An abnormality identification method and device for a liquid cooling machine and related equipment

Abstract

The application discloses an abnormality identification method and device of a liquid cooling machine and related equipment, wherein the method comprises the following steps: obtaining operation data time sequence and outlet water temperature time sequence of the liquid cooling machine; inputting the operation data time sequence into a pre-trained time sequence feature extraction model to obtain operation features, and determining temperature rise features of the liquid cooling machine based on the outlet water temperature time sequence; combining the operation features and the temperature rise features to obtain combined features; obtaining a normal sample library, wherein the normal sample library comprises a plurality of normal sample features; determining an abnormality score of the combined features based on the similarity between the combined features and the normal sample features; and determining that the liquid cooling machine is in an abnormal state if the abnormality score is greater than a dynamic threshold. The application is applied to the technical field of liquid cooling, and the accuracy of abnormality identification is improved by using the operation features and the temperature rise features of the liquid cooling machine as abnormality criteria.

Description

Technical Field

[0001] This application belongs to the field of liquid cooling technology, and in particular relates to a method, device and related equipment for anomaly identification of a liquid cooler. Background Technology

[0002] Liquid coolers are active temperature control devices that utilize liquid circulation and convection to remove heat. They are typically used in applications requiring precise heat dissipation, such as data centers, power electronics, and high-power-density energy storage systems. In large-scale energy storage projects, especially with the continuous increase in cell energy density and system charge-discharge rates, the compact arrangement of battery clusters within containers leads to severe localized heat accumulation. Traditional air cooling is no longer sufficient to meet the temperature uniformity requirements between cells. Therefore, adopting liquid cooling technology to achieve efficient heat exchange, precise temperature control, and system energy efficiency optimization has become the core technical approach for current large-scale energy storage thermal management solutions.

[0003] Current fault diagnosis for liquid chillers typically relies on traditional data statistical methods. This involves acquiring the chiller's current temperature rise status, such as the rate of temperature rise and peak deviation. When the temperature rise exceeds a pre-set fixed threshold, a temperature rise fault is identified and reported. However, this method depends on predefined fault modes and fixed thresholds, and can only identify defined fault types, lacking effective means to identify unforeseen abnormal states.

[0004] Therefore, how to effectively improve the accuracy of liquid chillers in identifying various abnormal states is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] This application provides a method, apparatus, and related equipment for anomaly identification of liquid chillers, which can overcome the limitations of fixed feature definitions in traditional statistical methods and identify various types of anomaly patterns.

[0006] In a first aspect, this application provides a method for anomaly identification of a liquid chiller, comprising: Acquire the operating data timing and outlet water temperature timing of the liquid chiller; The time series of the operating data is input into a pre-trained time series feature extraction model to obtain operating features, and the temperature rise features of the liquid chiller are determined based on the time series of the outlet water temperature. By combining the operational characteristics and the temperature rise characteristics, a combined characteristic is obtained; Obtain a normal sample library, which contains multiple normal sample features; Based on the similarity between the combined feature and the normal sample feature, the anomaly score of the combined feature is determined; If the abnormal score is greater than the dynamic threshold, the liquid chiller is determined to be in an abnormal state, wherein the dynamic threshold is determined based on the features of the plurality of normal samples.

[0007] In one possible implementation, acquiring the operating data timing and outlet water temperature timing of the liquid chiller includes: Based on a preset time window, the operating data and outlet water temperature of the liquid chiller are sampled to obtain multiple raw data. If the number of original data is less than the first number, then the multiple original data are supplemented to the first number based on linear interpolation. The multiple raw data are standardized to obtain the time series of the operating data and the time series of the outlet water temperature.

[0008] In one possible implementation, determining the temperature rise characteristics of the liquid chiller based on the outlet water temperature time sequence includes: Based on the original data corresponding to the start and end times of the time window, the temperature rise rate characteristics of the liquid chiller under the time window are determined. Based on the original data corresponding to the highest and lowest temperatures within the time window, the temperature rise characteristics of the liquid chiller within the time window are determined. Based on the second-order difference of the multiple raw data with respect to time, the characteristics of temperature rise acceleration are determined; Based on the original data corresponding to the highest and median temperatures within the time window, the relative peak deviation characteristics are determined. The temperature rise characteristics are obtained by combining the temperature rise rate characteristics, temperature rise amplitude characteristics, temperature rise acceleration characteristics, and relative peak deviation characteristics.

[0009] In one possible implementation, combining the operating characteristic and the temperature rise characteristic to obtain a combined characteristic includes: The operational features are normalized to obtain a first vector; Based on the temperature rise characteristics and the preset first weight, a second vector is determined; The first vector and the second vector are combined to obtain the combined feature.

[0010] In one possible implementation, determining the anomaly score of the combined feature based on the similarity between the combined feature and the normal sample feature includes: Calculate the feature space distance between the combined feature and each of the normal sample features; Based on the feature space distance, the second number of normal sample features are determined as nearest neighbor sample features; The average value of the feature space distances corresponding to the features of the nearest neighbor samples is determined as the anomaly score.

[0011] In one possible implementation, before determining that the liquid chiller is in an abnormal state if the anomaly score is greater than a dynamic threshold, the anomaly identification method for the liquid chiller further includes: Calculate the similarity between each of the normal sample features and other normal sample features in the normal sample library, and calculate the corresponding normal score; Based on the normal scores corresponding to each of the normal sample features, calculate the median and absolute median difference of the normal sample library; The dynamic threshold is determined based on the weighted sum of the median and the absolute median difference.

[0012] In one possible implementation, after determining the anomaly score of the combined feature based on the similarity between the combined feature and the normal sample feature, the anomaly identification method for the liquid chiller further includes: Obtain the anomaly score within a preset continuous time range; If the probability that the abnormal score within the continuous time range is not greater than the dynamic threshold is less than a preset fluctuation threshold, then the combined feature is determined as a normal sample feature and added to the normal sample library. Update the dynamic threshold.

[0013] In one possible implementation, before inputting the time series of the operating data into a pre-trained time series feature extraction model to obtain operating features, the anomaly identification method for the liquid chiller further includes: The time-series feature extraction model is pre-trained. The pre-training includes: freezing the embedding layer and the third-order encoder layer near the input side of the temporal feature extraction model, training the parameters of the other encoder layers and the classification layer of the temporal feature extraction model with a first learning rate; unfreezing all layers of the temporal feature extraction model, and adjusting all parameters of the temporal feature extraction model with a second learning rate, wherein the second learning rate is less than the first learning rate.

[0014] Secondly, this application provides an anomaly identification device for a liquid chiller, comprising: The timing acquisition module is used to acquire the operating data timing and outlet water temperature timing of the liquid chiller; The feature extraction module is used to input the time series of the operating data into the pre-trained time series feature extraction model to obtain the operating features, and to determine the temperature rise features of the liquid chiller based on the time series of the outlet water temperature. The feature combination module is used to combine operational features and temperature rise features to obtain combined features; The sample acquisition module is used to acquire a normal sample library, which contains multiple normal sample features. The similarity calculation module is used to determine the anomaly score of the combined feature based on the similarity between the combined feature and the normal sample feature; The anomaly detection module is used to determine that the liquid chiller is in an abnormal state if the anomaly score is greater than a dynamic threshold, wherein the dynamic threshold is determined based on the features of multiple normal samples.

[0015] In one possible implementation, the timing acquisition module specifically includes: The sampling submodule is used to sample the operating data and outlet water temperature of the liquid chiller based on a preset time window to obtain multiple raw data. The determination submodule is used to supplement multiple original data to the first number based on linear interpolation if the number of original data is less than the first number. The standardization module is used to perform standardization calculations on multiple raw data to obtain the runtime data time series and the outlet water temperature time series.

[0016] In one possible implementation, the feature extraction module is specifically used to determine the temperature rise rate characteristic of the liquid chiller within the time window based on the raw data corresponding to the start and end times of the time window; to determine the temperature rise amplitude characteristic of the liquid chiller within the time window based on the raw data corresponding to the highest and lowest temperatures within the time window; to determine the temperature rise acceleration characteristic based on the second-order difference of multiple raw data over time; to determine the relative peak deviation characteristic based on the raw data corresponding to the highest and median temperatures within the time window; and to obtain the temperature rise characteristic by combining the temperature rise rate characteristic, temperature rise amplitude characteristic, temperature rise acceleration characteristic, and relative peak deviation characteristic.

[0017] In one possible implementation, the feature combining module specifically includes: Normalization submodule 931 is used to normalize the running features to obtain the first vector; The weighted submodule 932 is used to determine the second vector based on the temperature rise characteristics and the preset first weight; Combined submodule 933 is used to combine the first vector and the second vector to obtain the combined feature.

[0018] In one possible implementation, the similarity calculation module is specifically used to calculate the feature space distance between the combined feature and each normal sample feature; based on the feature space distance, determine a second number of normal sample features as nearest neighbor sample features; and determine the average feature space distance corresponding to the nearest neighbor sample features as the anomaly score.

[0019] In one possible implementation, the anomaly detection device for the liquid chiller further includes: The threshold calculation module is used to calculate the similarity between each normal sample feature and other normal sample features in the normal sample library, and calculate the corresponding normal score; based on the normal scores corresponding to each normal sample feature, it calculates the median and absolute median difference corresponding to the normal sample library; and based on the weighted sum of the median and absolute median difference, it determines the dynamic threshold.

[0020] In one possible implementation, the anomaly detection device for the liquid chiller further includes: The update module is used to obtain abnormal scores within a preset continuous time range; if the probability that the abnormal score within the continuous time range is not greater than the dynamic threshold is less than the preset fluctuation threshold, the combined features will be determined as normal sample features and added to the normal sample library; the dynamic threshold will be updated.

[0021] In one possible implementation, the anomaly detection device for the liquid chiller further includes: The pre-training module is used to pre-train the temporal feature extraction model. The pre-training includes: freezing the embedding layer and the third-order encoder layer near the input side of the temporal feature extraction model, training the parameters of the other encoder layers and the classification layer of the temporal feature extraction model with a first learning rate; unfreezing all layers of the temporal feature extraction model, and adjusting all parameters of the temporal feature extraction model with a second learning rate, which is less than the first learning rate.

[0022] Thirdly, this application provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the method described in the first aspect.

[0023] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in the first aspect above.

[0024] Fifthly, this application provides a computer program product comprising a computer program that, when executed by one or more processors, implements the steps of the method described in the first aspect.

[0025] Existing fault diagnosis methods for liquid chillers rely on predefined fault modes and fixed thresholds, only able to identify defined fault types, lacking effective means to identify unforeseen abnormal states. This application obtains operational features by inputting the liquid chiller's operating data into a pre-trained time-series feature extraction model, and simultaneously determines temperature rise features based on the outlet water temperature time series. The two are then fused and compared with a normal sample library for similarity, using the degree of deviation from the normal distribution as the anomaly criterion, combined with a dynamic threshold determined based on normal samples. The operational features reflect the deep dynamic evolution of the operating data, while the temperature rise features reflect the physical mechanism state of the liquid chiller. The two complement each other, comprehensively characterizing the liquid chiller's operating state from different dimensions. Compared to a single feature, this approach is more sensitive in capturing subtle signs of deviation from normal operating conditions, thus effectively identifying whether the liquid chiller is abnormal. The solution provided by this application can effectively identify even undefined and unforeseen abnormal states, improving the accuracy of anomaly identification.

[0026] It is understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 A system architecture diagram of the anomaly identification method for the liquid chiller provided in the embodiments of this application; Figure 2 This is a flowchart illustrating the anomaly identification method for a liquid chiller provided in an embodiment of this application. Figure 3 A flowchart illustrating the preprocessing stage of the anomaly identification method for a liquid chiller provided in this application embodiment; Figure 4 A flowchart illustrating the method for determining temperature rise characteristics in the anomaly identification method for a liquid chiller provided in this application embodiment; Figure 5 A flowchart illustrating the feature fusion stage of the anomaly identification method for a liquid chiller provided in this application embodiment; Figure 6 A flowchart illustrating the method for determining the anomaly score in the anomaly identification method for a liquid chiller provided in this application embodiment; Figure 7A flowchart illustrating the dynamic threshold determination method for the anomaly identification method of the liquid chiller provided in this application embodiment; Figure 8 A flowchart illustrating the sample update method for the anomaly identification method of the liquid chiller provided in this application embodiment; Figure 9 This is a schematic diagram of the anomaly detection device for a liquid cooler provided in an embodiment of this application; Figure 10 This is a structural block diagram of a portion of the terminal implementing the anomaly identification method for a liquid chiller according to embodiments of this application; Figure 11 This is a structural block diagram of a portion of the server for anomaly detection of a liquid chiller implementing embodiments of this application; Figure 12 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0029] The embodiments of the technical solution of this application will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of this application and are therefore merely examples, and should not be used to limit the scope of protection of this application.

[0030] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0031] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly indicating the number, specific order, or primary and secondary relationship of the indicated technical features.

[0032] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0033] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0034] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), unless otherwise expressly and specifically defined.

[0035] Please see Figure 1 , Figure 1 The system architecture diagram used in the anomaly identification method for the liquid chiller provided in this application embodiment includes a server 110, the Internet, and a terminal 120, etc.

[0036] Terminal 120 is a device that displays the operating status and anomaly identification results of the liquid chiller. It can take various forms, including desktop computers, laptops, PDAs (personal digital assistants), mobile phones, vehicle-mounted terminals, and dedicated terminals. Furthermore, it can be a single device or a combination of multiple devices. For example, multiple desktop computers connected via a local area network, sharing a single monitor, can work collaboratively to form a single terminal 120. Terminal 120 can communicate and exchange data via wired or wireless means.

[0037] Server 110 refers to a computer system that can provide certain services to terminal 120. Compared to ordinary terminal 120, server 110 has higher requirements in terms of stability, security, and performance. Server 110 can be a single high-performance computer in a network platform, a cluster of multiple high-performance computers, a portion of a single high-performance computer (e.g., a virtual machine), or a combination of portions of multiple high-performance computers (e.g., virtual machines).

[0038] A gateway, also known as an internetwork connector or protocol converter, is a computer system or device that enables network interconnection at the transport layer, acting as a translator. It bridges the gap between two systems using different communication protocols, data formats, languages, or even completely different architectures. Gateways can also provide filtering and security functions. Messages sent from terminal 120 to server 110 must be forwarded to the appropriate server 110 through the gateway. Messages sent from server 110 to terminal 120 must also be forwarded to the appropriate terminal 120 through the gateway.

[0039] The anomaly identification method for the liquid chiller provided in this application embodiment can be executed on the terminal 120, or on the server 110, or partly on the terminal 120 and partly on the server 110.

[0040] According to one embodiment of this application, an anomaly identification method for liquid coolers is proposed. This method can be applied to scenarios where liquid cooling is used for heat dissipation, such as large energy storage equipment, data center computer rooms, and power electronic equipment cooling systems.

[0041] Please see Figure 2 , Figure 2 This is a flowchart of a method for anomaly identification of a liquid chiller provided in an embodiment of this application. The method may include: 210. Obtain the operating data timing and outlet water temperature timing of the liquid chiller; 220. Input the runtime data into a pre-trained time-series feature extraction model to obtain runtime features; 230. Determine the temperature rise characteristics of the liquid chiller based on the time sequence of the outlet water temperature; 240. Combining operational characteristics and temperature rise characteristics yields combined characteristics; 250. Obtain the normal sample library, which contains multiple normal sample features; 260. Determine the anomaly score of the combined feature based on the similarity between the combined feature and the normal sample feature; 270. If the abnormal score is greater than the dynamic threshold, the liquid chiller is determined to be in an abnormal state. The dynamic threshold is determined based on the features of multiple normal samples.

[0042] Steps 210-270 are described in detail below.

[0043] In step 210, the operational data sequence refers to the data sequence formed by arranging various sensor data generated during the operation of the liquid chiller in chronological order. This operational data may include, but is not limited to, one or more data such as the liquid chiller's outlet temperature, inlet temperature, input power, compressor frequency, pump speed, coolant flow rate, and ambient temperature; this embodiment does not specifically limit this. The outlet temperature sequence refers to the data sequence formed by arranging the coolant temperature at the liquid chiller's outlet in chronological order.

[0044] In practical implementation, the operating data and outlet water temperature of the liquid chiller can be continuously sampled within a certain time range to obtain multiple sampled values ​​arranged in chronological order within a preset time length, which constitute the aforementioned operating data time series and outlet water temperature time series. The time length corresponding to the time series can be set according to actual monitoring needs and computing resources, such as 3 minutes, 5 minutes, 10 minutes, or 15 minutes.

[0045] In step 220, the temporal feature extraction model is a pre-trained deep learning model used to extract deep dynamic features from temporal data. The temporal feature extraction model can be a model built on the Transformer architecture, or other neural network models with temporal modeling capabilities such as Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs). This application embodiment does not specifically limit this. In a preferred embodiment, the Large version of Chronos (710M parameters) can be used. Chronos-T5-Large has been pre-trained on over 55 public datasets (covering fields such as energy, transportation, healthcare, and finance), with a total data volume of 8.4 billion observations. Experiments show that its performance on unseen datasets (zero-sample setting) significantly outperforms smaller versions and traditional statistical models (such as ARIMA and ETS), and even surpasses dedicated deep learning models (such as PatchTST and N-BEATS) on some tasks, making it effectively applicable to the anomaly identification of the liquid chiller proposed in this application. Through a time-series feature extraction model, operational features reflecting the deep-seated patterns of the liquid chiller's operating status can be extracted from the time-series operational data. These operational features can be represented in the form of feature vectors. For example, after numerical projection, relative position encoding, multi-layer (e.g., 8-layer) Transformer self-attention, and mean pooling in the time-series feature extraction model, a 256-dimensional general time-series embedding vector is output. .

[0046] In one implementation, considering that the quality of the time-series feature extraction model directly affects the representation ability of operational features, and thus determines the accuracy of the entire anomaly identification method, in practical applications of energy storage systems, data from the liquid chiller under normal operating conditions often constitutes the vast majority, while anomaly data is scarce and difficult to label. If a small amount of labeled data is used directly for model training, it is very easy to lead to model overfitting and insufficient generalization ability. To solve the above problems, this application adopts a two-stage pre-training strategy to pre-train the time-series feature extraction model.

[0047] Specifically, prior to step 220, the following is also included: The temporal feature extraction model is pre-trained. The pre-training includes: freezing the embedding layer and the third-order encoder layer near the input side of the temporal feature extraction model, training the parameters of the remaining encoder layers and the classification layer of the temporal feature extraction model with a first learning rate; unfreezing all layers of the temporal feature extraction model, and adjusting all parameters of the temporal feature extraction model with a second learning rate, where the second learning rate is less than the first learning rate.

[0048] Pre-training refers to training a model on large-scale unlabeled time-series data to enable it to extract general time-series features. Based on this, adaptive adjustments are then made for the specific task of identifying anomalies in liquid chillers. In this embodiment, pre-training specifically includes a first stage and a second stage: Phase 1: Freeze the embedding layer and the third-order encoder layer near the input side of the temporal feature extraction model, and train the parameters of the other encoder layers and the classification layer of the temporal feature extraction model with the first learning rate.

[0049] The model consists of two layers: an embedding layer, the first layer, and an encoder layer. The embedding layer maps the input data to vector representations. The encoder layer is the core feature extraction module, extracting deep feature representations from the input sequence. The third parameter is a preset positive integer, such as 2, which freezes the parameters of the embedding layer and the first two encoder layers closest to the input, preventing them from participating in gradient updates during training and preserving the general temporal feature extraction capabilities learned during pre-training. The remaining encoder layers are located near the output, and their parameters are updated normally during training. The classification layer is the output layer, mapping the features extracted by the encoder to classification results (e.g., normal / abnormal). In the first stage, the parameters of the remaining encoder and classification layers are trained using a first learning rate. A relatively high first learning rate (e.g., 1e-3) can be used to accelerate parameter convergence of the model on the target task.

[0050] Second stage: Unfreeze all layers of the temporal feature extraction model, and adjust all parameters of the temporal feature extraction model with a second learning rate, which is less than the first learning rate.

[0051] Specifically, after completing the first stage of training, the embedding layer and the third-order encoder are unfrozen, allowing all layer parameters of the model to participate in training. A second learning rate is then used to fine-tune all parameters of the temporal feature extraction model. The second learning rate is an extremely low learning rate (e.g., 1e-5), much smaller than the first learning rate, to ensure that the learned general features are not drastically altered during global fine-tuning. Fine-tuning is performed only on the basis of the general features, thereby improving the model's performance on the specific task of liquid chiller anomaly identification while maintaining its general feature extraction capabilities.

[0052] Through the above two-stage pre-training strategy, the time-series feature extraction model retains the ability to extract general features at the bottom level in the early stage of training and adapts to the data distribution of specific tasks in the later stage of training. It can effectively extract operating features that reflect the deep-seated laws of the operating status when the operating data of the liquid chiller is limited.

[0053] In one specific implementation, the temporal feature extraction model is a large temporal model built on the Transformer architecture, comprising an embedding layer, an 8-layer encoder, and a classification head. In the first stage, the third number is set to 3, meaning all parameters of the embedding layer and the first 3 encoder layers are frozen. The parameters of the remaining 5 encoder layers (layers 4 to 8) and the classification head are trained with a first learning rate of 1e-3 for 50 training epochs. In the second stage, all layer parameters are unfrozen, and the parameters of all 8 encoder layers and the classification head are fine-tuned with a second learning rate of 1e-5 for 10 training epochs.

[0054] It should be noted that the specific values ​​of the third number of layers (3 layers), the first learning rate (1e-3), the second learning rate (1e-5), and the number of training epochs mentioned above are merely examples. Those skilled in the art can make adaptive adjustments based on the actual situation of the model size and dataset. For example, when the encoder has 12 layers, the third number of layers can be set to 4 or 5 layers; when the model size is large, the first learning rate can be adjusted to 5e-4. All the above adjustments fall within the scope of protection of this application.

[0055] The dataset used for pre-training can be large-scale unlabeled liquid chiller operation data or other industrial time-series data. Unsupervised pre-training strategies such as self-supervised learning or contrastive learning enable the model to initially acquire the ability to extract time-series features. Then, through the above two stages of fine-tuning, it is adapted to the anomaly recognition task.

[0056] In step 230, the temperature rise characteristic is determined based on the time sequence of the outlet water temperature and is used to reflect the multi-dimensional characteristics of the thermodynamic behavior of the liquid chiller. The temperature rise characteristic characterizes the temperature change pattern of the liquid chiller over a certain period of time from the perspective of physical mechanism. It belongs to a different dimension of characterization method from the above-mentioned operating characteristics, and the two can complement each other.

[0057] In step 240, "combination" refers to fusing the two features into a unified feature expression according to preset rules. Through feature combination, operational features extracted from a data-driven perspective and temperature rise features extracted from a physical mechanism perspective can be merged to form a more comprehensive representation of the operational status. In specific implementations, the operational feature vector and temperature rise feature vector can be directly concatenated into a longer feature vector using a splicing method; a weighted fusion method can also be used, assigning weights to the two features and then performing a linear combination; other appropriate feature fusion strategies can also be used, which are not specifically limited in this embodiment. Regardless of the combination method used, the final combined feature simultaneously contains both deep dynamic information of the operational data and physical change information of the outlet water temperature.

[0058] In step 250, the normal sample library refers to a pre-constructed collection storing feature data of the liquid chiller under normal operating conditions. Each normal sample feature is a combined feature obtained by processing the liquid chiller in the same manner as steps 210 to 240 when the liquid chiller was in normal operating conditions at a certain historical moment. That is, each sample feature in the normal sample library corresponds to the combined feature expression of the liquid chiller under normal operating conditions at a certain time window. The normal sample library can be constructed before the liquid chiller is put into operation by collecting sufficient sample data under normal operating conditions and processing it according to the method of this application; or it can be continuously supplemented and updated during the actual operation of the liquid chiller by verifying the operating status. The quantity and quality of samples in the normal sample library directly affect the accuracy of subsequent anomaly judgment. Generally speaking, the more samples there are and the more comprehensive the coverage of normal operating conditions, the higher the reliability of anomaly identification.

[0059] In some embodiments, the normal sample library can be stored in local storage, edge computing devices, or cloud servers, and can be read and accessed by the corresponding devices when anomaly identification is required.

[0060] In step 260, similarity is used to measure the proximity of the combined features of the current time window to the features of historical normal samples in the feature space. Higher similarity indicates that the current operating state of the liquid chiller is closer to its historical normal state, and the lower the probability of it being in an abnormal state; conversely, lower similarity indicates that the current operating state deviates more from the normal state, and the higher the probability of an anomaly. In specific implementations, similarity can be measured by calculating the distance in the feature space; the smaller the distance, the higher the similarity. After calculating the distances with all normal sample features, a quantitative index can be determined based on these distances as an anomaly score. The lower the anomaly score, the lower the probability that the liquid chiller is abnormal in its current state; conversely, the higher the anomaly score, the higher the probability that the liquid chiller is abnormal in its current state.

[0061] In some embodiments, to suppress the interference of noise points or boundary samples in the sample library on the similarity measurement, a nearest-neighbor strategy can be used to determine the anomaly score. Specifically, several normal sample features with the smallest distance from the current combined feature can be selected from the normal sample library as nearest-neighbor samples, and then the anomaly score is determined based on the distance to these nearest-neighbor samples. This method can more robustly reflect the degree of deviation between the current sample and the distribution of normal samples. For example, summation, averaging, weighted averaging, or density estimation can be performed on the examples corresponding to these neighboring samples. Taking density estimation as an example, the density distribution of normal sample features in the feature space is pre-calculated, and then the position of the current combined feature in the feature space is determined; if the combined feature is located in a high-density area, the anomaly score is low; if the combined feature is located in a low-density area, the anomaly score is high.

[0062] In step 270, the dynamic threshold is a judgment criterion determined based on multiple normal sample features in the normal sample library. Unlike a fixed threshold, the dynamic threshold is adaptively determined according to the distribution of the normal sample library, reflecting the dispersion and distribution range of sample features in the current normal sample library. When the abnormal score exceeds the dynamic threshold, it indicates that the deviation of the current combined feature from the normal sample library has exceeded the normal fluctuation range, thus determining that the liquid chiller is in an abnormal state.

[0063] In some embodiments, the dynamic threshold is determined by calculating the corresponding normal score for each normal sample feature in the normal sample library in the same way as in step 260, and then performing statistical analysis on these normal scores. The dynamic threshold obtained in this way is not sensitive to extreme values ​​in the sample library and has good robustness.

[0064] After determining that the liquid chiller is in an abnormal state, further abnormality alerts can be generated, such as through text, sound, light, or push notifications, to notify maintenance personnel to handle the situation promptly. In other embodiments, the abnormality identification results (including abnormality scores, judgment results, and corresponding time-series data) can also be stored for subsequent fault analysis and maintenance decisions.

[0065] In summary, this application provides an anomaly identification method for a liquid chiller. First, the operating data time series and outlet water temperature time series of the liquid chiller are acquired. The operating data time series are input into a pre-trained time series feature extraction model to obtain operating features. Simultaneously, the temperature rise feature of the liquid chiller is determined based on the outlet water temperature time series. The operating features and temperature rise feature are combined to obtain a combined feature. A normal sample library containing multiple normal sample features is acquired. An anomaly score is determined based on the similarity between the combined feature and the normal sample features. When the anomaly score is greater than a dynamic threshold determined based on the normal sample features, the liquid chiller is determined to be in an abnormal state. This method achieves a multi-dimensional representation of the liquid chiller's operating state by fusing operating features and temperature rise features, compensating for the insufficient representation of single features in complex anomaly scenarios. By comparing the similarity of the fused features with the normal sample library and combining it with a dynamic threshold for judgment, any abnormal state deviating from the normal distribution can be identified without relying on predefined fault modes. Furthermore, the dynamic threshold can adapt to normal fluctuations under different operating conditions, solving the problem that existing methods rely on fixed thresholds and predefined fault modes, leading to difficulty in adapting to changes in operating conditions and inability to identify unforeseen anomalies. This effectively improves the accuracy and robustness of anomaly identification.

[0066] In one implementation, please refer to Figure 3 , Figure 3 This is a flowchart of the preprocessing stage of the anomaly identification method for a liquid chiller provided in this application embodiment. The method corresponds to... Figure 2Step 210 in the illustrated embodiment specifically includes: 310. Based on a preset time window, the operating data and outlet water temperature of the liquid chiller are sampled to obtain multiple raw data. 320. If the number of original data is less than the first number, then the multiple original data are supplemented to the first number based on linear interpolation. 330. Standardize the multiple raw data to obtain the running data time series and the outlet water temperature time series.

[0067] Specifically, in step 310, the time window is a preset time interval used to segment and extract sensor data generated during the continuous operation of the liquid chiller. The length of the time window can be set according to actual monitoring needs, the rate of change of the liquid chiller's operating conditions, and the computational resources of the subsequent feature extraction model. For example, it can be set to 3 minutes, 5 minutes, 10 minutes, or 15 minutes, etc. This embodiment does not specifically limit this.

[0068] It is understandable that abnormal states of liquid chillers often manifest as continuous deviations or gradual deterioration of operating parameters over time, and such dynamic trends cannot be captured by instantaneous data at a single moment. Therefore, this application introduces a time window mechanism. By segmenting the liquid chiller's operating data over time, continuous long-term data is divided into several short time-series segments of equal length. The data sequence within each time window reflects the dynamic changes of the liquid chiller during that time period, providing a data foundation containing time evolution information for subsequent time-series feature extraction and temperature rise feature calculation.

[0069] In a specific implementation, the time window can be a sliding window that slides segment by segment on the time axis. There may be overlapping parts between adjacent windows, or there may be no overlapping parts, thereby realizing continuous monitoring of long-term operating data of the liquid chiller and real-time anomaly identification.

[0070] Within a time window, data from various sensors of the liquid chiller are collected at a preset sampling frequency, resulting in multiple raw data sets arranged in chronological order. These raw data sets may include, but are not limited to, the liquid chiller's input power, compressor frequency, pump speed, coolant flow rate, ambient temperature, inlet temperature, and outlet temperature. The collected raw data are organized in a time series format, with each sampling moment corresponding to a set of data, thus forming multivariate time-series data within the time window. It should be noted that the operating data and outlet temperature can be sampled synchronously within the same time window, or they can be collected separately at their respective sampling frequencies and then aligned to the same time axis, as long as the final result is time-aligned time-series data within the same time window.

[0071] Considering that the operating state of a liquid chiller is a dynamic process that changes continuously over time, its sensor data often suffers from problems such as inconsistent sampling frequencies, missing data, and inconsistent dimensions. For example, communication interruptions or momentary sensor failures may lead to missing data at certain times, and differences in sampling frequencies between different sensors may also result in misaligned data. To address this, this application further proposes a data preprocessing method. This method involves regularizing the original data through a preset time window, interpolating and filling in missing data, and standardizing the filled data to eliminate the impact of missing values ​​and dimensional differences on subsequent processing. This provides regular and reliable input data for the time-series feature extraction model and temperature rise feature calculation.

[0072] Therefore, in step 320, to address the issues of missing data and packet loss, this application embodiment sets a first number of data points that should be collected within each time window. For example, if the time window is set to 10 minutes and data is collected once per second, then there should be 600 sampling points within this window, with the first number being 600, corresponding to 600 original data points. If the actual number of samples is less than this first number, it indicates that there is missing data within the time window. If incomplete data is used directly for subsequent processing, it may cause the time series feature extraction model to fail to input correctly or cause deviations in the calculation of temperature rise features. Therefore, this application uses linear interpolation to fill in the missing data points. That is, for a data point at a missing moment within the time window, the two most recent valid data points before and after the missing moment are obtained, and linear interpolation is performed with time as the independent variable and data value as the dependent variable to estimate the data value at the missing moment. After being filled in by linear interpolation, the number of original data points within the time window reaches the first number, forming a complete and continuous time series, ensuring that subsequent steps can be executed normally.

[0073] In some embodiments, when there are too many missing data points (e.g., the missing proportion exceeds a preset threshold), the accuracy of linear interpolation may not meet the requirements. In this case, the data in that time window can be discarded, and the data in the next time window can be re-acquired for processing to ensure the quality of the input data. It is understandable that when there are multiple time windows within a certain period of time with excessive missing data, it indicates that there may be problems such as sensor failure or limited signal transmission, and it can be directly determined that the liquid chiller is in an abnormal state.

[0074] Considering that the dimensions and numerical ranges of data from different sensors may vary significantly (e.g., temperature ranges from -10℃ to 50℃, flow rate ranges from 0 to 100 L / min, and power ranges from 0 to 1000 W), directly inputting this data with different dimensions and numerical ranges into the subsequent time-series feature extraction model may lead to problems such as gradient instability and slow convergence during model training, affecting the feature extraction effect. Therefore, in step 330, this embodiment of the application also includes standardizing the raw data.

[0075] Standardization refers to mapping raw data to a uniform scale range through certain mathematical transformations to eliminate the influence of different dimensions and numerical ranges on subsequent processing. In some embodiments, taking the effluent temperature time series as an example: the standard fraction method is used to standardize the effluent temperature time series, and the calculation formula is as follows:

[0076] in, The original outlet water temperature value at time t within the preset time window. This is the average of all outlet water temperature data within the current time window. This represents the standard deviation of all outlet water temperature data within the current time window. This represents the standardized outlet water temperature value at time t. After the above standardization process, the standardized values ​​corresponding to each sampling time within the current time window are obtained, forming a standardized outlet water temperature time series. When the first number is 600, the standardized output sequence is: .

[0077] In other embodiments, the range normalization method can be used to map the data to the [0,1] interval, or other appropriate standardization methods can be used, as long as the technical objective of eliminating dimensional differences can be achieved.

[0078] For other physical quantities in the operational data (such as inlet temperature, input power, compressor frequency, pump speed, coolant flow rate, etc.), the same or corresponding standardization methods can be used to process them so that all data input to the time series feature extraction model are within a uniform scale range.

[0079] After standardization, data of different dimensions are on the same order of magnitude, which can be effectively processed by the time-series feature extraction model, thereby improving the stability and accuracy of feature extraction.

[0080] In one implementation, please refer to Figure 4 , Figure 4 This is a flowchart illustrating the method for determining temperature rise characteristics in the anomaly identification method for a liquid chiller provided in this application embodiment. The method corresponds to... Figure 2Step 230 in the illustrated embodiment specifically includes: 410. Based on the raw data corresponding to the start and end times of the time window, determine the temperature rise rate characteristics of the liquid chiller within the time window; 420. Based on the original data corresponding to the highest and lowest temperatures within the time window, determine the temperature rise characteristics of the liquid chiller within the time window; 430. Determine the characteristics of temperature rise acceleration based on the second-order difference of multiple raw data over time; 440. Based on the original data corresponding to the highest temperature and median temperature within the time window, determine the relative peak deviation characteristics; 450. By combining the characteristics of temperature rise rate, temperature rise amplitude, temperature rise acceleration, and relative peak deviation, the temperature rise characteristics are obtained.

[0081] During operation, the variation in the outlet water temperature of a liquid chiller directly reflects the system's heat exchange efficiency and heat dissipation capacity. When the liquid chiller malfunctions, the dynamic response of the outlet water temperature often changes accordingly—for example, the temperature rise rate accelerates when the coolant pump is insufficient, the temperature fluctuation amplitude increases when the heat exchanger is fouled, and the temperature exhibits non-stationary acceleration or deceleration changes when the refrigeration system oscillates. A single instantaneous temperature value is insufficient to reflect these complex dynamic anomaly patterns. Therefore, this application extracts multidimensional temperature rise characteristics from the outlet water temperature time series. These temperature rise characteristics characterize the dynamic behavior of the liquid chiller's temperature change within a time window, specifically including temperature rise rate characteristics, temperature rise amplitude characteristics, temperature rise acceleration characteristics, and relative peak deviation characteristics.

[0082] The temperature rise rate characteristic reflects the overall temperature rise trend of the window, and is determined based on the outlet water temperature corresponding to the start and end times of the time window. Specifically, the calculation formula is as follows:

[0083] in, This represents the outlet water temperature value at the end of the time window. This represents the outlet water temperature value at the beginning of the time window. The duration of the time window (e.g., 10 minutes). The temperature change per unit time obtained by dividing the difference between the outlet water temperature at the end and start times within the time window by the duration of the time window is used as the characteristic of the temperature rise rate.

[0084] The temperature rise amplitude characteristic represents the severity of temperature fluctuations within the time window, and is determined based on the highest and lowest outlet water temperatures within that time window. Specifically, the calculation formula is as follows:

[0085] in, and These represent the maximum and minimum outlet water temperatures within the time window, respectively. The difference between the two is used as a characteristic of the temperature rise amplitude.

[0086] Temperature rise acceleration characteristic represents the degree of acceleration / deceleration during temperature rise. Because the input is discrete time-series data, it needs to be approximated using numerical methods. Specifically, the calculation formula is as follows:

[0087] in, For the effluent temperature time series data within the time window, firstly, the first difference of the effluent temperature time series is calculated to obtain the temperature rise rate sequence. Then, the first difference of the temperature rise rate sequence is calculated to obtain the temperature rise acceleration characteristics. .

[0088] The relative peak deviation characteristic characterizes the degree to which the peak temperature deviates from the normal range. It is determined based on the highest outlet water temperature and the median temperature within the time window. This characteristic can robustly characterize the degree to which the peak temperature of the liquid chiller deviates from the normal range within the time window. Specifically, the calculation formula is as follows:

[0089] in, =10^-6 to prevent the denominator from being 0. The median temperature is the temperature within the time window. The relative peak deviation characteristic is obtained by calculating the ratio of the difference between the maximum and median values ​​to the median value of the absolute deviation of each temperature relative to the median value. This method eliminates the influence of the difference in absolute temperature values ​​under different operating conditions on the measurement of deviation. Compared with the traditional Z-score, this method is more resistant to extreme outlier interference.

[0090] By combining the characteristics of temperature rise rate, temperature rise amplitude, temperature rise acceleration, and relative peak deviation, the temperature rise characteristics are obtained. Specifically, these four characteristics are organized into a vector form, forming a four-dimensional temperature rise characteristic vector. This temperature rise feature vector comprehensively characterizes the thermodynamic state of the liquid chiller within the time window from four dimensions: temperature rise rate, fluctuation amplitude, acceleration characteristics, and peak deviation, serving as the basis for subsequent integration with operational characteristics.

[0091] It should be noted that the execution order of steps 410 to 440 is not limited and can be flexibly adjusted according to the actual calculation process, as long as the above four characteristics can be obtained and combined to form the temperature rise characteristics.

[0092] Among them, the temperature rise amplitude characteristic reflects the severity of temperature oscillation within the window, the temperature rise acceleration characteristic reflects the acceleration or deceleration characteristics of the temperature rise process, and the relative peak deviation characteristic measures the degree of deviation of the peak temperature from the normal range. These four characteristics characterize the dynamic behavior of the outlet water temperature from different perspectives, comprehensively representing the thermodynamic state of the liquid chiller within the time window. This provides rich physical information for subsequent anomaly identification, compensating for the shortcomings of single statistical features in characterizing complex anomaly scenarios.

[0093] In one implementation, please refer to Figure 5 , Figure 5 This is a flowchart of the feature fusion stage of the anomaly identification method for a liquid chiller provided in this application embodiment. The method corresponds to... Figure 2 Step 240 in the illustrated embodiment specifically includes: 510. Normalize the running features to obtain the first vector; 520. Based on the temperature rise characteristics and the preset first weight, determine the second vector; 530. Combine the first vector and the second vector to obtain the combined feature.

[0094] After obtaining operational and temperature rise characteristics, it is necessary to effectively integrate the two to fully utilize information from both data-driven and physical mechanism dimensions. However, direct concatenation or simple weighting may face the following problems: the high-dimensional vectors output by time-series feature extraction models (such as large time-series models) are usually located on a high-dimensional manifold, and their directional information carries rich time-series pattern characteristics (such as periodicity, trends, etc.), while the absolute magnitude of the values ​​does not directly reflect the importance of the features; if amplitude normalization (such as Min-Max normalization) is applied to each dimension of the high-dimensional vector, it will destroy the geometric topology of the vector in the high-dimensional space, causing subsequent distance-based similarity calculations to fail. On the other hand, temperature rise characteristics are close to zero under normal conditions, but will increase significantly when temperature rise anomalies occur, and have clear physical meaning and amplitude change patterns.

[0095] In view of this, the embodiments of this application perform different processing on the operating features and the temperature rise features before feature fusion.

[0096] For the running features: Since the running features are high-dimensional feature vectors output by the time-series feature extraction model, to preserve the directional information of these high-dimensional vectors and eliminate the influence of amplitude differences on subsequent distance calculations, L2 norm normalization is used to process the running features. This involves dividing the running feature vector by its L2 norm (Euclidean norm) so that the magnitude of the normalized first vector is 1, i.e., mapped to the unit hypersphere, thus avoiding the destruction of the geometric topology caused by amplitude differences in different dimensions in high-dimensional space. According to high-dimensional statistical laws, the expected value of a single-dimensional value is 1√d (where d is the vector dimension). When the dimension is 256, 1 / ≈0.06, the values ​​of each dimension are within a similar scale range.

[0097] For the temperature rise feature: A preset first weight is obtained, and the feature values ​​of each dimension of the temperature rise feature are multiplied by the first weight to obtain a weighted second vector. In some embodiments, the first weight is a uniform scalar, meaning that all dimensions of the temperature rise feature are multiplied by the same weight value. By setting the value of the first weight, the overall influence of the temperature rise feature on subsequent anomaly detection can be controlled. In other embodiments, the first weight is a weight vector, the dimension of which is the same as the dimension of the temperature rise feature, and each element in the weight vector corresponds to one dimension of the temperature rise feature. Using a weight vector approach allows for the configuration of different weight values ​​for different dimensions of the temperature rise feature. For example, a larger weight can be assigned to feature dimensions with a higher correlation to temperature rise faults, and a smaller weight can be assigned to feature dimensions with a lower correlation, thereby achieving differentiated sensitivity adjustment.

[0098] In one specific implementation, the first weight is a uniform scalar with a value range of 0.0 to 5.0, and a default value of 3.0. This default value is an empirically optimized value, which can be adjusted according to the false alarms and missed alarms of different energy storage power stations. When w > 1, the model's sensitivity to temperature rise-related faults increases.

[0099] In another specific implementation, the first weight is a weight vector w = [w1, w2, w3, w4], corresponding to the weights of the temperature rise rate feature, temperature rise amplitude feature, temperature rise acceleration feature, and relative peak deviation feature, respectively. For example, for a fault scenario where abnormal temperature rise rate is the primary characteristic, w1 can be set to a larger value, and the other weights can be set to smaller values; for a fault scenario where drastic temperature fluctuations are the primary characteristic, w2 can be set to a larger value. Through the above differentiated configuration, the fused features can be made more closely aligned with the anomaly detection requirements of specific application scenarios.

[0100] Under normal operating conditions, the values ​​of each dimension of the temperature rise feature are close to zero. At this point, the value of the first weight has negligible impact on the fused features, and the system primarily relies on operational features for monitoring. When an abnormal temperature rise occurs, the values ​​of each dimension of the temperature rise feature increase significantly. After being amplified by the first weight, the proportion of the temperature rise feature in the fused features can be rapidly increased, thereby effectively capturing temperature rise-related anomalies. In practical applications, the specific values ​​of the weight vector can be set and adjusted based on prior knowledge or historical data for different fault types.

[0101] The first vector, obtained based on normalization, and the second vector, obtained based on weighting, are concatenated to form a combined feature. Taking a first vector of 256 dimensions and a second vector of 4 dimensions as an example, the concatenated combined feature is a 260-dimensional vector. The combined feature simultaneously includes deep temporal features from a data-driven perspective and temperature rise features from a physical mechanism perspective, identifying any undefined abnormal patterns and specifically enhancing the feature response of core liquid chiller faults (such as continuous temperature rise), with the two complementing each other.

[0102] By using the above fusion method, only one parameter (or a set of parameters) of the first weight needs to be adjusted to flexibly control the system’s sensitivity to different types of anomalies, effectively overcoming the shortcomings of traditional methods that rely on multiple manual rules, have limited coverage, and are difficult to adapt to different working conditions.

[0103] In one implementation, please refer to Figure 6 , Figure 6 This is a flowchart illustrating the method for determining the anomaly score in the anomaly identification method for a liquid chiller provided in this application embodiment. The method corresponds to... Figure 2 Step 260 in the illustrated embodiment specifically includes: 610. Calculate the feature space distance between the combined feature and the features of each normal sample; 620. Based on the feature space distance, determine the features of the second number of normal samples as the features of the nearest neighbor samples; 630. The average value of the feature space distances corresponding to the features of neighboring samples is determined as the anomaly score.

[0104] In this embodiment, the similarity between the current combined feature and the features of each normal sample in the normal sample library is compared in the feature space, and the degree of deviation of the current sample from the distribution of normal samples is used as the anomaly criterion. This method does not require manually setting a fixed physical threshold, but rather adaptively generates anomaly scores based on the data distribution of the sample library itself, avoiding reliance on a fixed physical threshold and achieving generalized detection of any unknown anomaly pattern.

[0105] Specifically, the combined features of the current time window and the features of each normal sample in the normal sample library are obtained, and the distance between the combined features and each normal sample feature is calculated in the feature space. The feature space distance is used to measure the similarity between the current combined features and each normal sample feature. The smaller the distance, the closer the current running state is to the normal sample; the larger the distance, the greater the deviation between the current running state and the normal sample.

[0106] In some embodiments, the feature space distance is the L2 Euclidean distance, and its calculation formula is as follows:

[0107] in, The combined features of the current time window, Let M be the feature of the i-th normal sample in the normal sample library, and M be the total number of normal sample features in the normal sample library. Let L2 be the Euclidean distance between the current combined feature and the i-th normal sample feature.

[0108] It should be noted that in other embodiments, the feature space distance can also be other distance measurement methods such as Mahalanobis distance and cosine distance, as long as they can effectively measure the similarity between feature vectors. This application does not make any specific limitations on this.

[0109] In one implementation, the anomaly score can be directly equal to the minimum feature space distance between the combined feature and the normal sample feature; that is, the minimum value among all distances is taken as the anomaly score. However, relying solely on a single minimum distance has significant limitations: the sample library may contain well-labeled normal samples, but their positions in the feature space may be located at the boundaries of the normal distribution or in isolated regions. If this boundary sample is used as the sole reference, the current sample may be assigned an excessively low anomaly score, resulting in missed detection.

[0110] Therefore, in this embodiment, all distances are sorted in ascending order of value, and the top few normal sample features with the smallest distance are selected as the nearest neighbor sample features. The average value of the feature space distances corresponding to the nearest neighbor sample features is determined as the anomaly score.

[0111] The second number is a preset positive integer, for example, it can be set to 5, indicating that the features of the 5 nearest normal samples are selected as the features of the nearest neighbor samples. The specific value of the second number can be adjusted adaptively according to the size of the sample database. When the sample database is large, this value can be increased appropriately to improve statistical stability, and when the sample database is small, this value can be decreased appropriately to avoid introducing too many distant samples.

[0112] In one implementation, please refer to Figure 7 , Figure 7 This is a flowchart illustrating the dynamic threshold determination method for the anomaly identification method of a liquid chiller provided in this application embodiment. Specifically, in... Figure 2 Before step 270 in the illustrated embodiment, the method further includes: 710. Calculate the similarity between each normal sample feature and other normal sample features in the normal sample library, and calculate the corresponding normal score; 720. Based on the normal scores corresponding to the characteristics of each normal sample, calculate the median and absolute median difference of the normal sample library; 730. Determine the dynamic threshold based on the weighted sum of the median and the absolute median difference.

[0113] Traditional fault diagnosis methods typically use pre-set fixed thresholds as judgment criteria, such as fixed temperature difference thresholds or temperature change rate thresholds. However, this application's embodiments consider that using fixed thresholds is difficult to adapt to normal fluctuations under different operating conditions—a threshold that is too loose can lead to missed alarms, while a threshold that is too tight can lead to false alarms. In view of this, this application proposes a method for adaptively determining dynamic thresholds based on a normal sample library. This method calculates the corresponding normal score for each normal sample feature in the normal sample library in the same way, and determines the dynamic threshold based on the statistical distribution characteristics of these normal scores.

[0114] Specifically, for each normal sample feature in the normal sample library, it is treated as a test sample. Following the same method as determining the abnormal score in step 260, the similarity between this normal sample feature and other normal sample features in the normal sample library is calculated to obtain the normal score corresponding to that normal sample feature. All normal scores can constitute a normal score set. .

[0115] The median is the middle value when all scores in the normal range are arranged in ascending order. The absolute median deviation is the median of the absolute deviations of each normal score from the median, and its calculation formula is:

[0116] med is the median of the set of normal scores. For the i-th normal score, Let be the absolute deviation of the i-th normal score relative to the median. This indicates the operation of taking the median.

[0117] The absolute median is a robust measure of dispersion. Compared to the standard deviation, it is less sensitive to outliers or extreme values ​​and can more stably reflect the dispersion of the normal score distribution.

[0118] The calculated median and absolute median difference are linearly combined according to preset weighting coefficients to obtain the dynamic threshold:

[0119] in, This is a preset weighting coefficient (also known as a threshold coefficient). This dynamic threshold uses the median as the reference center and the absolute median difference as a measure of dispersion. The weighting coefficient controls the degree of deviation of the threshold from the center. When... When the value is large, only samples with a large deviation are judged as abnormal; when When the value is small, samples with smaller deviations are judged as anomalies. Preferably, the weighting coefficient... The value of this factor is a positive number between 2 and 3, for example, it can be set to 3.0. This weighting factor can be adaptively adjusted according to different tolerance levels for false alarm rate and false negative rate in actual application scenarios.

[0120] The dynamic threshold determined by the above method is not a fixed constant, but is adaptively calculated based on the distribution of the normal sample library itself. When the normal sample library is updated (e.g., new normal samples are added), the median and absolute median difference are updated accordingly, and the dynamic threshold is also adaptively adjusted accordingly. Therefore, it can always match the statistical distribution of the current normal sample library and adapt to the normal operating fluctuations of the liquid chiller under different operating conditions.

[0121] Therefore, this application also provides a normal sample update mechanism. Please refer to... Figure 8 , Figure 8 This is a flowchart of a sample update method for an anomaly identification method of a liquid chiller provided in an embodiment of this application. Specifically, after step 730, it further includes: 810. Obtain the abnormal scores within a preset continuous time range; 820. If the probability that the abnormal score within a continuous time range is not greater than the dynamic threshold is less than the preset fluctuation threshold, then the combined feature will be determined as a normal sample feature and added to the normal sample library. 830. Update dynamic threshold.

[0122] It is understandable that during long-term operation, the operating conditions of a liquid chiller may change slowly, or the operating mode may switch due to control strategy adjustments or equipment maintenance, leading to changes in the data distribution under normal operating conditions. This application's embodiment automatically identifies whether the liquid chiller is continuously operating in a normal state by statistically monitoring abnormal scores within a continuous time window. It adds the characteristics of the normal window that meet the conditions to the normal sample library and updates the dynamic threshold, achieving online iteration of the normal sample library. This allows the system to adapt to changes in operating conditions and control strategy switching during long-term operation, effectively suppressing false alarms.

[0123] Specifically, the preset continuous time range can be a historical period ending at the current time, such as the past 24 hours. Obtain the outlier scores corresponding to all time windows within this continuous time range. Calculate the proportion of windows with outlier scores not exceeding the current dynamic threshold within the continuous time range out of the total number of windows. If this proportion is lower than a preset fluctuation threshold (e.g., 5%), it indicates that the liquid chiller has been operating normally within this continuous time range without abnormal fluctuations. Therefore, the combined feature corresponding to the current time window is determined to be a normal sample, and it is added to the normal sample library as a new normal sample feature. After adding the new normal sample feature to the normal sample library, return to step 820, recalculate the median and absolute median difference based on the updated normal sample library, and update the dynamic threshold according to the weighted sum of the recalculated median and absolute median difference, ensuring that the dynamic threshold always matches the statistical distribution of the current normal sample library.

[0124] In one specific implementation, the preset continuous time range is the past 24 hours, and the preset fluctuation threshold is 5%. That is, the abnormal scores of all time windows in the past 24 hours are counted. If the proportion of windows with abnormal scores not exceeding the current dynamic threshold is less than 5%, it indicates that the liquid chiller has been operating normally in the past 24 hours and no noteworthy abnormal fluctuations have occurred. Then, the combined features of the current window are added to the normal sample library, and the dynamic threshold is recalculated.

[0125] It should be noted that the specific values ​​of 24 hours and 5% mentioned above are only examples. In actual applications, other values ​​can be set according to the operating characteristics of the liquid chiller and monitoring requirements. For example, the continuous time range can be set to 12 hours, 48 ​​hours or 72 hours, and the fluctuation threshold can be set to 3%, 10% etc. This application embodiment does not make specific limitations on this.

[0126] The updated normal sample library is invoked during subsequent anomaly identification, and the dynamic threshold is updated accordingly. Through the above online update mechanism, even if the liquid chiller undergoes control strategy switching (such as adjusting the coolant flow setpoint) or slow degradation during long-term operation (such as slight fouling of the heat exchanger leading to a decrease in heat dissipation efficiency), the normal sample library can adaptively track this controlled change or slow drift by accumulating new normal samples, enabling the system to accurately identify real abnormal states rather than normal operating condition changes. The time-series feature extraction model extracts semantic information at the "operating mode" level (such as stable plateau periods, minor fluctuations, no sudden changes, etc.) rather than specific temperature values. Therefore, even if the temperature rise rate feature may be temporarily larger at the moment of control strategy switching, other features can still reflect that the change is a controlled normal adjustment. The update mechanism can include such samples in the normal library, thereby effectively suppressing false alarms.

[0127] Corresponding to the anomaly identification method for liquid chillers provided above, this application also provides an anomaly identification device for liquid chillers. Please refer to [link to relevant documentation]. Figure 9 The anomaly detection device 900 of the liquid chiller includes: The timing acquisition module 910 is used to acquire the operating data timing and outlet water temperature timing of the liquid chiller; The feature extraction module 920 is used to input the time series of the running data into the pre-trained time series feature extraction model to obtain the running features, and to determine the temperature rise features of the liquid chiller based on the time series of the outlet water temperature. The feature combination module 930 is used to combine the operating features and the temperature rise features to obtain the combined features; The sample acquisition module 940 is used to acquire a normal sample library, which contains multiple normal sample features. The similarity calculation module 950 is used to determine the anomaly score of the combined feature based on the similarity between the combined feature and the normal sample feature; The anomaly detection module 960 is used to determine that the liquid chiller is in an abnormal state if the anomaly score is greater than a dynamic threshold, wherein the dynamic threshold is determined based on the features of multiple normal samples.

[0128] In one embodiment, the timing acquisition module 910 specifically includes: The sampling submodule 911 is used to sample the operating data and outlet water temperature of the liquid chiller based on a preset time window to obtain multiple raw data. The determination submodule 912 is used to supplement multiple original data to the first number based on linear interpolation if the number of original data is less than the first number. The standardization module 913 is used to perform standardization calculations on multiple raw data to obtain the running data time series and the outlet water temperature time series.

[0129] In one implementation, the feature extraction module 920 is specifically used to determine the temperature rise rate characteristic of the liquid chiller within a time window based on the original data corresponding to the start and end times of the time window; to determine the temperature rise amplitude characteristic of the liquid chiller within the time window based on the original data corresponding to the highest and lowest temperatures within the time window; to determine the temperature rise acceleration characteristic based on the second-order difference of multiple original data over time; to determine the relative peak deviation characteristic based on the original data corresponding to the highest and median temperatures within the time window; and to obtain the temperature rise characteristic by combining the temperature rise rate characteristic, temperature rise amplitude characteristic, temperature rise acceleration characteristic, and relative peak deviation characteristic.

[0130] In one implementation, the feature combining module 930 specifically includes: Normalization submodule 931 is used to normalize the running features to obtain the first vector; The weighted submodule 932 is used to determine the second vector based on the temperature rise characteristics and the preset first weight; Combined submodule 933 is used to combine the first vector and the second vector to obtain the combined feature.

[0131] In one implementation, the similarity calculation module 950 is specifically used to calculate the feature space distance between the combined feature and each normal sample feature; based on the feature space distance, determine a second number of normal sample features as nearest neighbor sample features; and determine the average feature space distance corresponding to the nearest neighbor sample features as the anomaly score.

[0132] In one embodiment, the anomaly detection device 900 for the liquid chiller further includes: The threshold calculation module 970 is used to calculate the similarity between each normal sample feature and other normal sample features in the normal sample library, and calculate the corresponding normal score; based on the normal scores corresponding to each normal sample feature, calculate the median and absolute median difference corresponding to the normal sample library; and determine the dynamic threshold based on the weighted sum of the median and absolute median difference.

[0133] In one embodiment, the anomaly detection device 900 for the liquid chiller further includes: The update module 980 is used to obtain abnormal scores within a preset continuous time range; if the probability that the abnormal score within the continuous time range is not greater than the dynamic threshold is less than the preset fluctuation threshold, the combined features will be determined as normal sample features and added to the normal sample library; the dynamic threshold will be updated.

[0134] In one embodiment, the anomaly detection device 900 for the liquid chiller further includes: The pre-training module 990 is used to pre-train the temporal feature extraction model. The pre-training includes: freezing the embedding layer and the third-order encoder layer near the input side of the temporal feature extraction model, training the parameters of the other encoder layers and the classification layer of the temporal feature extraction model with a first learning rate; unfreezing all layers of the temporal feature extraction model, and adjusting all parameters of the temporal feature extraction model with a second learning rate, which is less than the first learning rate.

[0135] Reference Figure 10 , Figure 10 The structural block diagram of a portion of the terminal 120 for implementing the anomaly identification method of the liquid chiller according to an embodiment of this application includes: a radio frequency (RF) circuit 2810, a memory 2815, an input unit 2830, a display unit 2840, a sensor 2850, an audio circuit 2860, a wireless fidelity (WiFi) module 2870, a processor 2880, and a power supply 2890, etc. Those skilled in the art will understand that... Figure 10 The terminal 120 structure shown does not constitute a limitation on a mobile phone or computer, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0136] The RF circuit 2810 can be used to receive and transmit signals during information transmission or calls. In particular, it receives downlink information from the base station and processes it with the processor 2880; in addition, it transmits uplink data to the base station.

[0137] The memory 2815 can be used to store software programs and modules. The processor 2880 executes various terminal functions and data processing by running the software programs and modules stored in the memory 2815.

[0138] The input unit 2830 can be used to receive input numeric or character information, and to generate key signal inputs related to the terminal's settings and function control. Specifically, the input unit 2830 may include a touch panel 2831 and other input devices 2832.

[0139] The display unit 2840 can be used to display input or provided information, as well as various menus of the terminal. The display unit 2840 may include a display panel 2841.

[0140] Audio circuitry 2860, speaker 2861, and microphone 2862 provide an audio interface.

[0141] In this embodiment, the processor 2880 included in the terminal 120 can execute the skill retrieval method of the previous embodiment.

[0142] The terminal 120 in this application embodiment includes, but is not limited to, mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle terminals, aircraft, etc. The embodiments of this invention can be applied to various scenarios, including but not limited to artificial intelligence, information processing, etc.

[0143] Figure 11This is a partial structural block diagram of a server 110 implementing the skill retrieval method of this application embodiment. The server 110 can vary significantly due to different configurations or performance, and may include one or more central processing units (CPUs) 2922 (e.g., one or more processors) and memory 2932, and one or more storage media 2930 (e.g., one or more mass storage devices) for storing application programs 2942 or data 2944. The memory 2932 and storage media 2930 can be temporary or persistent storage. The program stored in the storage media 2930 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the server 2900. Furthermore, the CPU 2922 may be configured to communicate with the storage media 2930 and execute the series of instruction operations in the storage media 2930 on the server 2900.

[0144] Server 2900 may also include one or more power supplies 2926, one or more wired or wireless network interfaces 2950, ​​one or more input / output interfaces 2958, and / or one or more operating systems 2941, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0145] The central processing unit 2922 in the server 2900 can be used to execute the anomaly identification method of the liquid chiller according to the embodiments of this application.

[0146] Please see Figure 12 This application also provides an electronic device 2, including a memory 21, a processor 20, and a computer program 22 stored in the memory 21 and executable on the processor 20. When the processor executes the computer program, it implements the abnormal identification method of the liquid cooler in the aforementioned embodiments.

[0147] This application also provides a computer-readable storage medium for storing program code for executing the anomaly identification method for the liquid chiller in the foregoing embodiments.

[0148] This application also provides a computer program product, which includes a computer program. A processor of a computer device reads and executes the computer program, causing the computer device to perform the above-described anomaly identification method for a liquid chiller.

[0149] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the above device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0150] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0151] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of external device software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0152] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For instance, the division of modules or units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between devices or units through some interfaces, and may be electrical, mechanical, or other forms.

[0153] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0154] If the integrated units described above are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing associated hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable storage medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer-readable storage device, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc. It should be noted that the contents of the aforementioned computer-readable storage media may be appropriately added to or subtracted from the contents according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable storage media may not include electrical carrier signals and telecommunication signals.

[0155] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for anomaly identification in a liquid chiller, characterized in that, include: Acquire the operating data timing and outlet water temperature timing of the liquid chiller; The time series of the operating data is input into a pre-trained time series feature extraction model to obtain operating features, and the temperature rise features of the liquid chiller are determined based on the time series of the outlet water temperature. By combining the operational characteristics and the temperature rise characteristics, a combined characteristic is obtained; Obtain a normal sample library, which contains multiple normal sample features; Based on the similarity between the combined feature and the normal sample feature, the anomaly score of the combined feature is determined; If the abnormal score is greater than the dynamic threshold, the liquid chiller is determined to be in an abnormal state, wherein the dynamic threshold is determined based on the features of the plurality of normal samples.

2. The anomaly identification method for a liquid chiller according to claim 1, characterized in that, The acquisition of the operating data timing and outlet water temperature timing of the liquid chiller includes: Based on a preset time window, the operating data and outlet water temperature of the liquid chiller are sampled to obtain multiple raw data. If the number of original data is less than the first number, then the multiple original data are supplemented to the first number based on linear interpolation. The multiple raw data are standardized to obtain the time series of the operating data and the time series of the outlet water temperature.

3. The anomaly identification method for a liquid chiller according to claim 2, characterized in that, The determination of the temperature rise characteristics of the liquid chiller based on the outlet water temperature time sequence includes: Based on the original data corresponding to the start and end times of the time window, the temperature rise rate characteristics of the liquid chiller under the time window are determined. Based on the original data corresponding to the highest and lowest temperatures within the time window, the temperature rise characteristics of the liquid chiller within the time window are determined. Based on the second-order difference of the multiple raw data with respect to time, the characteristics of temperature rise acceleration are determined; Based on the original data corresponding to the highest and median temperatures within the time window, the relative peak deviation characteristics are determined. The temperature rise characteristics are obtained by combining the temperature rise rate characteristics, temperature rise amplitude characteristics, temperature rise acceleration characteristics, and relative peak deviation characteristics.

4. The anomaly identification method for a liquid chiller according to claim 1, characterized in that, The combination of the operational characteristics and the temperature rise characteristics to obtain the combined characteristics includes: The operational features are normalized to obtain a first vector; Based on the temperature rise characteristics and the preset first weight, a second vector is determined; The first vector and the second vector are combined to obtain the combined feature.

5. The anomaly identification method for a liquid chiller according to claim 1, characterized in that, The step of determining the anomaly score of the combined feature based on the similarity between the combined feature and the normal sample feature includes: Calculate the feature space distance between the combined feature and each of the normal sample features; Based on the feature space distance, the second number of normal sample features are determined as nearest neighbor sample features; The average value of the feature space distances corresponding to the features of the nearest neighbor samples is determined as the anomaly score.

6. The anomaly identification method for a liquid chiller according to claim 1, characterized in that, Before determining that the liquid chiller is in an abnormal state if the anomaly score is greater than the dynamic threshold, the anomaly identification method for the liquid chiller further includes: Calculate the similarity between each of the normal sample features and other normal sample features in the normal sample library, and calculate the corresponding normal score; Based on the normal scores corresponding to each of the normal sample features, calculate the median and absolute median difference of the normal sample library; The dynamic threshold is determined based on the weighted sum of the median and the absolute median difference.

7. The anomaly identification method for a liquid chiller according to claim 6, characterized in that, After determining the anomaly score of the combined feature based on the similarity between the combined feature and the normal sample feature, the anomaly identification method for the liquid chiller further includes: Obtain the anomaly score within a preset continuous time range; If the probability that the abnormal score within the continuous time range is not greater than the dynamic threshold is less than a preset fluctuation threshold, then the combined feature is determined as a normal sample feature and added to the normal sample library. Update the dynamic threshold.

8. The anomaly identification method for a liquid chiller according to any one of claims 1 to 7, characterized in that, Before inputting the time series of the operating data into the pre-trained time series feature extraction model to obtain operating features, the anomaly identification method for the liquid chiller further includes... The time-series feature extraction model is pre-trained. The pre-training includes: freezing the embedding layer and the third-order encoder layer near the input side of the temporal feature extraction model, training the parameters of the other encoder layers and the classification layer of the temporal feature extraction model with a first learning rate; unfreezing all layers of the temporal feature extraction model, and adjusting all parameters of the temporal feature extraction model with a second learning rate, wherein the second learning rate is less than the first learning rate.

9. An anomaly detection device for a liquid chiller, characterized in that, include: The timing acquisition module is used to acquire the operating data timing and outlet water temperature timing of the liquid chiller; The feature extraction module is used to input the time series of the operating data into a pre-trained time series feature extraction model to obtain operating features, and to determine the temperature rise features of the liquid chiller based on the time series of the outlet water temperature. A feature combination module is used to combine the operational feature and the temperature rise feature to obtain a combined feature; The sample acquisition module is used to acquire a normal sample library, which contains multiple normal sample features; A similarity calculation module is used to determine the anomaly score of the combined feature based on the similarity between the combined feature and the normal sample feature; An anomaly determination module is used to determine that the liquid chiller is in an abnormal state if the anomaly score is greater than a dynamic threshold, wherein the dynamic threshold is determined based on the features of the plurality of normal samples.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 8.

Images