system failures, etc. This step can effectively narrow the scope of analysis and ensure the focus and accuracy of subsequent fault diagnosis. The GRU-SA chiller fault key data is compared with historical chiller operating data for fault diagnosis, generating GRU-SA chiller fault diagnosis comparison data. By comparing historical data with current fault data, the system can identify similarities and differences between the current fault and historical anomalies, thereby determining whether it is a similar fault or a change in system status. This comparative analysis not only helps identify fault types but also provides the model with more real-world operating data, improving the accuracy and reliability of fault diagnosis. Subsequently, a fault diagnosis evaluation was conducted on the GRU-SA chiller fault diagnosis comparative data, generating the GRU-SA chiller evaluation data. Based on multi-dimensional comparison and analysis, this evaluation data provides detailed fault diagnosis results, enabling decision makers to respond quickly and effectively prevent the expansion and spread of equipment failures. Finally, a fault diagnosis report was constructed based on the GRU-SA chiller evaluation data, generating a GRU-SA chiller evaluation report. This report not only provides comprehensive fault diagnosis information but also includes detailed evaluation results, recommended maintenance plans, and equipment status predictions, providing a scientific basis and decision support for equipment management and maintenance. Through this process, the system effectively supports fault early warning, maintenance scheduling, and optimized control strategies, helping to improve chiller reliability, reduce unexpected failures, and extend equipment life. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A schematic flow chart of the steps of a chiller fault prediction and diagnosis method based on the GRU-SA model;

[0062] Figure 2 for Figure 1 Detailed implementation steps of step S2 in FIG.

[0063] Figure 3 for Figure 1 Detailed implementation steps of step S3 in FIG.

[0064] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0065] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative work are within the scope of protection of the present invention.

[0066] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0067] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0068] To achieve this, please refer to Figures 1 to 3 A chiller fault prediction and diagnosis method based on the GRU-SA model, the method comprising the following steps:

[0069] Step S1: obtaining a multimodal chiller dataset; performing data preprocessing on the multimodal chiller dataset to generate a multimodal chiller preprocessed dataset;

[0070] Step S2: Performing physical knowledge virtual feature enhancement on the multimodal chiller preprocessed dataset to generate a multimodal chiller virtual feature set; performing dynamic adversarial data sample construction on the multimodal chiller virtual feature set to generate multimodal chiller model input data;

[0071] Step S3: The multimodal chiller model input data is input to the preset GRU convolutional model, and the self-attention layer is used to calculate the feature correlation matrix to generate a GRU-SA chiller preliminary model; the GRU-SA chiller preliminary model is adaptively adjusted to generate a GRU-SA chiller fault judgment model;

[0072] Step S4: Deploy the GRU-SA chiller fault judgment model on the cloud, perform edge-cloud collaborative reasoning, and generate GRU-SA chiller fault judgment data; draw a contribution heat map of the GRU-SA chiller model output data, thereby completing the chiller fault prediction and diagnosis based on the GRU-SA model.

[0073] The present invention achieves the beneficial effect of ensuring data consistency and accuracy by acquiring and preprocessing a multimodal chiller dataset, enabling subsequent analysis to be based on high-quality data. Data preprocessing not only eliminates noise but also converts the data into a standardized format, providing reliable input for feature enhancement and analysis. Next, a physical knowledge-based virtual feature enhancement technique is employed to generate a multimodal chiller virtual feature set by integrating the physical characteristics of chiller operation. This process provides the model with richer feature information, enabling it to capture deeper operational patterns, particularly in complex environments, effectively extracting potential fault-related patterns. A dynamic adversarial data sample construction method is employed to expand and optimize the multimodal chiller virtual feature set. The generation of adversarial samples enables the model to better cope with various disturbances and improves its ability to identify abnormal conditions and potential faults. This approach enhances the model's robustness, ensuring its stability and accuracy in actual operation. A preliminary GRU-SA chiller model is generated by inputting the constructed input data into a pre-set GRU convolutional model and combining it with a self-attention (SA) mechanism to calculate the feature correlation matrix. The model can further optimize the feature weights through adaptive parameter adjustment, thereby forming a fault diagnosis model with higher accuracy. The trained GRU-SA chiller fault diagnosis model is deployed on the cloud and combined with edge computing for collaborative reasoning, which further improves the data processing efficiency and real-time performance. The collaborative reasoning between the cloud and the edge enables the model to achieve efficient data transmission and processing between different computing nodes, thereby quickly responding to the fault prediction needs in actual operation. In addition, by drawing a contribution heat map, the output results of the model are visualized, and the contribution of each feature to fault diagnosis is intuitively displayed, which helps decision makers to quickly locate the source of the problem. Therefore, the present invention solves the problems of insufficient accuracy and poor robustness of traditional chiller fault diagnosis models through technical means such as multimodal data processing, physical knowledge enhancement, dynamic adversarial data sample generation, GRU-SA model optimization, and cloud-edge collaborative reasoning, thereby improving the accuracy and real-time performance of fault prediction.

[0074] In the embodiment of the present invention, reference Figure 1 FIG. 1 is a flow chart of a method for predicting and diagnosing chiller faults based on a GRU-SA model according to the present invention. In this example, the method for predicting and diagnosing chiller faults based on a GRU-SA model includes the following steps:

[0075] Step S1: obtaining a multimodal chiller dataset; performing data preprocessing on the multimodal chiller dataset to generate a multimodal chiller preprocessed dataset;

[0076] In an embodiment of the present invention, a multimodal chiller dataset is acquired. This dataset contains data from multiple sensors or measurement devices, such as vibration signals, temperature signals, and pressure data. These data often have different acquisition frequencies, noise characteristics, and physical units, requiring effective processing to provide valid information for subsequent analysis. During the data preprocessing stage, multiple data cleaning and transformation methods are used to standardize the raw data to eliminate inconsistencies and ensure that the data is analyzed on a consistent basis. Specifically, for vibration signals, wavelet packet transform denoising can be used to remove high-frequency noise and low-frequency drift, extracting valuable vibration features. For temperature and pressure signals, time series drift correction and normalization can be performed to ensure consistency across time points and eliminate systematic errors caused by factors such as temperature fluctuations and equipment operation. Furthermore, data preprocessing includes missing value imputation, outlier detection and processing, and data smoothing. Missing value imputation methods can include mean interpolation, linear interpolation, or multiple interpolation, depending on the data distribution and missingness mechanism. Outlier detection can combine statistical methods and physical models to improve data reliability by identifying data points that deviate from the normal range. Data smoothing methods such as sliding window averaging can reduce random fluctuations in the data and improve signal stability.

[0077] Step S2: Performing physical knowledge virtual feature enhancement on the multimodal chiller preprocessed dataset to generate a multimodal chiller virtual feature set; performing dynamic adversarial data sample construction on the multimodal chiller virtual feature set to generate multimodal chiller model input data;

[0078] In an embodiment of the present invention, preprocessed data can be supplemented by introducing physical knowledge about the unit's operating principles, thermodynamic models, and system states, thereby generating a multimodal virtual feature set for the chiller. For example, for the unit's temperature, pressure, and vibration signals, physical formulas can be embedded into the data based on thermodynamic equilibrium, refrigeration cycle efficiency, and mechanical vibration models. Through mathematical derivation and simulation, these raw data can be transformed into features with higher recognition and explanatory power. This enhancement process not only improves the physical relevance of the data but also ensures the physical rationality of the model input features, enabling the model to understand and capture complex patterns in the data at a deeper level. The process of constructing dynamic adversarial data samples aims to improve the robustness and generalization ability of the model by performing adversarial data enhancement on the generated virtual feature set. Dynamic adversarial sample construction utilizes perturbation techniques, typically combined with gradient-generated adversarial examples (GANs) or optimization-based adversarial training methods, to perturb the raw data and generate challenging samples. These adversarial samples can explore within the data space, allowing the model to not only learn the characteristics of conventional data during training but also cope with noise, bias, and other anomalies.

[0079] Step S3: The multimodal chiller model input data is input to the preset GRU convolutional model, and the self-attention layer is used to calculate the feature correlation matrix to generate a GRU-SA chiller preliminary model; the GRU-SA chiller preliminary model is adaptively adjusted to generate a GRU-SA chiller fault judgment model;

[0080] In this embodiment of the present invention, the input data of a multimodal chiller model is fed into a preset GRU (Gated Recurrent Unit) convolutional model. The core of this process is the input of time series features into the GRU model. The GRU effectively captures long-term dependencies in time series data through its internal gating mechanisms (such as reset and update gates). Because chiller data exhibits significant temporal characteristics, such as the changing trends of temperature, pressure, and vibration signals, the GRU model, through its recursive structure, can better understand the dynamic characteristics and trends in the data, demonstrating its superiority in processing nonlinear and non-stationary time series data. The convolution operation further enhances the model's ability to express spatial features, effectively extracting local patterns and important features in the input data. This allows the model to not only capture temporal dependencies but also handle different types of data features. Subsequently, a self-attention mechanism is introduced to calculate the feature correlation matrix. The self-attention mechanism performs weighted aggregation on the relationships between different time steps of the input data, calculating the importance of each feature within the entire sequence, thereby generating a feature correlation matrix. This mechanism helps strengthen the model's focus on key features and improves its understanding of the relationships between features of different dimensions, further enhancing prediction accuracy. After generating the preliminary model, adaptive parameter adjustment of the GRU-SA chiller preliminary model is another important step. Adaptive parameter adjustment dynamically adjusts the model's weights and hyperparameters using optimization algorithms (such as gradient descent and the Adam optimizer) to minimize the loss function and improve the model's generalization. This process automatically adjusts the model's learning rate, network structure, and other hyperparameters based on the data distribution and task requirements, thereby continuously optimizing model performance during training, avoiding overfitting, and improving its predictive ability on unseen data. After adaptive adjustment, the resulting GRU-SA chiller fault diagnosis model can more accurately identify chiller fault modes in practical applications, providing efficient fault prediction and diagnosis capabilities.

[0081] Step S4: Deploy the GRU-SA chiller fault judgment model on the cloud, perform edge-cloud collaborative reasoning, and generate GRU-SA chiller fault judgment data; draw a contribution heat map of the GRU-SA chiller model output data, thereby completing the chiller fault prediction and diagnosis based on the GRU-SA model.

[0082] In this embodiment of the present invention, the GRU-SA chiller fault diagnosis model is deployed in the cloud. This process uploads the trained GRU-SA model to the cloud platform for real-time computation and inference. Cloud deployment fully leverages the advantages of cloud computing resources, such as powerful computing and storage capabilities, to enable large-scale data processing and model inference. Once deployed, the GRU-SA model can receive data input from the chiller in real time and perform fault diagnosis in the cloud. Furthermore, through the edge-cloud collaborative inference mechanism, some computational tasks are assigned to edge devices for local inference, reducing data transmission latency and computing pressure on the cloud, thereby improving system response speed and efficiency. Edge devices typically have lower computing resources and are primarily used for preliminary data preprocessing and simple inference, while the cloud is responsible for complex computational tasks and large-scale model inference. This collaborative work enables efficient and low-latency fault prediction. The generated GRU-SA chiller fault diagnosis data contains the types and severity of faults that occurred during chiller operation. Based on this output data, the next step is to create a contribution heat map of the GRU-SA chiller model output data. The contribution heatmap analyzes the weights of each feature in the model's output data, displaying the degree of influence of each feature on the final fault diagnosis. The heatmap visually demonstrates the contribution of different features (such as temperature, pressure, and vibration) to the fault prediction results, helping engineers understand the root cause of the fault and providing guidance for subsequent fault prevention and maintenance. In this way, the heatmap not only improves the interpretability of the model but also provides data support for chiller fault diagnosis and maintenance decisions. Ultimately, by combining the GRU-SA chiller fault prediction and diagnosis system with the heatmap, a complete chiller fault prediction and diagnosis system based on the GRU-SA model has been established.

[0083] Preferably, step S1 includes the following steps:

[0084] Step S11: Acquire a multimodal chiller dataset, wherein the multimodal chiller dataset includes a chiller vibration signal, a chiller temperature signal, and a chiller pressure data;

[0085] Step S12: performing wavelet packet transform denoising on the unit vibration signal to generate unit wavelet transform denoised data; performing data standardization processing on the unit wavelet transform denoised data to generate unit vibration standardised data;

[0086] Step S13: performing time series drift processing on the unit temperature signal and the unit pressure data to generate unit temperature normalized data and unit pressure normalized data;

[0087] Step S14: synchronize the unit vibration normalized data, the unit temperature normalized data, and the unit pressure normalized data in a time window to generate a multi-modal chiller preprocessing data set.

[0088] In this embodiment of the present invention, a multimodal chiller dataset, comprising vibration signals, temperature signals, and pressure data, is acquired to provide foundational data for subsequent analysis and modeling. A multimodal dataset encompasses data from different physical quantities, such as vibration, temperature, and pressure. These data are complementary and diverse in reflecting the chiller's operating status, revealing more system characteristics. Wavelet packet transform is then used to denoise the chiller vibration signal. Wavelet packet transform is a time-frequency analysis method that effectively decomposes signals and extracts features from different frequency bands, thereby removing noise and improving signal quality. Wavelet transform denoising technology selects appropriate wavelet basis functions to perform multi-level decomposition and reconstruction of the signal, filtering out high-frequency noise while preserving the signal's key features. The wavelet packet transformed signal is then normalized to ensure uniform dimensionality across different feature data and eliminate dimensional differences between variables, ensuring data comparability and consistency in subsequent analysis. This normalization process involves linearly transforming the data to a uniform standard range (e.g., a mean of 0 and a variance of 1), thereby enhancing model training effectiveness. The unit temperature and pressure data require time series drift correction. Time series drift refers to the trend or drift in signal changes over time, which can affect signal analysis and model prediction. By removing or adjusting this drift trend, the signal's periodic changes and transient fluctuations can be better captured, thereby improving the accuracy of subsequent processing. The drift-adjusted data is then standardized to ensure that the temperature and pressure signals have a uniform scale and standard, facilitating subsequent multimodal data fusion and model training. The standardized data of the unit vibration, temperature, and pressure signals requires time window synchronization. Time window synchronization ensures that multimodal signals are aligned at the same time point, facilitating multivariate analysis and modeling. During chiller operation, the changes in vibration, temperature, and pressure data are interrelated. However, these signals typically have different sampling frequencies and time series. Therefore, time window synchronization technology is required to ensure consistent alignment of multimodal data at each time point, providing a complete, multi-dimensional system status dataset for subsequent analysis.

[0089] Preferably, step S13 includes the following steps:

[0090] Step S131: Performing a two-dimensional linear projection on the unit temperature signal and the unit pressure data to generate a two-dimensional projection of the unit temperature and pressure; performing linear interpolation outlier correction on the two-dimensional projection of the unit temperature and pressure to generate unit temperature and pressure outlier processed data;

[0091] Step S132: performing white correlation trend analysis on the unit temperature-pressure abnormal value processing data to generate unit temperature-pressure white correlation trend data;

[0092] Step S133: performing time series drift processing on the unit temperature-pressure white correlation trend data to generate unit temperature-pressure time series drift data; performing second-order difference normalization processing on the unit temperature-pressure time series drift data to generate unit temperature normalized data and unit pressure normalized data.

[0093] In this embodiment of the present invention, a two-dimensional linear projection is performed on the unit temperature signal and unit pressure data. This is the process of mapping high-dimensional data onto a two-dimensional plane. This linear projection compresses the original multidimensional data into low-dimensional data that is easier to analyze, while preserving the key information of the original data. The core technology of this step is matrix operations in linear algebra. Using weight coefficients, the changing trends and relationships of various variables can be visualized and presented in two-dimensional space, thereby reducing data complexity and facilitating subsequent analysis. Next, during the linear interpolation outlier correction process on the two-dimensional unit temperature and pressure projections, linear interpolation techniques are used to fill or correct outliers in the data. Outliers are caused by noise, sensor failure, or acquisition errors, and can negatively impact subsequent data analysis and modeling. Linear interpolation infers the values ​​of missing data points using the linear relationships between known data points. This ensures data integrity while minimizing distortion of the original data trend. This process ensures data continuity and consistency. White correlation trend analysis is performed on the corrected unit temperature and pressure data. White correlation analysis aims to identify potential noise or trend components in the data. Time series data often contains white noise (random fluctuations in the data) and systematic trend components. White noise analysis can effectively separate these trend components from the data, thereby enhancing the extraction of valid information. This process helps more accurately describe the intrinsic correlation between unit temperature and pressure data, removes irrelevant noise, and improves analysis quality. Time series drift processing is performed on the unit temperature-pressure white noise trend data. Time series drift refers to systematic changes in data over time, which often affect data stability and forecast accuracy. Drift processing can remove these time-dependent trends, making the data more stable and easier to analyze. The temperature time series drift data is then subjected to second-order difference normalization. This process uses second-order differencing to remove trend components from the data and emphasize the fluctuations, making the data more stable and suitable for subsequent statistical modeling or machine learning analysis. Second-order difference normalization converts the data rate of change into unit standard deviation, ensuring that temperature and pressure data are compared and integrated in the same dimension, improving data consistency and operability.

[0094] Preferably, the physical knowledge virtual feature enhancement of the multimodal chiller preprocessing data set comprises the following steps:

[0095] Conduct correlation analysis of the refrigeration cycle physical process based on the unit vibration standardization data to generate the unit cold cycle vibration analysis data; conduct indirect refrigeration cycle efficiency impact analysis on the unit cold cycle vibration analysis data to generate the unit vibration theoretical energy efficiency ratio;

[0096] Based on the unit temperature standardization data, the temperature fluctuation pattern is used to identify potential equipment failures and generate the unit temperature fluctuation pattern data; the unit temperature fluctuation pattern data is subjected to least squares regression physical constraint dimensionality reduction to generate the unit temperature subcooling deviation;

[0097] Perform thermodynamic state analysis on the standardized unit pressure data to generate the unit pressure thermodynamic state curve; perform outlier peak analysis on the unit pressure thermodynamic state curve to generate unit pressure thermodynamic abnormal peak data; perform temperature-pressure correlation analysis on the unit pressure thermodynamic abnormal peak data and the unit temperature subcooling deviation to generate the unit temperature-pressure subcooling deviation.

[0098] One-hot encoding is performed based on the temperature-pressure subcooling deviation of the unit and the theoretical energy efficiency ratio of the unit vibration, and physical prior processing is performed to generate a multi-modal chiller virtual feature set.

[0099] In an embodiment of the present invention, by performing correlation analysis on the refrigeration cycle physical processes of standardized unit vibration data and applying the relationship between physical models and vibration data, the unit's refrigeration cycle characteristics can be extracted from the vibration data. This correlation analysis typically relies on physical modeling and signal processing techniques. By establishing a mathematical relationship between the refrigeration cycle and vibration, it reveals the impact of the unit's vibration characteristics on refrigeration efficiency, thereby generating unit refrigeration cycle vibration analysis data. This process achieves a quantitative correlation between vibration data and refrigeration performance through mathematical modeling. An indirect refrigeration cycle efficiency impact analysis is performed to evaluate the impact of vibration on chiller efficiency by calculating the theoretical energy efficiency ratio (EER) based on the vibration data. This analysis relies on thermodynamic and kinetic models, combining vibration characteristics to derive the unit's theoretical EER, thereby determining the unit's efficiency performance. This technology can provide effective predictions of the unit's operating status. For standardized unit temperature data, temperature fluctuation pattern recognition is first performed. Using a machine learning algorithm, fluctuation patterns in the temperature data are identified. These patterns can reveal potential equipment failures, such as temperature instability, which can signal a cooling system failure. Temperature fluctuation pattern recognition is performed through time series analysis or cluster analysis, generating unit temperature fluctuation pattern data. Least squares regression and physical constraint dimensionality reduction are used to reduce the dimensionality of unit temperature fluctuation data, reducing the complexity of the feature space and making the data more concise and easier to analyze. This dimensionality reduction method effectively extracts the unit's temperature subcooling deviation by retaining key physical features and removing redundant data. This technical approach, based on a regression model, maps complex temperature data into a low-dimensional space, facilitating subsequent evaluation and analysis. When processing standardized unit pressure data, a unit pressure thermodynamic state curve is generated through thermodynamic state analysis. Thermodynamic state analysis typically uses state parameters such as pressure, temperature, and volume to generate equipment operating state curves based on thermodynamic equations, reflecting the unit's operating efficiency and performance. When pressure data is abnormal, peak analysis can identify potential anomalies, such as signs of system overload or failure. Analysis of abnormal peak data allows for timely detection and correction of anomalies in pressure data. Temperature-pressure correlation analysis of abnormal unit pressure thermodynamic peak data and temperature subcooling deviation reveals the mutual influence between temperature and pressure, thereby analyzing their combined impact on the unit's operating status. This analysis helps reveal the overall performance of the unit and improves the accuracy of fault prediction. Finally, by performing one-hot encoding on the temperature-pressure subcooling deviation and the vibration energy efficiency ratio, combined with physical priors, a multimodal chiller virtual feature set was generated. One-hot encoding, a common feature engineering technique, converts discrete features of different categories into binary vectors. Here, it is used to standardize multiple features, making them suitable for machine learning models.

[0100] As an example of the present invention, refer to Figure 2 As shown, in this example, step S2 includes:

[0101] Step S21: performing physical knowledge virtual feature enhancement on the multimodal chiller preprocessing data set to generate a multimodal chiller virtual feature set;

[0102] Step S22: performing random noise perturbation on the multimodal chiller virtual feature set to generate multimodal random noise perturbation data; performing gradient descent on the multimodal random noise perturbation data to generate multimodal gradient descent data;

[0103] Step S23: scramble the multimodal gradient descent data to generate multimodal gradient descent scrambled data; perform time series perturbation analysis on the multimodal gradient descent scrambled data, thereby completing the construction of dynamic adversarial data samples and generating multimodal chiller model input data.

[0104] In this embodiment of the present invention, a multimodal chiller preprocessed dataset is enhanced with virtual features based on physical knowledge. This virtual feature enhancement utilizes known physical laws and system behavior characteristics based on the chiller's physical model and expert knowledge to generate virtual features that complement the original dataset. These virtual features enhance the dataset's expressiveness by mapping the chiller's operating principles, enabling it to cover a wider range of operating scenarios and system states, thus facilitating accurate prediction and diagnosis in practical applications. The generated virtual feature set is perturbed with random noise to enhance data diversity. Noise perturbation, a common enhancement method, blurs the precise features of the original data by adding a certain range of random noise, allowing the model to adapt to a wider variety of input scenarios during training. This method helps improve the model's robustness to data fluctuations and external interference, thereby avoiding overfitting and enhancing prediction performance on unknown data. Subsequently, the multimodal random noise-perturbed data is further optimized using the gradient descent algorithm. Gradient descent, a widely used optimization method, effectively minimizes the error function, helping the model better learn the underlying laws in the data. In this way, the perturbed data is guided toward a state that is more consistent with reality, further enhancing the dataset's practicality. The multimodal gradient descent data is further scrambled, subjecting it to a certain level of perturbation and reconstruction before being input into the model. This process, by nonlinearly scrambling the data, enhances the complexity and nonlinear characteristics of the dataset, enabling the model to better capture complex patterns. Furthermore, the data is subjected to time series perturbation analysis. This technique primarily addresses the dynamic patterns of time series data, analyzing and capturing trends in data over time, thereby improving the temporal dependence and temporal nature of the dataset.

[0105] Preferably, step S3 includes the following steps:

[0106] Step S31: Input the multimodal chiller model input data into the preset GRU convolutional model, which contains a GRU layer with 64 hidden units, and calculates the feature correlation matrix of the self-attention layer, outputs the hidden state of the GRU layer and the self-attention layer, and constructs a preliminary model of the GRU-SA chiller. The calculation formula for the feature correlation matrix includes:

[0107] ;

[0108] in, , , are query, key, and value matrices respectively, is the feature dimension;

[0109] Step S32: performing cross entropy loss function expansion on the GRU-SA chiller preliminary model to generate a GRU-SA energy regularization loss function;

[0110] Step S33: Adaptively adjust the parameters of the GRU-SA chiller preliminary model and the GRU-SA energy regularization loss function to generate a GRU-SA chiller fault judgment model, wherein the weight of the GRU-SA model is adjusted by the Adam optimizer, and the learning rate is set to 0.001.

[0111] In the embodiment of the present invention, at the data level, the technical means of step S3 realizes time series feature extraction and model optimization through multi-stage data processing: First, the multimodal chiller model input data (including multi-source time series signals such as vibration, temperature, and pressure) is input as a three-dimensional tensor (dimension: time step x feature dimension x number of modes). The tensor input is a preset GRU convolutional model, the core of which is a GRU layer containing 64 hidden units, and the hidden state is calculated through a gating mechanism (reset gate and update gate) in To update the gate, At the same time, the self-attention layer maps the hidden state sequence output by the GRU into the query matrix Q, the key matrix K, and the value matrix S, and calculates the feature correlation matrix through the scaled dot product attention mechanism:

[0112]

[0113] in The softmax function normalizes the attention weights for the feature dimension (typical value is set to 64) to highlight the feature contributions of the key time steps. The fusion of the GRU layer and the self-attention layer is achieved through splicing or weighted summation to generate a hidden state output that has both local temporal dependencies and global feature associations, and to construct a preliminary GRU-SA model. In step S32, based on the classification output of the preliminary model, the standard cross entropy loss function is expanded to an energy regularization loss function. This regularization term suppresses the model's tendency to overfit to noisy data and improves generalization ability by constraining the parameter norm. Step S33 adopts an adaptive parameter adjustment mechanism: the GRU-SA model weights are iteratively updated through the Adam optimizer (combined with momentum and adaptive learning rate), and its parameter update rule is:

[0114]

[0115] in is a fixed learning rate, and are the bias-corrected estimates of the first and second moments of the gradient, is a very small constant (usually The GRU convolution kernel weights, attention projection matrix, and fully connected layer parameters are adjusted to ultimately generate a GRU-SA fault diagnosis model that accurately captures fault characteristics. The entire process completes a closed loop of parameter optimization, from raw multimodal data to high-dimensional feature representation, and then to regularized loss-guided parameter optimization.

[0116] Preferably, the self-attention layer performs feature correlation matrix calculation including the following steps:

[0117] Obtain chiller refrigeration drawings and real-time refrigeration operating parameters;

[0118] Perform 3D modeling on the chiller refrigeration drawings to generate the chiller refrigeration cycle physical structure; construct the refrigeration system topology nodes on the chiller refrigeration cycle physical structure to generate the chiller topology nodes;

[0119] Perform real-time working condition encoding based on the chiller topology nodes and real-time refrigeration working condition parameters to generate dynamic unit working condition drive encoding data; perform sparse matrix correlation matrix synthesis on the dynamic unit working condition drive encoding data to generate a chiller sparse correlation matrix;

[0120] The sparse correlation matrix of the chiller is calculated using the physical constraint self-attention process and irrelevant noise masking is performed to complete the feature correlation matrix calculation of the attention layer.

[0121] In this embodiment of the present invention, by obtaining chiller refrigeration drawings and real-time operating parameters, combined with engineering design drawings and real-time operating data, basic data support is provided for subsequent modeling and analysis. Based on this, 3D modeling technology is used to process the chiller refrigeration drawings and generate the physical structure of the chiller's refrigeration cycle. The core of this step is to convert the 2D design drawings into a 3D spatial model, facilitating detailed physical process analysis and topological structure construction. The chiller's refrigeration system topology nodes are established. This process models the interconnections between the chiller's components (such as the cooling tower, compressor, and evaporator) to form a complete topological network. This topological node construction provides a systematic foundation for subsequent real-time operating condition encoding. Real-time operating condition encoding is then generated by combining real-time refrigeration condition parameters with the topological nodes. This encoding process converts the chiller's operating status into digital signals suitable for processing by machine learning models, reflecting the real-time nature of the dynamic interaction between the machine and the system. A sparse matrix correlation matrix is ​​generated for the chiller by synthesizing the generated dynamic unit operating condition drive code data. Sparse matrices are used to describe the non-zero interactions between nodes in a system. Sparse matrix synthesis effectively reduces computational complexity while preserving useful relational information within the system, laying the foundation for subsequent model optimization. A physically constrained self-attention process is performed on the chiller's sparse association matrix. The self-attention mechanism calculates the relative importance weights between features in the input data, enabling the model to effectively capture long-range dependencies and the dynamic interactions between different features in the system. The introduction of physical constraints at this stage ensures that the model adheres to physical laws when calculating feature associations, preventing irrational predictions or judgments in real-world applications.

[0122] Preferably, step S33 includes the following steps:

[0123] Step S331: performing time dimension association on the GRU-SA chiller preliminary model and the GRU-SA energy regularization loss function to generate GRU-SA chiller time dimension data; performing feature dimension association on the GRU-SA chiller preliminary model and the GRU-SA energy regularization loss function to generate GRU-SA chiller feature dimension data;

[0124] Step S332: performing multi-level association aggregation on the GRU-SA chiller time dimension data and the GRU-SA chiller feature dimension data to generate GRU-SA chiller time-feature dimension data;

[0125] Step S333: inject the fault type prior into the GRU-SA chiller time-feature dimension data to generate chiller fault type data; perform temporal fault causality protection on the chiller fault type data, and perform adaptive parameter adjustment on the refrigerant evaporator and compressor suction pipe to generate a GRU-SA chiller fault judgment model.

[0126] In this embodiment of the present invention, the model's timeliness is modeled by integrating time series data. Time features are extracted based on the dynamic changes in the time series, providing data support with time-series relevance for subsequent analysis. Furthermore, feature dimensions are correlated, and by aggregating the model's feature data, the inherent connections between multidimensional features are further explored, providing more comprehensive data support for modeling complex system behavior. The combination of time and feature dimensions enables the model to capture behavioral patterns at different time points and across different feature dimensions during chiller operation, thereby improving prediction accuracy. By performing multi-level correlation and aggregation of time and feature dimension data, time-feature dimension data for the GRU-SA chiller is generated. This process hierarchically processes data, deeply integrating the relationships between different dimensions through a hierarchical structure, enabling the model to more comprehensively understand the dynamic changes within the system. This aggregation process not only optimizes the efficiency of information transfer but also enhances the model's expressiveness, allowing it to simultaneously consider time changes and feature characteristics, improving the model's ability to fit complex system behavior. By injecting prior information about fault types, chiller fault type data is generated. This technical approach utilizes prior knowledge to inject different types of fault information into the model training process, which helps the model better identify and distinguish different fault types during the learning process and provide more accurate guidance for subsequent fault prediction.

[0127] As an example of the present invention, refer to Figure 3 As shown, in this example, step S4 includes:

[0128] Step S41: Deploy the GRU-SA chiller fault judgment model on the cloud to generate a GRU-SA chiller fault cloud judgment model; perform edge-cloud collaborative reasoning on the GRU-SA chiller fault cloud judgment model to generate GRU-SA chiller fault judgment data;

[0129] Step S42: Draw a contribution heat map for the output data of the GRU-SA chiller model to generate a GRU-SA chiller fault heat map;

[0130] Step S43: performing chiller fault prediction and diagnosis evaluation based on the GRU-SA chiller fault heat map, thereby completing chiller fault prediction and diagnosis based on the GRU-SA model.

[0131] In this embodiment of the present invention, the GRU-SA chiller fault diagnosis model is deployed to the cloud and edge-cloud collaborative reasoning is performed. Cloud deployment primarily migrates complex computing tasks and model processing to the cloud platform, leveraging cloud computing's high-performance computing resources for efficient data processing and analysis. Cloud deployment offers the advantage of dynamically scalable computing power to meet large-scale data processing requirements. Edge-cloud collaborative reasoning, through division of labor between edge devices and the cloud, effectively distributes computing tasks based on real-time requirements and data size. Preliminary data processing and reasoning are performed on the edge, centralizing complex computing tasks and large-scale data processing in the cloud, reducing latency and optimizing resource utilization. A heat map is also created for the GRU-SA chiller model output data. This heat map technique visually displays each indicator output by the model through color coding, helping to identify the chiller's fault area and severity. Contribution analysis of the model output data effectively determines the impact of each feature on fault diagnosis. Heat map visualization then visually displays the location and cause of the fault, providing maintenance personnel with accurate fault location and priority decision-making. Fault prediction and diagnosis assessment evaluates the current chiller failure risk and warning level through analysis of thermal maps, combined with historical fault data and changes in the operating environment.

[0132] Preferably, step S43 includes the following steps:

[0133] Step S431: Obtain historical chiller operation data;

[0134] Step S432: extracting key faults from the GRU-SA chiller fault heat map to generate GRU-SA chiller fault key data;

[0135] Step S433: performing fault diagnosis comparison on the GRU-SA chiller fault key data and the historical chiller operation data to generate GRU-SA chiller fault diagnosis comparison data; performing fault judgment evaluation on the GRU-SA chiller fault diagnosis comparison data to generate GRU-SA chiller evaluation data;

[0136] Step S434: constructing a fault judgment report based on the GRU-SA chiller evaluation data, generating a GRU-SA chiller evaluation report, thereby completing the chiller fault prediction and diagnosis based on the GRU-SA model.

[0137] In this embodiment of the present invention, historical chiller operating data is acquired. This process involves collecting various types of chiller equipment operating data, such as temperature, pressure, vibration, and other multi-dimensional physical signals. This data provides a rich foundation for subsequent fault prediction and diagnosis. Step S432 generates GRU-SA chiller fault key data by extracting key faults from the GRU-SA chiller fault heat map. This process utilizes a self-attention mechanism and heat map analysis techniques to filter out high-impact fault modes or abnormal behaviors from large-scale operating data. The key technologies in this step are heat map generation and automatic identification of focus points, which highlight the areas most likely to fail and provide target areas for further analysis. The GRU-SA chiller fault key data is then compared with historical chiller operating data for fault diagnosis to generate fault diagnosis comparison data. The core technology of this process is data comparison and analysis. Machine learning models are used to compare the extracted fault data with historical data. By identifying similarities and differences, the device is determined to be in a fault state, and the specific nature and location of the fault can be inferred. In addition, by comparing data, it can also provide a data basis for the next step of fault judgment and evaluation, and generate GRU-SA chiller evaluation data to quantitatively evaluate the severity and risk of the fault. The technical means used in this step include weight adjustment of model output, deep fusion analysis of historical data and existing data, etc. By constructing a fault judgment report based on the evaluation data, a GRU-SA chiller evaluation report is generated to complete fault prediction and diagnosis. The report generation process combines the evaluation results with the actual operation scenario to form a diagnostic report with operational and early warning functions. This process combines data visualization and report generation technology. By generating specific evaluation reports, it not only helps operation and maintenance personnel identify the type of fault, but also provides targeted maintenance suggestions, thereby improving the intelligence level and response speed of equipment management.

[0138] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0139] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A chiller fault prediction and diagnosis method based on the GRU-SA model, characterized in that: The following steps are involved: Step S1: Obtain a multimodal chiller dataset; Perform data preprocessing on the multimodal chiller dataset to generate a multimodal chiller preprocessed dataset; Step S2: Performing physical knowledge virtual feature enhancement on the multimodal chiller preprocessing dataset to generate a multimodal chiller virtual feature set; Dynamically construct adversarial data samples for the multimodal chiller virtual feature set to generate multimodal chiller model input data; Step S3: The multimodal chiller model input data is input to the preset GRU convolutional model, and the self-attention layer is used to calculate the feature correlation matrix to generate a GRU-SA chiller preliminary model; the GRU-SA chiller preliminary model is adaptively adjusted to generate a GRU-SA chiller fault judgment model; Step S4: Deploy the GRU-SA chiller fault judgment model on the cloud, perform edge-cloud collaborative reasoning, and generate GRU-SA chiller fault judgment data; draw a contribution heat map for the GRU-SA chiller model output data.

2. The chiller fault prediction and diagnosis method based on the GRU-SA model according to claim 1 is characterized in that: Step S1 includes the following steps: Step S11: Acquire a multimodal chiller dataset, wherein the multimodal chiller dataset includes a chiller vibration signal, a chiller temperature signal, and a chiller pressure data; Step S12: performing wavelet packet transform denoising on the unit vibration signal to generate unit wavelet transform denoised data; performing data standardization processing on the unit wavelet transform denoised data to generate unit vibration standardised data; Step S13: performing time series drift processing on the unit temperature signal and the unit pressure data to generate unit temperature normalized data and unit pressure normalized data; Step S14: synchronize the unit vibration normalized data, the unit temperature normalized data, and the unit pressure normalized data in a time window to generate a multi-modal chiller preprocessing data set.

3. The chiller fault prediction and diagnosis method based on the GRU-SA model according to claim 2 is characterized in that: Step S13 includes the following steps: Step S131: Performing a two-dimensional linear projection on the unit temperature signal and the unit pressure data to generate a two-dimensional projection of the unit temperature and pressure; performing linear interpolation outlier correction on the two-dimensional projection of the unit temperature and pressure to generate unit temperature and pressure outlier processed data; Step S132: performing white correlation trend analysis on the unit temperature-pressure abnormal value processing data to generate unit temperature-pressure white correlation trend data; Step S133: performing time series drift processing on the unit temperature-pressure white correlation trend data to generate unit temperature-pressure time series drift data; performing second-order difference normalization processing on the unit temperature-pressure time series drift data to generate unit temperature normalized data and unit pressure normalized data.

4. The chiller fault prediction and diagnosis method based on the GRU-SA model according to claim 1 is characterized in that: The method of enhancing the multimodal chiller preprocessing dataset with virtual features based on physical knowledge includes the following steps: Conduct correlation analysis of the refrigeration cycle physical process based on the unit vibration standardization data to generate the unit cold cycle vibration analysis data; conduct indirect refrigeration cycle efficiency impact analysis on the unit cold cycle vibration analysis data to generate the unit vibration theoretical energy efficiency ratio; Based on the unit temperature standardization data, the temperature fluctuation pattern is used to identify potential equipment failures and generate the unit temperature fluctuation pattern data; the unit temperature fluctuation pattern data is subjected to least squares regression physical constraint dimensionality reduction to generate the unit temperature subcooling deviation; Perform thermodynamic state analysis based on the standardized unit pressure data to generate the unit pressure thermodynamic state curve; perform outlier peak analysis on the unit pressure thermodynamic state curve to generate unit pressure thermodynamic abnormal peak data; perform temperature-pressure correlation analysis on the unit pressure thermodynamic abnormal peak data and the unit temperature subcooling deviation to generate the unit temperature-pressure subcooling deviation; One-hot encoding is performed based on the temperature-pressure subcooling deviation of the unit and the theoretical energy efficiency ratio of the unit vibration, and physical prior processing is performed to generate a multi-modal chiller virtual feature set.

5. The chiller fault prediction and diagnosis method based on the GRU-SA model according to claim 1 is characterized in that: Step S2 includes the following steps: Step S21: performing physical knowledge virtual feature enhancement on the multimodal chiller preprocessing data set to generate a multimodal chiller virtual feature set; Step S22: performing random noise perturbation on the multimodal chiller virtual feature set to generate multimodal random noise perturbation data; performing gradient descent on the multimodal random noise perturbation data to generate multimodal gradient descent data; Step S23: scramble the multimodal gradient descent data to generate multimodal gradient descent scrambled data; perform time series perturbation analysis on the multimodal gradient descent scrambled data, thereby completing the construction of dynamic adversarial data samples and generating multimodal chiller model input data.

6. The chiller fault prediction and diagnosis method based on the GRU-SA model according to claim 1 is characterized in that: Step S3 includes the following steps: Step S31: Input the multimodal chiller model input data into the preset GRU convolutional model, which contains a GRU layer with 64 hidden units, and calculates the feature correlation matrix of the self-attention layer, outputs the hidden state of the GRU layer and the self-attention layer, and constructs a preliminary model of the GRU-SA chiller. The calculation formula for the feature correlation matrix includes: ； in, , , are query, key, and value matrices respectively, is the feature dimension; Step S32: performing cross entropy loss function expansion on the GRU-SA chiller preliminary model to generate a GRU-SA energy regularization loss function; Step S33: Adaptively adjust the parameters of the GRU-SA chiller preliminary model and the GRU-SA energy regularization loss function to generate a GRU-SA chiller fault judgment model, wherein the weight of the GRU-SA model is adjusted by the Adam optimizer, and the learning rate is set to 0.

001.

7. The chiller fault prediction and diagnosis method based on the GRU-SA model according to claim 1 is characterized in that: The self-attention layer performs feature correlation matrix calculation including the following steps: Obtain chiller refrigeration drawings and real-time refrigeration operating parameters; Perform 3D modeling on the chiller refrigeration drawings to generate the chiller refrigeration cycle physical structure; construct the refrigeration system topology nodes on the chiller refrigeration cycle physical structure to generate the chiller topology nodes; Perform real-time working condition encoding based on the chiller topology nodes and real-time refrigeration working condition parameters to generate dynamic unit working condition drive encoding data; perform sparse matrix correlation matrix synthesis on the dynamic unit working condition drive encoding data to generate a chiller sparse correlation matrix; The sparse correlation matrix of the chiller is calculated using the physical constraint self-attention process and irrelevant noise masking is performed to complete the feature correlation matrix calculation of the attention layer.

8. The chiller fault prediction and diagnosis method based on the GRU-SA model according to claim 6 is characterized in that: Step S33 The following steps are involved: Step S331: performing time dimension association on the GRU-SA chiller preliminary model and the GRU-SA energy regularization loss function to generate GRU-SA chiller time dimension data; performing feature dimension association on the GRU-SA chiller preliminary model and the GRU-SA energy regularization loss function to generate GRU-SA chiller feature dimension data; Step S332: performing multi-level association aggregation on the GRU-SA chiller time dimension data and the GRU-SA chiller feature dimension data to generate GRU-SA chiller time-feature dimension data; Step S333: inject the fault type prior into the GRU-SA chiller time-feature dimension data to generate chiller fault type data; perform temporal fault causality protection on the chiller fault type data, and perform adaptive parameter adjustment on the refrigerant evaporator and compressor suction pipe to generate a GRU-SA chiller fault judgment model.

9. The chiller fault prediction and diagnosis method based on the GRU-SA model according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: Deploy the GRU-SA chiller fault judgment model on the cloud to generate a GRU-SA chiller fault cloud judgment model; perform edge-cloud collaborative reasoning on the GRU-SA chiller fault cloud judgment model to generate GRU-SA chiller fault judgment data; Step S42: Draw a contribution heat map for the output data of the GRU-SA chiller model to generate a GRU-SA chiller fault heat map; Step S43: performing a chiller fault prediction diagnosis assessment based on the GRU-SA chiller fault heat map, thereby generating a GRU-SA chiller assessment report.

10. The chiller fault prediction and diagnosis method based on the GRU-SA model according to claim 9, characterized in that: Step S43 includes the following steps: Step S431: Obtain historical chiller operation data; Step S432: extracting key faults from the GRU-SA chiller fault heat map to generate GRU-SA chiller fault key data; Step S433: performing fault diagnosis comparison on the GRU-SA chiller fault key data and the historical chiller operation data to generate GRU-SA chiller fault diagnosis comparison data; performing fault judgment evaluation on the GRU-SA chiller fault diagnosis comparison data to generate GRU-SA chiller evaluation data; Step S434: constructing a fault judgment report based on the GRU-SA chiller evaluation data to generate a GRU-SA chiller evaluation report.

Images