Water chilling unit diagnosis method and equipment based on Transform and medium

Abstract

The invention discloses a water chilling unit diagnosis method and device based on Transform and a medium, and relates to the technical field of water chilling units. The method comprises the following steps: acquiring original time series data of the water chilling unit through a multi-dimensional sensor, wherein the original time series data comprises temperature data, pressure data, flow data, current data and liquid level data; enhancing and fusing the time sequence data to obtain enhanced time sequence data; carrying out modeling on the enhanced time sequence data by adopting an improved Transform architecture to obtain an abnormality diagnosis model of the water chilling unit; and automatic hyper-parameters are introduced into the anomaly diagnosis model to calculate an optimal parameter combination of the anomaly diagnosis model, and anomaly diagnosis is carried out on the operation data of the water chilling unit. By means of the method, early recognition and early warning of the abnormal state are achieved, the equipment fault risk is reduced, the detection precision is remarkably improved, false alarm and missing alarm are reduced, and a brand new technical scheme is provided for intelligent operation and maintenance of the water chilling unit.

Description

Technical Field

[0001] The present application relates to the technical field of chillers, and in particular to a Transformer-based chiller diagnosis method, device and medium. Background Art

[0002] In the traditional chiller operation and maintenance practice, equipment status monitoring and fault diagnosis mainly rely on manual experience judgment, which makes it difficult to make quick and accurate judgments on complex abnormal patterns and deep-seated system problems, and also prolongs the downtime of equipment. In recent years, with the development of deep learning technology, some anomaly detection methods based on deep neural networks have been proposed and applied to time series data analysis, such as autoencoders, convolutional neural networks (CNN) and long short-term memory networks (LSTM). Although these methods have improved the automation level of anomaly detection to a certain extent, there are still the following key problems:

[0003] First, the data generated during the operation of the chiller is high-dimensional and strongly coupled. The operation of the equipment involves multiple physical quantities such as temperature, pressure, and flow. There are complex nonlinear relationships between these parameters. Traditional deep learning methods often find it difficult to effectively model such complex dependencies, especially when processing long sequence data. Existing models are prone to the problem of long-range dependency information loss, which affects the accuracy of anomaly detection.

[0004] Secondly, in the actual operating environment, abnormal data samples are extremely scarce compared to normal samples, resulting in serious data imbalance problems. This data distribution makes it difficult for the model to fully learn the characteristics of abnormal patterns, and is prone to underreporting or false alarms. At the same time, factors such as environmental noise and sensor drift will also interfere with the model's judgment and reduce the reliability of detection. Obtaining a large amount of labeled abnormal data is not only time-consuming but also costly, which further limits the application of supervised learning methods.

[0005] Third, although the Transformer architecture performs well in sequence modeling, when it is directly applied to the multivariate time series data of chillers, the computational complexity grows quadratically with the length of the sequence, and the computational overhead is huge when processing long time series; the standard Transformer has difficulty capturing local time series features, affecting the ability to detect short-term anomalies; parameter optimization is difficult, and model performance is highly sensitive to hyperparameter selection.

[0006] In addition, most of the current anomaly detection methods use fixed model structures and hyperparameter configurations, which are difficult to adapt to the dynamic changes in equipment operating status and environmental conditions. This static detection strategy often exhibits poor generalization performance and environmental adaptability in practical applications, especially in complex industrial systems such as chillers. The equipment operating status will change significantly with factors such as load changes and seasonal changes. Fixed model configurations are difficult to maintain good detection performance continuously.

[0007] Through the above analysis, the problems and defects of the prior art are as follows:

[0008] The existing technology does not take into account the complex relationship between the operating data of the chiller and the small number of normal samples. In addition, the parameter configuration of the Transformer architecture is fixed during calculation, which makes it difficult to adapt to the dynamic changes of the equipment based on environmental conditions. Summary of the invention

[0009] The embodiments of the present application provide a Transformer-based chiller diagnosis method, device and medium, which can solve the problems in the prior art that the complex relationship between the operating data of the chiller and the small number of normal samples are not taken into account, and the parameter configuration of the Transformer architecture is fixed during calculation, making it difficult to adapt to the dynamic changes of the equipment based on environmental conditions.

[0010] In the first aspect, an embodiment of the present application provides a Transformer-based chiller diagnosis method, the method comprising: obtaining original time series data of the chiller through a multidimensional sensor, the original time series data including temperature data, pressure data, flow data, current data and liquid level data; enhancing and fusing the time series data to obtain enhanced time series data; using an improved Transformer architecture to model the enhanced time series data to obtain an abnormal diagnosis model for the chiller; introducing automated hyperparameters to the abnormal diagnosis model, calculating the optimal parameter combination of the abnormal diagnosis model, and performing abnormal diagnosis on the operating data of the chiller.

[0011] In one implementation of the present application, before obtaining the original time series data of the chiller through a multidimensional sensor, the method also includes: obtaining historical abnormal data and calculating the probability distribution of the historical abnormal data of each multidimensional sensor, and obtaining the contribution of each multidimensional sensor according to the probability distribution; learning a first weight coefficient according to the contribution, and assigning it to the corresponding multidimensional sensor.

[0012] In one implementation of the present application, an improved Transformer architecture is used to model the enhanced time series data to obtain an abnormal diagnosis model for a chiller, specifically including: slicing the enhanced time series data through overlapping sliding windows to obtain data segments, the temperature data including meteorological temperature, and evaporator inlet and outlet water temperature, condenser inlet and outlet water temperature, compressor suction and exhaust temperature; based on the long-term dependency between the evaporator inlet and outlet water temperature and the evaporation pressure, the condenser inlet and outlet water temperature, and the compressor suction and exhaust temperature, as well as the local characteristics of the chilled water inlet and outlet temperature difference, inter-slice global attention and intra-slice local attention calculations are performed on the data segments to obtain global and local first feature information; through adaptive weights, the global and local feature information are dynamically fused to obtain the second feature information of the chiller.

[0013] In one implementation of the present application, after the enhanced time series data is segmented by overlapping sliding windows to obtain data segments, the method also includes: calculating the correlation coefficient of any two data segments, if the correlation coefficient exceeds a preset threshold, the corresponding data segment is regarded as a positive sample; the temperature data, pressure data, flow data, current data and liquid level data are divided into intervals according to numerical values ​​to obtain interval data respectively; a preset time distance threshold is set, and the interval data exceeding the time distance threshold is defined as a negative sample.

[0014] In one implementation of the present application, after defining the interval data exceeding the time distance threshold as a negative sample, the method further includes: calculating the weight according to the ratio of the number of positive samples and negative samples; modeling according to the first weight coefficient, the second feature information and the weights of the positive and negative samples to obtain an abnormal diagnosis model of the chiller.

[0015] In one implementation of the present application, automated hyperparameters are introduced into the anomaly diagnosis model to calculate the optimal parameter combination of the anomaly diagnosis model, specifically including: based on the anomaly diagnosis model including structural parameters, the structural parameters include the number of Transformer layers, the number of attention heads, and the hidden layer dimension; using a progressive search method to identify the sensitivity of the structural parameters, dynamically adjusting the structural parameters according to the sensitivity, and performing iterative optimization to obtain the optimal parameter combination of the anomaly diagnosis model.

[0016] In one implementation of the present application, time series data is enhanced and fused to obtain enhanced time series data, specifically including: based on the seasonal characteristics of the time series data of the chiller, a preset proportion of time steps are selected in the time series data for masking, and noise is injected to obtain masked data and noisy data, and the masking includes continuous time period masking; based on historical abnormal data, abnormal samples are generated when the compressor exhaust temperature is higher than the temperature threshold when the compressor fails, and when the condenser flow is lower than the flow threshold, and the condensing pressure is higher than the pressure threshold; the masked data, the noisy data and the abnormal samples are combined to obtain enhanced time series data; the enhanced time series data is normalized by mean variance to obtain normally distributed data; the normally distributed data is timestamp aligned, and a weighted average calculation is performed according to a first weight coefficient to obtain enhanced and fused time series data.

[0017] In one implementation of the present application, after performing abnormal diagnosis on the operating data of the chiller, the method further includes: obtaining real-time operating data and meteorological data of the chiller, and continuously monitoring the chiller; and predicting the probability and time of occurrence of abnormal data through an abnormal diagnosis model.

[0018] In the second aspect, an embodiment of the present application also provides a Transformer-based chiller diagnostic device, the device comprising at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can: obtain original time series data of the chiller through a multidimensional sensor, the original time series data comprising temperature data, pressure data, flow data, current data and liquid level data; fuse and enhance the time series data to obtain enhanced time series data; use the improved Transformer architecture to model the enhanced time series data to obtain an abnormal diagnosis model for the chiller; introduce automated hyperparameters into the abnormal diagnosis model to calculate the optimal parameter combination of the abnormal diagnosis model, and perform abnormal diagnosis on the operating data of the chiller.

[0019] In the third aspect, an embodiment of the present application also provides a Transformer-based non-volatile computer storage medium for chiller diagnosis, which stores computer executable instructions, and the computer executable instructions are set to: obtain the original time series data of the chiller through a multi-dimensional sensor, the original time series data including temperature data, pressure data, flow data, current data and liquid level data; fuse and enhance the time series data to obtain enhanced time series data; use the improved Transformer architecture to model the enhanced time series data to obtain an abnormal diagnosis model for the chiller; introduce automated hyperparameters to the abnormal diagnosis model to calculate the optimal parameter combination of the abnormal diagnosis model, and perform abnormal diagnosis on the operating data of the chiller.

[0020] The Transformer-based chiller diagnosis method, equipment and medium provided in the embodiment of the present application not only significantly improve the accuracy and recall rate of anomaly detection, but also reduce the computational overhead through automated parameter optimization and incremental update mechanisms, thus realizing the transition from traditional manual monitoring to intelligent anomaly detection; the sample imbalance problem is solved through a specially designed data enhancement strategy, thereby improving the generalization ability of the model. The organic combination of these three core technologies can not only realize the early identification and early warning of abnormal conditions and reduce the risk of equipment failure, but also significantly improve the detection accuracy, reduce false alarms and missed alarms, and provide a new technical solution for the intelligent operation and maintenance of chillers. It can be widely used in the intelligent operation and maintenance and fault diagnosis of various chiller systems, which is of great significance to improving equipment operation efficiency and reducing maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0022] Figure 1 A flowchart of a Transformer-based chiller diagnosis method provided in an embodiment of the present application;

[0023] Figure 2 A schematic diagram of the overall architecture of a Transformer-based chiller diagnosis method provided in an embodiment of the present application;

[0024] Figure 3 A schematic diagram of a multi-segment Transformer comparative learning architecture of a Transformer-based chiller diagnosis method provided in an embodiment of the present application;

[0025] Figure 4A schematic diagram of the internal structure of a Transformer-based chiller diagnostic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0027] The embodiments of the present application provide a Transformer-based chiller diagnosis method, device and medium, which solve the problems in the prior art that the complex relationship between the operating data of the chiller and the small number of normal samples are not taken into account, and the parameter configuration of the Transformer architecture is fixed during calculation, making it difficult to adapt to the dynamic changes of the equipment based on environmental conditions.

[0028] The technical solution proposed in the embodiments of the present application is described in detail below with reference to the accompanying drawings.

[0029] Figure 1 A flow chart of a chiller diagnosis method based on Transformer provided in an embodiment of the present application. Figure 1 As shown, a Transformer-based chiller diagnosis method provided in an embodiment of the present application specifically includes the following steps:

[0030] Step 10: Obtain the original time series data of the chiller through a multi-dimensional sensor. The original time series data includes temperature data, pressure data, flow data, current data, and liquid level data.

[0031] In this step, first, the chiller mainly includes a compressor, an evaporator, a condenser and an expansion valve. Its working process is that the compressor sucks in low-temperature and low-pressure refrigerant gas, compresses it into high-temperature and high-pressure gas, and discharges it into the condenser. The condenser cools and liquefies the high-temperature and high-pressure refrigerant gas, releases heat, and the liquid refrigerant throttles and reduces pressure when flowing through the expansion valve, and becomes low-pressure and low-temperature wet steam to enter the evaporator. The refrigerant liquid in the evaporator absorbs the heat of the chilled water and then vaporizes, cooling the chilled water. The vaporized refrigerant gas is sucked in by the compressor and compressed again to start a new refrigeration cycle. The temperature data includes the evaporator temperature, the condensation temperature, the compressor suction temperature, the exhaust temperature, etc. The pressure parameters include the evaporation pressure, the condensation pressure, etc. The flow parameters include the cooling water flow, etc. The liquid level parameters include the refrigerant liquid level, etc.

[0032] Step 20: Enhance and fuse the time series data to obtain enhanced time series data.

[0033] As an optional embodiment, the time series data is enhanced and fused to obtain enhanced time series data, which may specifically include: Step 201: Based on the seasonal characteristics of the time series data of the chiller, a preset proportion of time steps are selected in the time series data for masking, and noise is injected to obtain masked data and noisy data, and the masking includes continuous time period masking.

[0034] In this step, a hierarchical time series masking scheme is adopted to target the periodic characteristics of the chiller time series data. Random time step masking: a certain proportion of time steps are randomly selected in the time series for masking. The masking probability distribution is:

[0035]

[0036] where p m is the basic masking probability, r is a random number, and α is the masking ratio threshold. Considering the continuity of the device state, such as the repeated summer and winter of the air conditioner, a masking strategy for continuous time periods is proposed: for a sequence of length L, a continuous segment of length l is selected for masking, where l obeys the following distribution:

[0037] l~Uniform(l min ,l max )

[0038] This strategy can simulate situations such as data loss or temporary sensor failure during equipment operation, and improve the model's ability to handle incomplete data.

[0039] In addition, for the multi-dimensional sensor data of the chiller, a masking scheme for feature correlation is proposed: 1. Independent feature masking: independent masking of different sensor data to simulate the failure of a single sensor; 2. Associated feature masking: based on expert knowledge, physically related sensor combinations such as inlet and outlet water temperature and pressure are masked at the same time; 3. Conditional feature masking: selectively mask the key features of a specific working stage according to the operating conditions of the equipment.

[0040] Then, according to the common noise types in the actual operation of the chiller: Gaussian white noise: simulates sensor random error; mutation noise: simulates sensor instantaneous interference; drift noise: simulates sensor zero drift; periodic noise: simulates external environmental interference. For each noise mode, a corresponding mathematical model is designed, taking drift noise as an example:

[0041] x′ t =x t +d·t+∈

[0042] Where d is the drift rate and ∈ is the random disturbance term.

[0043] In order to ensure the rationality of the injected noise, an adaptive intensity adjustment mechanism based on data characteristics can be adopted: intensity control based on signal-to-noise ratio: dynamically adjust the noise amplitude according to the signal strength of the original data; condition-aware noise adjustment: use different noise intensities under different operating conditions; progressive noise injection: use a smaller noise intensity at the beginning of training and gradually increase it as the training progresses.

[0044] Step 202: Based on historical abnormal data, an abnormal sample is generated for when the compressor discharge temperature is higher than a temperature threshold, when the compressor fails, when the condenser flow rate is lower than a flow rate threshold, and when the condensing pressure is higher than a pressure threshold.

[0045] In this step, the equipment-level anomaly can be a compressor failure mode; a heat exchanger efficiency reduction mode; a refrigerant leakage mode; a valve failure mode. The system-level anomaly can be a load imbalance mode; a control system abnormality mode; a system efficiency degradation mode. Based on the abnormal mode, an abnormal data sample that conforms to the laws of physics is generated. First, a quantitative indicator of the degree of abnormality is defined:

[0046]

[0047] where w i is the weight of each feature, μ i and σ i It is the statistical characteristics of normal operation.

[0048] Step 203: combining the mask data, the noise data and the abnormal samples to obtain enhanced time series data;

[0049] Step 204: normalize the mean and variance of the enhanced time series data to obtain normally distributed data.

[0050] In this step, in order to eliminate the dimensional differences between different sensor data and improve the stability of model training, a multi-level feature preprocessing strategy is adopted. This process includes the following key steps: First, standardization is performed to normalize the mean variance of each feature dimension, ensuring that different types of sensor data are mapped to the same numerical range, which facilitates unified processing by the model. For the d-th dimension feature, its standardization process can be expressed as:

[0051]

[0052] where μ d and σ d are the mean and standard deviation of the dimension features respectively.

[0053] Secondly, considering the common outlier problem in industrial data, a quantile-based robust normalization method is introduced, using the 25th percentile and the 75th percentile as the normalization benchmark. Compared with the traditional maximum and minimum value normalization method, it can better handle the impact of outliers:

[0054]

[0055] where q 0.25 and q 0.75 denote the 25th and 75th quantiles respectively, which are not overly affected by extreme values.

[0056] Step 205: aligning the timestamps of the normally distributed data, and performing weighted average calculation according to the first weight coefficient to obtain enhanced and fused time series data.

[0057] Step 30: Use the improved Transformer architecture to model the enhanced time series data and obtain the abnormality diagnosis model of the chiller.

[0058] As an optional embodiment, the enhanced time series data is modeled using an improved Transformer architecture to obtain a chiller abnormality diagnosis model, which may specifically include: Step 301: Slice the enhanced time series data through overlapping sliding windows to obtain data segments, the temperature data includes meteorological temperature, evaporator inlet and outlet water temperature, condenser inlet and outlet water temperature, compressor suction and exhaust temperature, and chilled water inlet and outlet temperature difference, and the pressure data includes evaporation pressure and condensing pressure.

[0059] In this step, an innovative overlapping sliding window sharding strategy is adopted to effectively solve the computational complexity problem of traditional Transformer when processing long sequences. It not only maintains the continuity of time series data, but also captures feature correlations across segments through segment overlap.

[0060] Specifically, for the original time series data of the chiller, which contains multidimensional sensor data (including temperature, pressure, flow, etc.) with T time steps, in practical applications, the window size w and step size s can be used for sequence slicing. Each data fragment contains w consecutive time steps. The degree of overlap between adjacent fragments is determined by the step size s. This slicing method ensures the continuity of information and also provides sufficient context information. The overlap rate is an important indicator to measure the degree of information sharing between adjacent fragments. It is calculated as follows:

[0061]

[0062] A large number of experiments have verified that the best effect can be achieved when the overlap rate is set within the range of 0.3-0.5. This range ensures sufficient information redundancy while avoiding excessive computational overhead.

[0063] In addition, an adaptive sharding mechanism can be introduced to dynamically adjust the window size according to the time-varying characteristics of the data. Specifically, during periods of drastic changes in device load, the system will automatically reduce the window size to capture more fine-grained change characteristics; while during stable device operation, the window size will be appropriately increased to obtain more contextual information.

[0064] Step 302: Based on the long-term dependency between the evaporator inlet and outlet water temperature and evaporation pressure, the condenser inlet and outlet water temperature and condensation pressure, the compressor suction and exhaust temperatures, and the local characteristics of the chilled water inlet and outlet temperature difference, inter-slice global attention and intra-slice local attention calculations are performed on the data segments to obtain global and local first feature information.

[0065] In this step, the inter-slice global attention mechanism aims to capture the long-range dependencies between different time segments. For the sliced ​​sequence data, the query, key value, and value matrices are first obtained through linear projection. This process can be expressed as:

[0066]

[0067] Where Q g and K g are query and key-value matrices respectively, and d is the dimension of the attention head, which enables the model to adaptively focus on time segments with strong correlation and effectively capture long-term dependencies. In order to further improve the expressive power of the attention mechanism, a multi-head attention design is adopted to project the input features into multiple subspaces, calculate the attention independently in each subspace, and finally merge the results of each head, so that the model can understand the temporal relationship of the data from multiple perspectives at the same time, significantly improving the comprehensiveness of feature extraction.

[0068] Then, considering that the abnormality of the chiller unit often manifests as a change in the local time series pattern, we first enhance the local features of each data segment, introduce position encoding to retain the position information of the sequence, and limit the calculation scope of attention by designing the local receptive field matrix:

[0069]

[0070] Where M is the local receptive field matrix, which is used to control the calculation range of attention, ensuring that the model can accurately capture local time series features and has a strong recognition ability for short-term abnormal patterns. In practical applications, the local attention mechanism is particularly effective in identifying short-term events such as compressor start and stop, valve switching, etc. By adjusting the local window size, the computational efficiency and feature extraction capabilities can be flexibly balanced. For example, for high-frequency sampled sensor data, a larger local window can be used; for low-frequency sampled data, the window size can be appropriately reduced.

[0071] Step 303: Dynamically fuse the global and local feature information through adaptive weights to obtain the second feature information of the chiller.

[0072] In this step, in order to make full use of the feature information at both the global and local levels, an adaptive feature fusion mechanism is adopted to achieve dynamic fusion of global and local features by learning adaptive weights:

[0073] F=α☉F g +(1-α)☉F l

[0074] where F g and F l are the global and local attention features respectively, and α is the adaptive weight coefficient, which allows the model to automatically adjust the importance of global and local features according to different operating scenarios.

[0075] As an optional embodiment, after the enhanced time series data is segmented by overlapping sliding windows to obtain data segments, the method may further include: step 304: calculating the correlation coefficient between any two data segments, and if the correlation coefficient exceeds a preset threshold, the corresponding data segment is regarded as a positive sample;

[0076] In this step, for any two time segments, their correlation coefficient is calculated:

[0077]

[0078] When the correlation coefficient exceeds a preset threshold, the corresponding fragment pair is regarded as a positive sample. This correlation-based selection strategy ensures the consistency of the temporal pattern of the positive sample pair.

[0079] Step 305: The temperature data, pressure data, flow data, current data and liquid level data are divided into intervals according to the values ​​to obtain interval data respectively.

[0080] In this step, for example, the current range is divided into several sub-intervals, such as a low current interval, a medium current interval, and a high current interval.

[0081] Step 306: Preset a time distance threshold and define interval data exceeding the time distance threshold as negative samples.

[0082] In this step, for the data that has been divided into intervals, a hierarchical negative sample selection strategy can be adopted to ensure the diversity and representativeness of negative samples. The basic negative sample set is defined based on the time distance, and then the most challenging samples are selected through the difficult example mining mechanism. This strategy not only improves the model's discrimination ability, but also accelerates the training convergence speed. For example, selecting data fragments under different operating conditions as negative samples and selecting fragments under different intervals can help the model learn more robust feature representations. At the same time, by dynamically adjusting the difficulty of negative samples, the model training can be gradually optimized.

[0083] As an optional embodiment, after defining the interval data exceeding the time distance threshold as negative samples, the method may further include: Step 307: calculating weights according to the ratio of the number of positive samples to the number of negative samples.

[0084] In this step, the improved InfoNCE loss function is used to introduce temperature parameters and sample weights to optimize the feature learning process:

[0085]

[0086] Where τ is the temperature parameter, which is used to adjust the distribution of the feature space, w j and w k are the weights of positive and negative samples respectively, which not only improves the model's representation learning ability, but also enhances its ability to recognize abnormal patterns.

[0087] In this way, the multi-segment Transformer contrastive learning architecture can effectively process the long sequence time data of the chiller and learn discriminative feature representations, providing a reliable foundation for subsequent anomaly detection. Experimental results show that the architecture has achieved significant improvements in multiple indicators such as anomaly detection accuracy and recall rate.

[0088] Step 308: Modeling is performed according to the first weight coefficient, the second feature information, and the weights of the positive and negative samples to obtain an abnormality diagnosis model for the chiller.

[0089] Step 40: Introduce automated hyperparameters into the anomaly diagnosis model to calculate the optimal parameter combination of the anomaly diagnosis model, and perform anomaly diagnosis on the operating data of the chiller.

[0090] As an optional embodiment, automatic hyperparameters are introduced into the anomaly diagnosis model to calculate the optimal parameter combination of the anomaly diagnosis model, which may specifically include: Step 401: based on the anomaly diagnosis model including structural parameters, the structural parameters include the number of Transformer layers, the number of attention heads, and the hidden layer dimension.

[0091] In this step, we mainly perform adaptive adjustments on the key architectural parameters of the multi-segment Transformer model. First, we define a search space containing the core parameters of the model structure, including the number of Transformer layers, the number of attention heads, the hidden layer dimension, etc.

[0092] Step 402: Using a progressive search method to identify the sensitivity of the structural parameters, dynamically adjusting the structural parameters according to the sensitivity, and performing iterative optimization to obtain the optimal parameter combination of the abnormality diagnosis model.

[0093] In this step, in order to improve the search efficiency, a parameter importance evaluation method based on performance sensitivity is adopted:

[0094]

[0095] Where S i represents the sensitivity of the i-th parameter, ΔP j represents the performance change after the jth disturbance, Δθ i The parameter changes can identify the structural parameters that have the most significant impact on the model performance, so that these parameters can be optimized first. In the initial stage, a larger search step is used to quickly locate the reasonable range of parameters; as the optimization process progresses, the search step is gradually reduced to achieve fine adjustment of the parameters; an early stopping mechanism is introduced to terminate the search in time when there is no obvious improvement after multiple rounds of optimization.

[0096] As an optional embodiment, before obtaining the original time series data of the chiller through the sensor, the method may further include: Step 101: obtaining historical abnormal data, calculating the probability distribution of the normal distribution data of each multidimensional sensor in the historical abnormal data, and obtaining the contribution of each sensor according to the probability distribution;

[0097] Step 102: Learn a first weight coefficient according to the contribution and assign it to the corresponding multi-dimensional sensor.

[0098] As an optional embodiment, after performing abnormal diagnosis on the operating data of the chiller, the method also includes: obtaining real-time operating data and meteorological data of the chiller, and continuously monitoring the chiller; and predicting the probability and time of occurrence of abnormal data through an abnormal diagnosis model.

[0099] In summary, an innovative multi-slice Transformer architecture is used to model the multivariate time series data of the chiller. High-precision anomaly detection is achieved through a dual attention mechanism (including inter-slice global attention and intra-slice local attention) and a contrastive learning strategy. Automated hyperparameter optimization technology is introduced to ensure that the model maintains optimal performance under different operating environments through a dynamic parameter search space.

[0100] The above is an embodiment of the method proposed in this application. Based on the same inventive concept, the embodiment of this application also provides a Transformer-based chiller diagnostic device, whose structure is as follows Figure 4 shown.

[0101] Figure 4 The internal structure diagram of a Transformer-based chiller diagnostic device provided in the embodiment of the present application. Figure 4 As shown, the device includes:

[0102] at least one processor 401;

[0103] and, a memory 402 communicatively coupled to the at least one processor;

[0104] Among them, the memory 402 stores instructions that can be executed by at least one processor, and the instructions are executed by at least one processor 401 so that the at least one processor 401 can: obtain original time series data of the chiller through a multidimensional sensor, the original time series data including temperature data, pressure data, flow data, current data and liquid level data; enhance and fuse the time series data to obtain enhanced time series data; use the improved Transformer architecture to model the enhanced time series data to obtain an abnormal diagnosis model of the chiller; introduce automated hyperparameters into the abnormal diagnosis model to calculate the optimal parameter combination of the abnormal diagnosis model, and perform abnormal diagnosis on the operating data of the chiller.

[0105] Some embodiments of the present application provide corresponding Figure 1 A non-volatile computer storage medium for chiller diagnosis based on Transformer is provided, which stores computer executable instructions, and the computer executable instructions are set to: obtain original time series data of the chiller through a multi-dimensional sensor, and the original time series data includes temperature data, pressure data, flow data, current data and liquid level data; enhance and fuse the time series data to obtain enhanced time series data; use the improved Transformer architecture to model the enhanced time series data to obtain an abnormal diagnosis model of the chiller; introduce automated hyperparameters to the abnormal diagnosis model to calculate the optimal parameter combination of the abnormal diagnosis model, and perform abnormal diagnosis on the operation data of the chiller.

[0106] Each embodiment in this application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the IoT device and medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0107] The system and medium provided in the embodiments of the present application correspond one-to-one to the method. Therefore, the system and medium also have similar beneficial technical effects to the corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the system and medium will not be repeated here.

[0108] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0109] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0110] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0111] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0112] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0113] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0114] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0115] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0116] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A Transformer-based chiller diagnosis method, characterized in that: The method comprises: Acquire original time series data of the chiller through a multi-dimensional sensor, wherein the original time series data includes temperature data, pressure data, flow data, current data and liquid level data; Enhance and fuse the time series data to obtain enhanced time series data; The enhanced time series data is modeled using the improved Transformer architecture to obtain the abnormal diagnosis model of the chiller; An automated hyperparameter is introduced into the abnormality diagnosis model to calculate an optimal parameter combination of the abnormality diagnosis model and perform abnormality diagnosis on the operating data of the chiller.

2. A Transformer-based chiller diagnosis method according to claim 1, characterized in that: Before acquiring the original time series data of the chiller through the multi-dimensional sensor, the method further includes: Acquire historical abnormal data, calculate the probability distribution of the historical abnormal data of each multidimensional sensor, and obtain the contribution of each multidimensional sensor according to the probability distribution; A first weight coefficient is learned according to the contribution degree and allocated to the corresponding multi-dimensional sensor.

3. A Transformer-based chiller diagnosis method according to claim 2, characterized in that: The enhanced time series data is modeled using the improved Transformer architecture to obtain the abnormal diagnosis model of the chiller, which includes: The enhanced time series data is sliced ​​by overlapping sliding windows to obtain data segments, wherein the temperature data includes meteorological temperature, evaporator inlet and outlet water temperature, condenser inlet and outlet water temperature, compressor suction and exhaust temperature, and chilled water inlet and outlet temperature difference, and the pressure data includes evaporation pressure and condensation pressure; Based on the long-term dependency between the evaporator inlet and outlet water temperature and evaporation pressure, the condenser inlet and outlet water temperature and condensation pressure, the compressor suction and exhaust temperatures, and the local characteristics of the chilled water inlet and outlet temperature difference, the data segments are subjected to inter-slice global attention and intra-slice local attention calculations to obtain global and local first feature information; The global and local feature information are dynamically fused through adaptive weights to obtain second feature information of the chiller.

4. A Transformer-based chiller diagnosis method according to claim 3, characterized in that: After the enhanced time series data is fragmented by overlapping sliding windows to obtain data fragments, the method further includes: Calculate the correlation coefficient between any two data segments. If the correlation coefficient exceeds a preset threshold, the corresponding data segment is regarded as a positive sample. The temperature data, pressure data, flow data, current data and liquid level data are divided into intervals according to the values ​​to obtain interval data respectively; A time distance threshold is preset, and interval data exceeding the time distance threshold is defined as a negative sample.

5. A Transformer-based chiller diagnosis method according to claim 4, characterized in that: After defining the interval data exceeding the time distance threshold as negative samples, the method further includes: Calculate the weight according to the ratio of the number of positive samples to the number of negative samples; Modeling is performed based on the first weight coefficient, the second feature information, and the weights of the positive and negative samples to obtain an abnormality diagnosis model for the chiller.

6. The Transformer-based chiller diagnosis method according to claim 1, characterized in that: Introducing automated hyperparameters into the anomaly diagnosis model to calculate the optimal parameter combination of the anomaly diagnosis model, specifically including: Based on that the abnormal diagnosis model includes structural parameters, the structural parameters include the number of Transformer layers, the number of attention heads, and the hidden layer dimension; The sensitivity of the structural parameters is identified by using a progressive search method, the structural parameters are dynamically adjusted according to the sensitivity, and iterative optimization is performed to obtain the optimal parameter combination of the abnormality diagnosis model.

7. The Transformer-based chiller diagnosis method according to claim 2, characterized in that: The time series data is enhanced and fused to obtain enhanced time series data, specifically including: Based on the seasonal characteristics of the time series data of the chiller, a preset proportion of time steps are selected in the time series data for masking, and noise is injected to obtain masked data and noisy data, wherein the masking includes continuous time period masking; Based on the historical abnormal data, generating abnormal samples for when the compressor exhaust temperature is higher than a temperature threshold value when the compressor fails, and when the condenser flow rate is lower than a flow rate threshold value, and the condensing pressure is higher than a pressure threshold value; Combining the mask data, the noised data and the abnormal samples to obtain enhanced time series data; Normalizing the mean and variance of the enhanced time series data to obtain normally distributed data; The normally distributed data are timestamp aligned, and weighted average calculation is performed according to the first weight coefficient to obtain enhanced and fused time series data.

8. The Transformer-based chiller diagnosis method according to claim 1, characterized in that: After performing abnormal diagnosis on the operating data of the chiller, the method further includes: Obtain real-time operating data and meteorological data of the chiller and continuously monitor the chiller; The occurrence probability and time of abnormal data are predicted by the abnormal diagnosis model.

9. A Transformer-based chiller diagnostic device, characterized in that: The device comprises: at least one processor; and, a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to: Acquire original time series data of the chiller through a multi-dimensional sensor, wherein the original time series data includes temperature data, pressure data, flow data, current data and liquid level data; Enhance the time series data to obtain enhanced time series data; The enhanced time series data is modeled using the improved Transformer architecture to obtain the abnormal diagnosis model of the chiller; An automated hyperparameter is introduced into the abnormality diagnosis model to calculate an optimal parameter combination of the abnormality diagnosis model, and abnormality diagnosis is performed on the operating data of the chiller.

10. A non-volatile computer storage medium for Transformer-based chiller diagnosis, storing computer executable instructions, characterized in that: The computer executable instructions are configured to: Acquire original time series data of the chiller through a multi-dimensional sensor, wherein the original time series data includes temperature data, pressure data, flow data, current data and liquid level data; Enhance the time series data to obtain enhanced time series data; The enhanced time series data is modeled using the improved Transformer architecture to obtain the abnormal diagnosis model of the chiller; An automated hyperparameter is introduced into the abnormality diagnosis model to calculate an optimal parameter combination of the abnormality diagnosis model, and abnormality diagnosis is performed on the operating data of the chiller.

Images