Method, device, medium and product for detecting abnormality of multi-connected system

Abstract

The application belongs to the technical field of intelligent household appliances, and particularly relates to an abnormality detection method, device, medium and product of a multi-connected system. The method obtains operation parameters and equipment parameters of the multi-connected system; generates a dynamic feature vector and a static condition vector of the multi-connected system based on the operation parameters and the equipment parameters; inputs the dynamic feature vector and the static condition vector into a feature reconstruction model to obtain a reconstructed feature vector corresponding to the dynamic feature vector; determines a reconstruction error of the operation parameters based on the reconstructed feature vector and the dynamic feature vector, and determines that the multi-connected system is abnormal in the case that the reconstruction error exceeds a preset error threshold. The method constructs a feature reconstruction model through unsupervised learning, realizes fault detection and accurate diagnosis of the multi-connected system, effectively improves the detection capability of early faults, realizes high-precision detection of the abnormality of the multi-connected system, and reduces the possibility of false positives and false negatives.

Description

Technical Field

[0001] This application belongs to the field of smart home appliance technology, specifically relating to an anomaly detection method, device, medium, and product for a multi-split air conditioning system. Background Technology

[0002] Multi-split air conditioning systems are widely used in industrial plants, commercial complexes (such as shopping malls and office buildings), large residential communities, and data centers. Their core function is to achieve efficient regulation of temperature, humidity, and air quality in different areas through the coordinated control of the outdoor unit and multiple indoor units.

[0003] In actual operation and maintenance, the complexity of multi-split systems leads to highly coupled operating parameters. Once equipment failure occurs (such as compressor wear, refrigerant leakage, sensor malfunction, etc.), it can easily cause a sharp drop in the energy efficiency of the multi-split system, equipment damage, or even safety accidents.

[0004] However, existing fault detection methods for multi-unit air conditioning systems rely on labeled fault data. Due to the scarcity of fault samples in actual operation and maintenance, the model training is insufficient, which significantly reduces the accuracy of fault detection results for multi-unit air conditioning systems. Summary of the Invention

[0005] This application provides an anomaly detection method, device, medium, and product for multi-split air conditioning systems, which solves the technical problem that existing fault detection methods for multi-split air conditioning systems rely on labeled fault data. Due to the scarcity of fault samples in actual operation and maintenance, the model training is insufficient, which significantly reduces the accuracy of fault detection results for multi-split air conditioning systems.

[0006] In a first aspect, this application provides an anomaly detection method for a multi-unit air conditioning system, comprising:

[0007] Obtain the operating parameters and equipment parameters of the multi-split air conditioning system;

[0008] Based on operating parameters and equipment parameters, dynamic feature vectors and static condition vectors of the multi-split air conditioning system are generated. The dynamic feature vectors are used to reflect the operating status of the multi-split air conditioning system under different feature dimensions, while the static condition vectors are used to reflect the inherent attributes of the multi-split air conditioning system under different feature dimensions.

[0009] The dynamic feature vector and static condition vector are input into the feature reconstruction model to obtain the reconstructed feature vector corresponding to the dynamic feature vector. The feature reconstruction model is used to learn the feature distribution law of the operating parameters corresponding to different equipment parameters in a multi-unit system under healthy operating conditions.

[0010] Based on the reconstructed feature vector and the dynamic feature vector, the reconstruction error of the operating parameters is determined. If the reconstruction error exceeds the preset error threshold, it is determined that there is an anomaly in the multi-unit system. The reconstruction error is used to quantify the degree of deviation between the dynamic feature vector and the reconstructed feature vector.

[0011] In one possible implementation, corresponding dynamic feature vectors and static condition vectors are generated based on operating parameters and device parameters;

[0012] The operating parameters are preprocessed and then standardized to obtain the corresponding dynamic feature vectors.

[0013] The equipment parameters are preprocessed and then normalized to obtain the corresponding static condition vector.

[0014] In one possible implementation, the feature reconstruction model includes an encoding branch, a decoding branch, and an attention branch; the encoding branch is used to extract the latent space representation of the dynamic feature vector, the decoding branch is used to reconstruct the features of the dynamic feature vector, and the attention branch is used to generate the attention weights of the dynamic feature vector.

[0015] Inputting the dynamic feature vector and static conditional vector into the feature reconstruction model yields the reconstructed feature vector corresponding to the dynamic feature vector, including:

[0016] Input the dynamic feature vector and the static condition vector into the encoding branch to obtain the latent vector after fusion encoding. The latent vector is a low-dimensional fusion feature representation vector generated after fusing multi-source high-dimensional features and performing dimensionality reduction encoding.

[0017] The latent vector is input into the attention branch to obtain the weight vector corresponding to the dynamic feature vector. The weight vector has the same number of dimensions as the dynamic feature vector.

[0018] The latent vector and static conditional vector are input into the decoding branch to obtain the reconstructed feature vector corresponding to the dynamic feature vector. The reconstructed feature vector has the same number of dimensions as the dynamic feature vector.

[0019] In one possible implementation, the reconstruction error of the running parameters is determined based on the reconstructed feature vector and the dynamic feature vector, including:

[0020] Determine the difference vector between the reconstructed feature vector and the dynamic feature vector;

[0021] The reconstruction error of the running parameters is determined based on the weight vector and the difference vector.

[0022] In one possible implementation, the method further includes:

[0023] Obtain historical operating parameters and historical equipment parameters of the multi-split air conditioning system;

[0024] Based on historical operating parameters and historical equipment parameters, generate historical dynamic feature vectors and historical static condition vectors for the multi-unit system.

[0025] A loss function based on an attention mechanism is constructed, and the autoencoder is pre-trained based on historical dynamic feature vectors and historical static condition vectors with the optimization objective of minimizing the loss function, thus obtaining a feature reconstruction model.

[0026] In one possible implementation, the method further includes:

[0027] Based on the training data of the feature reconstruction model, a benchmark knowledge base for the multi-split air conditioning system is generated. The training data is used to represent the sample set of historical normal operating conditions of the multi-split air conditioning system under healthy operating conditions. The benchmark knowledge base includes the dynamic feature statistical benchmark, attention weight benchmark distribution, and normal reconstruction error distribution of the multi-split air conditioning system under healthy operating conditions.

[0028] Among them, the dynamic feature statistical benchmark includes the benchmark mean and benchmark standard deviation for each feature dimension;

[0029] The baseline weight distribution includes the mean vector and standard deviation vector of the historical weight vectors;

[0030] The normal reconstruction error distribution includes the set distribution of historical reconstruction errors corresponding to any historical dynamic feature vector in the training data.

[0031] In one possible implementation, after determining that an anomaly exists in the multi-split system, the method further includes:

[0032] For each feature dimension, based on operating parameters and dynamic feature statistical benchmarks, the statistical deviation of the corresponding feature dimension of the multi-split system is determined;

[0033] Based on the weight vector and the weight baseline distribution, the weight deviation of the corresponding feature dimension of the multi-unit system is determined;

[0034] Based on statistical bias and weight bias, the anomaly level of the corresponding feature dimension of the multi-split air conditioning system is determined. The anomaly level is used to quantify the degree of anomaly of different feature dimensions of the multi-split air conditioning system.

[0035] Secondly, this application provides an anomaly detection device for a multi-unit air conditioning system, comprising:

[0036] The acquisition module is used to acquire the operating parameters and equipment parameters of the multi-unit system.

[0037] The processing module is used to generate dynamic feature vectors and static condition vectors for the multi-split air conditioning system based on operating parameters and equipment parameters. The dynamic feature vectors are used to reflect the operating status of the multi-split air conditioning system under different feature dimensions, while the static condition vectors are used to reflect the inherent attributes of the multi-split air conditioning system under different feature dimensions.

[0038] The processing module is also used to input dynamic feature vectors and static condition vectors into the feature reconstruction model to obtain the reconstructed feature vectors corresponding to the dynamic feature vectors. The feature reconstruction model is used to learn the feature distribution law of the operating parameters corresponding to different equipment parameters in the multi-unit system under healthy operating conditions.

[0039] The processing module is also used to determine the reconstruction error of the operating parameters based on the reconstructed feature vector and the dynamic feature vector, and to determine that there is an anomaly in the multi-unit system when the reconstruction error exceeds the preset error threshold. The reconstruction error is used to quantify the degree of deviation between the dynamic feature vector and the reconstructed feature vector.

[0040] In one possible implementation, the processing module is further configured to preprocess the operating parameters and standardize the preprocessed operating parameters to obtain the corresponding dynamic feature vector.

[0041] The processing module is also used to preprocess the device parameters and normalize the preprocessed device parameters to obtain the corresponding static condition vector.

[0042] In one possible implementation, the processing module is further configured to input the dynamic feature vector and the static condition vector into the encoding branch to obtain the fused encoded latent vector, wherein the latent vector is a low-dimensional fused feature representation vector generated after fusing and dimensionality-reducing encoding of multi-source high-dimensional features.

[0043] The processing module is also used to input the latent vector into the attention branch to obtain the weight vector corresponding to the dynamic feature vector. The weight vector has the same number of dimensions as the dynamic feature vector.

[0044] The processing module is also used to input the latent vector and static conditional vector into the decoding branch to obtain the reconstructed feature vector corresponding to the dynamic feature vector. The reconstructed feature vector has the same number of dimensions as the dynamic feature vector.

[0045] In one possible implementation, the processing module is further configured to determine the difference vector between the reconstructed feature vector and the dynamic feature vector.

[0046] The processing module is also used to determine the reconstruction error of the running parameters based on the weight vector and the difference vector.

[0047] In one possible implementation, the acquisition module is also used to acquire historical operating parameters and historical equipment parameters of the multi-unit system.

[0048] The processing module is also used to generate historical dynamic feature vectors and historical static condition vectors of the multi-unit system based on historical operating parameters and historical equipment parameters.

[0049] The processing module is also used to construct a loss function based on an attention mechanism, and to pre-train the autoencoder based on historical dynamic feature vectors and historical static condition vectors, with minimizing the loss function as the optimization objective, to obtain a feature reconstruction model.

[0050] In one possible implementation, the processing module is further configured to generate a benchmark knowledge base for the multi-unit system based on the training data of the feature reconstruction model. The training data is used to represent a sample set of historical normal operating conditions of the multi-unit system under healthy operating conditions. The benchmark knowledge base includes dynamic feature statistical benchmarks, attention weight benchmark distributions, and normal reconstruction error distributions of the multi-unit system under healthy operating conditions.

[0051] Among them, the dynamic feature statistical benchmark includes the benchmark mean and benchmark standard deviation for each feature dimension;

[0052] The baseline weight distribution includes the mean vector and standard deviation vector of the historical weight vectors;

[0053] The normal reconstruction error distribution includes the set distribution of historical reconstruction errors corresponding to any historical dynamic feature vector in the training data.

[0054] In one possible implementation, the processing module is further configured to determine the statistical deviation of the corresponding feature dimension of the multi-unit system for each feature dimension, based on the operating parameters and dynamic feature statistical benchmarks.

[0055] The processing module is also used to determine the weight deviation of the corresponding feature dimensions of the multi-unit system based on the weight vector and the weight benchmark distribution.

[0056] The processing module is also used to determine the anomaly level of the corresponding feature dimension of the multi-split system based on statistical bias and weight bias. The anomaly level is used to quantify the degree of anomaly of different feature dimensions of the multi-split system.

[0057] Thirdly, this application provides an electronic device, comprising:

[0058] The processor, and the memory that is in communication with the processor;

[0059] The memory stores instructions that the computer executes;

[0060] The processor executes computer execution instructions stored in memory to implement the anomaly detection method for a multi-connected system as described in the first aspect and various possible implementations of the first aspect.

[0061] Fourthly, this application provides a computer storage medium storing computer execution instructions, which are executed by a processor to implement an anomaly detection method for a multi-connected system as described in the first aspect and various possible implementations of the first aspect.

[0062] Fifthly, this application provides a program product, including a computer program, which, when executed by a processor, implements the anomaly detection method for a multi-connected system as described in the first aspect and various possible implementations of the first aspect.

[0063] The anomaly detection method for multi-split air conditioning systems provided in this application involves: acquiring the operating parameters and equipment parameters of the multi-split air conditioning system; preprocessing the operating parameters and standardizing them to obtain corresponding dynamic feature vectors; preprocessing the equipment parameters and normalizing them to obtain corresponding static condition vectors; inputting the dynamic feature vectors and static condition vectors into the encoding branch to obtain fused encoded latent vectors; inputting the latent vectors into the attention branch to obtain weight vectors corresponding to the dynamic feature vectors; inputting the latent vectors and static condition vectors into the decoding branch to obtain reconstructed feature vectors corresponding to the dynamic feature vectors; determining the difference vector between the reconstructed feature vectors and the dynamic feature vectors; and determining the reconstruction error of the operating parameters based on the weight vectors and the difference vectors; and determining that the multi-split air conditioning system has an anomaly if the reconstruction error exceeds a preset error threshold. This method constructs a healthy baseline model through unsupervised learning, combines dynamic feature importance allocation through an attention mechanism with a multi-dimensional anomaly evaluation system, and achieves fault detection and accurate diagnosis of multi-split air conditioning systems. This significantly reduces data acquisition costs, effectively improves the ability to detect early faults, achieves high-precision detection of anomalies in multi-split air conditioning systems, and reduces the possibility of false alarms and missed alarms. Attached Figure Description

[0064] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0065] Figure 1 This is a flowchart illustrating the anomaly detection method for the multi-split air conditioning system provided in this application. Figure 1 ;

[0066] Figure 2 This is a flowchart illustrating the anomaly detection method for the multi-split air conditioning system provided in this application. Figure 2 ;

[0067] Figure 3 This is a schematic diagram of the anomaly detection device for the multi-unit air conditioning system provided in this application;

[0068] Figure 4 This is a schematic diagram of the structure of the electronic device provided in this application.

[0069] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0071] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented, for example, in orders other than those illustrated or described herein.

[0072] In this application, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0073] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with relevant laws, regulations and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0074] First, the terms used in this application will be explained.

[0075] SVM: Support Vector Machine, or SVM for short, is a supervised machine learning algorithm based on statistical learning theory. It achieves sample classification and regression fitting by constructing an optimal classification hyperplane. It is suitable for pattern recognition tasks in high-dimensional feature spaces and can be used as an auxiliary model for feature classification and anomaly detection.

[0076] Random Forest: A supervised machine learning model based on ensemble learning, consisting of multiple independent decision trees. It outputs the final classification or regression result through a voting mechanism. It has the characteristics of being resistant to overfitting and has strong robustness. It can be used for feature selection and parameter classification.

[0077] Deep neural networks are neural network structures containing multiple layers of nonlinear transformation hidden layers. They extract high-order abstract features from data layer by layer to complete complex feature learning, fitting, and generation tasks.

[0078] Variational Autoencoder (VAE) is a deep generative model based on variational inference. It consists of an encoder and a decoder and achieves feature encoding and data reconstruction by learning the latent variable probability distribution of the input data.

[0079] Conditional Variational Autoencoder (CVAE) is an improved deep generative model that introduces conditional variables into VAE. By embedding conditional variables and original input data together in the encoding process, the probability distribution of latent variables is constrained, enabling the decoder to generate target data that meets expectations based on specified conditions, thus achieving directional controllability of the generated results.

[0080] Conditional Attention Variational Autoencoder (CAVAE) is a generative model that integrates CVAE and attention mechanisms. Its core is to guide generation through conditional variables and use attention to accurately capture key features of the input and conditions, thereby improving the controllability of generation and the quality of representation.

[0081] Z-score standardization: A feature standardization method based on mean and standard deviation, which transforms the original data into standardized data with zero mean and unit variance, eliminating the differences in the dimensionality and numerical scale of different feature dimensions, and achieving feature distribution normalization.

[0082] LeakyReLU activation function: An improved nonlinear activation function that modifies the linear unit. It preserves small gradients when the input is negative, solves the neuron inactivation problem of the traditional ReLU activation function, and ensures the continuity of gradients during backpropagation in deep neural networks.

[0083] Mean Squared Error (MSE) is a loss metric that measures the deviation between predicted and actual values. It quantifies the model fitting accuracy by calculating the average of the squared errors and is widely used in regression tasks and reconstruction loss calculation scenarios.

[0084] EVT: Extreme Value Theory, is a branch of statistics that studies the extreme tail behavior of random variables and the patterns of rare extreme events. It focuses only on extreme observations such as maxima and minima.

[0085] POT: Peaks Over Threshold (PCT) is a core method in EVT used for modeling and predicting extreme events. It characterizes the tail behavior of data distribution by fitting a generalized Pareto distribution to peak values ​​(over-threshold values) that exceed a set threshold. It is suitable for analyzing short-sequence extreme events without obvious seasonality.

[0086] GPD: Generalized Pareto Distribution, abbreviated as GPD, is the core distribution used in EVT for POT. It can flexibly fit different tail characteristics such as heavy tail, exponential tail, and light tail, and is used for calculating extreme quantiles, return periods, and extreme risk values.

[0087] Multi-split air conditioning systems are widely used in industrial plants, commercial complexes (such as shopping malls and office buildings), large residential communities, and data centers. Their core function is to achieve efficient regulation of temperature, humidity, and air quality in different areas through the coordinated control of outdoor units and multiple indoor units.

[0088] This multi-split system typically includes a complex thermodynamic cycle loop (such as compressor, condenser, evaporator, electronic expansion valve, etc.), a multi-level sensor network (real-time acquisition of parameters such as temperature, pressure, current, frequency, etc.), and dynamic operating condition adaptability (cooling / heating mode switching, load changes, etc.).

[0089] In existing technologies, the complexity of multi-split air conditioning systems leads to highly coupled operating parameters. For example, parameters such as compressor discharge temperature, condensing pressure, electronic expansion valve opening, and indoor unit coil temperature all affect system stability. In actual operation and maintenance, equipment failures (such as compressor wear, refrigerant leakage, sensor malfunction, etc.) may lead to a sharp drop in energy efficiency, equipment damage, or even safety accidents.

[0090] Existing fault detection methods for multi-split air conditioning systems include: expert systems based on rules or physical models and fault classification methods based on supervised learning. Among them, expert systems based on rules or physical models perform fault diagnosis by manually setting thresholds or logical rules (such as "compressor current exceeding X indicates a fault"); and fault classification methods based on supervised learning collect labeled fault data (such as "compressor fault", "refrigerant leakage", etc.) and train classification models (such as SVM, random forest, deep neural networks).

[0091] However, rule-based or physical model-based expert systems rely on the experience of domain experts, have high rule base maintenance costs, and are difficult to adapt to the complex operating conditions of multi-unit systems under different operating conditions (such as seasonal changes and load fluctuations), resulting in poor flexibility and high false alarm rates. In actual operation and maintenance, fault classification methods based on supervised learning have scarce fault samples (the fault occurrence rate is usually less than 0.1%), and the labeling requires professional personnel, resulting in high data costs and poor model generalization ability.

[0092] Therefore, existing fault detection methods rely on human experience or limited fault samples, making it difficult to meet the requirements of real-time performance, accuracy, and interpretability.

[0093] Furthermore, the dynamic fluctuation range of system operating parameters (such as temperature, pressure, current, etc.) is large, and the sensitivity of different parameters to faults varies significantly. For example, abnormal current may be an early signal of compressor failure, while fluctuations in ambient temperature may be misjudged as abnormal. In the absence of sufficient fault samples, existing technologies struggle to build diagnostic models with high generalization capabilities and cannot automatically distinguish between key features and redundant parameters, resulting in high false alarm rates and inaccurate fault location, posing a significant challenge to maintenance personnel.

[0094] To address the aforementioned issues, this application provides an anomaly detection method for multi-split air conditioning systems. It trains a Virtual Application Engine (VAE) using massive amounts of normal data to learn the joint probability distribution of the system's operating parameters. An attention mechanism is then used to quantify the sensitivity of each parameter to the system state, thereby constructing a digital model of the baseline characteristics of the operating parameters under normal operating conditions—a health baseline model. This health baseline model is used to detect faults in the real-time operating status of the multi-split system. Through multi-dimensional evaluation of reconstruction error, statistical bias, and attention weight deviation, accurate anomaly identification and automatic location of key parameters are achieved. This method constructs a health baseline model through unsupervised learning, combines dynamic feature importance allocation using an attention mechanism with a multi-dimensional anomaly evaluation system, and achieves accurate fault detection and diagnosis of multi-split systems. It significantly reduces data acquisition costs, effectively improves the ability to detect early faults, achieves high-precision anomaly detection in multi-split systems, and reduces the possibility of false alarms and missed alarms.

[0095] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0096] Figure 1 A flowchart illustrating the anomaly detection method for a multi-unit air conditioning system provided in this application embodiment. Figure 1The executing entity in this embodiment can be, for example, the control center of a multi-unit system. Figure 1 As shown, the anomaly detection method for a multi-unit air conditioning system provided in this embodiment includes:

[0097] S101: Obtain the operating parameters and equipment parameters of the multi-split air conditioning system.

[0098] Understandably, operating parameters refer to the dynamic monitoring data collected in real time during the operation of the multi-split air conditioning system, which is the direct basis for fault diagnosis. These operating parameters include, for example, compressor discharge temperature, total system input current, and indoor unit coil temperature. Equipment parameters refer to the inherent attributes or static configuration information of the multi-split air conditioning system, which describes the inherent attributes and operating background of the equipment in the multi-split air conditioning system and helps the model understand the current state of the multi-split air conditioning system. These equipment parameters include, for example, the rated cooling / heating capacity (horsepower) of the equipment and fan specifications.

[0099] In this embodiment, after obtaining the operating parameters and equipment parameters, the control center of the multi-unit system stores them in a one-dimensional vector storage format. This one-dimensional vector storage format is a standardized data storage format with a fixed order, fixed length, pure numerical type, and one-dimensional linear structure. It is specifically used to store the real-time dynamic operating parameters and static equipment parameters of the multi-unit system. It is different from storage formats such as structured dictionaries and two-dimensional tables, and is a dedicated format for time-series data processing, model inference, and edge computing scenarios.

[0100] For example, the operating parameters specifically include: "Thermodynamic cycle parameters: indoor unit coil temperature (supply / return water or refrigerant temperature), outdoor unit coil temperature (condensing / evaporating temperature), environmental and demand parameters (indoor temperature in each area, outdoor ambient temperature), control execution parameters (opening of indoor unit electronic expansion valve, opening of outdoor unit main valve), electrical and core performance parameters (total system input current, voltage, real-time power, compressor discharge temperature, compressor operating frequency)". The operating parameters acquired in this instance are collected in a fixed order and then concatenated into a one-dimensional numerical vector to obtain the operating parameters of the multi-split air conditioning system in vector storage format, which can also be called the initial feature vector, and the corresponding vector dimension is d; the equipment parameters specifically include: "Equipment capacity parameters: rated cooling / heating capacity (horsepower) of each indoor unit, total capacity (horsepower) of the outdoor unit; fan specification parameters: rated speed range or design speed of indoor and outdoor fans". The equipment parameters acquired in this instance are collected in a fixed order and then concatenated into a one-dimensional numerical vector to obtain the equipment parameters of the multi-split air conditioning system in vector storage format, which can also be called the initial condition vector, and the corresponding vector dimension is c.

[0101] S102: Generate dynamic feature vectors and static condition vectors for the multi-unit system based on operating parameters and equipment parameters.

[0102] Among them, the dynamic feature vector is used to reflect the operating status of the multi-split system under different feature dimensions; the static condition vector is used to reflect the inherent attributes of the multi-split system under different feature dimensions.

[0103] Understandably, a dynamic feature vector is a one-dimensional numerical vector composed of standardized operating parameters, used to characterize the real-time operating status of a multi-split air conditioning system under specific operating conditions. All elements in this dynamic feature vector are dimensionless standardized values, and the dimension is consistent with the one-dimensional vector corresponding to the original operating parameters (for example, it can be 12-dimensional). It can reflect the dynamic operating characteristics of the multi-split air conditioning system under specific operating conditions (such as the real-time relative status of temperature, pressure, and frequency), including standardized values ​​of parameters such as temperature, pressure, and current.

[0104] The static condition vector is a one-dimensional numerical vector composed of normalized equipment parameters. It is used to describe the inherent properties or operating background of the system, that is, it can reflect the static operating background of the multi-split system, such as the hardware foundation, performance benchmark, etc., including equipment capacity, fan rated speed, etc.

[0105] In this embodiment, the feature dimension refers to the index and classification dimension of a single independent element in a dynamic feature vector or a static condition vector, that is, the column dimension / index dimension of a one-dimensional vector; the feature dimension corresponds one-to-one with the element in the vector that represents the single independent operation attribute / inherent attribute of the multi-unit system, so each dimension of the vector corresponds to a feature dimension.

[0106] Because the original operating parameters have different dimensions (°C, MPa, r / min) and the original equipment parameters have different numerical magnitudes (e.g., rated voltage 380V, rated cooling capacity 45kW, number of indoor units 2), it is impossible to perform cross-dimensional feature comparison and analysis directly using the original operating parameters and equipment parameters (e.g., it is impossible to directly compare the relative states of temperature 26°C and pressure 2.45MPa). Therefore, standardization / normalization is used to eliminate the influence of dimensions and magnitudes, so that each element in the vector is converted to the same scale, thereby improving the performance and stability of the model.

[0107] For example, for each dimension of the data in the running parameters of the vector storage format, Z-score standardization is performed to obtain the corresponding dynamic feature vector, and the value range of the dynamic feature vector is [-1,1]; for each dimension of the data in the device parameters of the vector storage format, normalization is performed to obtain the corresponding static condition vector, and the value range of the static condition vector is [0,1].

[0108] S103: Input the dynamic feature vector and the static condition vector into the feature reconstruction model to obtain the reconstructed feature vector corresponding to the dynamic feature vector.

[0109] Among them, the feature reconstruction model is used to learn the feature distribution pattern of the operating parameters corresponding to different equipment parameters in a multi-unit system under healthy operating conditions.

[0110] Understandably, the feature reconstruction model is a generative model based on unsupervised learning, used to learn the joint probability distribution of dynamic feature vectors under healthy operating conditions, and generate reconstructed feature vectors through the decoding process; the learning objective of the feature reconstruction model is to make the reconstructed feature vectors as close as possible to the original dynamic feature vectors, thereby capturing the parameter distribution pattern under healthy conditions; this feature reconstruction model can be, for example, trained based on CAVAE.

[0111] In this embodiment of the application, the feature reconstruction model further includes an attention module. The attention module learns and outputs a one-dimensional weight vector that is completely consistent with the dimension of the input feature vector. Each element of the one-dimensional weight vector takes a value between [0,1], representing the importance score of the corresponding feature dimension. Specifically, the closer the weight is to 1, the greater the contribution of the feature dimension to the system state representation and feature reconstruction. The closer the weight is to 0, the more redundant information / noise the feature dimension is, and the lower its contribution is.

[0112] The dynamic feature vector is concatenated with the static conditional vector and then input into the feature reconstruction model (such as a learning model trained based on CAVAE). The model extracts the latent space representation through the encoder and reconstructs the dynamic feature vector through the decoder to generate the reconstructed feature vector. The latent space representation is then input into the attention module to generate the corresponding one-dimensional weight vector.

[0113] For example, dynamic features ( After concatenating the static condition (C), a low-dimensional latent vector (z) is extracted through a fully connected layer and the LeakyReLU activation function, and the mean is output. ) and variance Using an attention module, a corresponding attention weight vector (A) is generated based on the latent vector (z) to quantify the importance of each dynamic feature to the operating state of the multi-connected system; then, the latent vector (z) is concatenated with the static condition vector (C), and the dynamic features are reconstructed through a reverse fully connected layer to obtain the reconstructed feature vector. ).

[0114] S104: Based on the reconstructed feature vector and dynamic feature vector, determine the reconstruction error of the operating parameters, and if the reconstruction error exceeds the preset error threshold, determine that there is an anomaly in the multi-unit system.

[0115] The reconstruction error is used to quantify the degree of deviation between the dynamic feature vector and the reconstructed feature vector.

[0116] Understandably, the preset error threshold is the critical value set by the feature reconstruction model to determine the reconstruction error. It is a quantitative standard to distinguish whether the operating state of the multi-unit system is normal or abnormal. When the real-time reconstruction error output by the model exceeds this value, the operating state of the multi-unit system can be determined to be abnormal. This preset error threshold can be the maximum historical reconstruction error that the multi-unit system has under normal operating conditions (healthy operating state) during the feature reconstruction model training process, which also represents the maximum reconstruction error that the model may produce under healthy conditions.

[0117] Reconstruction error is a quantitative indicator calculated by comparing the difference between the dynamic feature vector and the reconstructed feature vector. It is used to characterize the degree of deviation between the current operating state and the healthy baseline. For example, it can be the mean squared error (MSE) based on attention weights.

[0118] For example, the reconfiguration error of the multi-unit system is calculated using the following formula 1, as detailed below:

[0119]

[0120] in, For real-time weighted reconstruction error; The first dynamic feature vector Each feature dimension The attention weight vector corresponding to the current sampling time is the th One element, This represents the dynamic feature vector at the current sampling time. This is the real-time reconstructed feature vector output by the feature reconstruction model.

[0121] If the preset error threshold is ,and Greater than This indicates that the operating status of the multi-unit system at the current sampling time deviates from the characteristic distribution of normal operating conditions, and there is a possibility of abnormal operation or equipment failure.

[0122] In this embodiment, the operating parameters and equipment parameters acquired by the multi-split air conditioning system have unique sampling timestamps, meaning that these operating parameters and equipment parameters represent the instantaneous operating state and inherent configuration information of the multi-split air conditioning system at any sampling time. In other words, in the process of determining the reconstruction error of the multi-split air conditioning system, the data used at that time is time-sensitive and can only represent the instantaneous operating state and inherent configuration information of the multi-split air conditioning system at the corresponding sampling time. Furthermore, the anomaly detection processes at different sampling times are independent of each other.

[0123] The anomaly detection method for multi-split air conditioning systems provided in this embodiment acquires the operating parameters and equipment parameters of the multi-split air conditioning system; based on the operating parameters and equipment parameters, it generates dynamic feature vectors and static condition vectors for the multi-split air conditioning system; it inputs the dynamic feature vectors and static condition vectors into a feature reconstruction model to obtain the reconstructed feature vectors corresponding to the dynamic feature vectors; based on the reconstructed feature vectors and dynamic feature vectors, it determines the reconstruction error of the operating parameters, and if the reconstruction error exceeds a preset error threshold, it determines that there is an anomaly in the multi-split air conditioning system. This method constructs a feature reconstruction model through unsupervised learning, realizing fault detection and accurate diagnosis of multi-split air conditioning systems, effectively improving the detection capability of early faults, achieving high-precision detection of anomalies in multi-split air conditioning systems, and reducing the possibility of false alarms and missed alarms.

[0124] Figure 2 A flowchart illustrating the anomaly detection method for a multi-unit air conditioning system provided in this application embodiment. Figure 2 .like Figure 2 As shown, in this embodiment... Figure 1 Based on the embodiments, the anomaly detection method for multi-split air conditioning systems is described in detail. The anomaly detection method for multi-split air conditioning systems shown in this embodiment includes:

[0125] S201: Obtain the operating parameters and equipment parameters of the multi-split air conditioning system.

[0126] Step S201 is similar to step S101 above, and will not be repeated here.

[0127] S202: Preprocess the operating parameters and standardize the preprocessed operating parameters to obtain the corresponding dynamic feature vector.

[0128] S203: Preprocess the equipment parameters and normalize the preprocessed equipment parameters to obtain the corresponding static condition vector.

[0129] In this embodiment of the application, the preprocessing of operating parameters and equipment parameters involves cleaning, aligning, and feature engineering the data acquired at a single sampling moment. The goal of the preprocessing is to remove noise and correct parameter deviations in the data acquired at the current sampling moment, and to convert the original data of a single sampling moment with defects into data that can be standardized / normalized.

[0130] Data cleaning addresses outliers, missing values, invalid values, and noisy data in the raw data at a single sampling time. Data alignment corrects for the instantaneous bias in data acquired from different feature dimensions at the same sampling time, as the raw data originates from multiple different sensors and acquisition modules. This correction ensures the complete operational status of the multi-unit system at that sampling time. Feature engineering extracts effective features from the cleaned and aligned data that characterize the operational status of the multi-unit system at that sampling time, reflect the instantaneous characteristics of parameters, and have discriminative power. This simplifies the data dimensions and improves the efficiency of anomaly detection.

[0131] For example, the running parameters are cleaned, aligned, and feature-engineered, and the preprocessed running parameters are standardized. Specifically, for each feature dimension of the running parameters in the preprocessed vector storage format, the mean and standard deviation of each feature dimension obtained during model training on the training set are used to calculate the difference between the data corresponding to each feature dimension and the mean, and the ratio of the corresponding difference to the standard deviation of the corresponding feature dimension is calculated to complete the Z-score standardization of each feature dimension, resulting in a standardized dynamic feature vector, and the value range of the dynamic feature vector is [-1, 1]. The device parameters are cleaned, aligned, and feature-engineered, and the preprocessed device parameters are normalized. Specifically, for each dimension of the device parameters in the preprocessed vector storage format, the device parameter is normalized by dividing it by the maximum nominal value of the system, resulting in a corresponding static condition vector, and the value range of the static condition vector is [0, 1].

[0132] S204: Input the dynamic feature vector and the static condition vector into the encoding branch to obtain the fused encoded latent vector.

[0133] S205: Input the latent vector into the attention branch to obtain the weight vector corresponding to the dynamic feature vector.

[0134] S206: Input the latent vector and static conditional vector into the decoding branch to obtain the reconstructed feature vector corresponding to the dynamic feature vector.

[0135] Among them, the latent vector is a low-dimensional fused feature representation vector generated after fusing and dimensionality-reducing high-dimensional features from multiple sources; the weight vector has the same number of dimensions as the dynamic feature vector, and the reconstructed feature vector has the same number of dimensions as the dynamic feature vector.

[0136] In this embodiment, the feature reconstruction model includes an encoding branch, a decoding branch, and an attention branch; the encoding branch is used to extract the latent space representation of the dynamic feature vector, the decoding branch is used to reconstruct the dynamic feature vector, and the attention branch is used to generate the attention weights of the dynamic feature vector.

[0137] Specifically, the encoding branch is a neural network module used to map the input data into a low-dimensional latent space representation. For example, it can be a multi-layer network composed of fully connected layers and the LeakyReLU activation function. The latent vector is a low-dimensional feature vector output by the encoding branch, used to represent the core features of the input data, such as a vector containing key information such as compressor status and condensing pressure. The attention branch is a module used to generate dynamic feature weights. For example, it can be a network composed of fully connected layers and Softmax normalization, which can output the importance weights of each dynamic feature. The decoding branch is a neural network module used to reconstruct the latent space representation and static conditional vector into a dynamic feature vector. For example, it can be a multi-layer network composed of fully connected layers and the LeakyReLU activation function.

[0138] The encoding branch compresses the concatenated vector of dynamic feature vector and static condition vector into a latent vector to extract the core features of the operating state of the multi-connected system at the corresponding sampling time. The attention branch generates an attention weight vector based on the latent space representation to quantify the importance of each dynamic feature to the operating state of the multi-connected system. Finally, the decoding branch reconstructs the concatenated vector of latent vector and static condition vector into a dynamic feature vector, generating a reconstructed feature vector. In this way, through the encoding-attention-decoding structure, unsupervised modeling of the health state of the multi-connected system is achieved.

[0139] For example, the input to the encoding branch is a standardized dynamic feature. (Dimension d) and conditional features The concatenated vector of dimension c has a dimension of (d+c). This encoding branch is the encoder in CAVAE. The encoder network consists of an input layer, multiple hidden layers, and an output layer. The input layer receives a concatenated vector of dimension (d+c). The hidden layers are multi-layer networks consisting of three fully connected layers and a LeakyReLU activation function. The first layer is a fully connected layer with an output dimension of 128, followed by the LeakyReLU activation function to accelerate model training. The second layer is a fully connected layer with an output dimension of 64, followed by the LeakyReLU activation function. The third layer is a fully connected layer with an output dimension of 32, followed by the LeakyReLU activation function. The formula for LeakyReLU is... ,in It is the input of the function (i.e., the output value of the previous layer of neurons). It is a preset fixed slope coefficient that takes values ​​in the (0,1) interval, usually set to a small value, such as 0.01; the output layer takes the 32-dimensional features output by the above hidden layer as input, and outputs the mean vector through two parallel, fully connected sub-layers without activation functions. And the dimension is (e.g., 16), and the log-variance vector The dimension is also Then, through reparameterization from the distribution Mid-sampling yields latent variables (dimension) ), which is also known as implicit vector.

[0140] The attention branch takes the high-level features or latent variable z from the encoder output as input. This attention branch consists of a fully connected layer and a softmax layer, and outputs a dynamic feature vector. Same attention weight vector And satisfy ; The magnitude of the value represents the model's ability to reconstruct the feature in the current context ( ) and state ( Under the following circumstances, it is believed that the first The importance of each dynamic feature in accurately reconstructing the overall state.

[0141] This decoding branch is also the decoder in CAVAE, which uses the sampled latent variables. and static condition vector As a joint input, its dimension is (d+c); its network structure is as follows: Input layer receives... ( (Wei) and A concatenated vector of (c-dimensional) dimensions, with dimensions ( +c); The first hidden layer is a fully connected layer, input ( The first layer has 32-dimensional input and 32-dimensional output, followed by a Leaky ReLU activation function; the second layer is a fully connected layer with 32-dimensional input and 64-dimensional output, followed by a Leaky ReLU activation function; the third layer is a fully connected layer with 64-dimensional input and 128-dimensional output, followed by a Leaky ReLU activation function; the output layer is a fully connected layer with 128-dimensional input and d-dimensional output, using a linear activation function to generate reconstructed dynamic features. .

[0142] Understandably, the joint modeling of dynamic features and static conditions is achieved through the collaborative processing of the encoder, attention branch, and decoder. Specifically, the encoder extracts low-dimensional latent space representations to reduce redundant information; the attention branch dynamically allocates feature weights to improve the reconstruction accuracy of key parameters; and the decoder, combined with static condition vectors, enhances the model's adaptability to the inherent attributes of the device, enabling the feature reconstruction model to learn the parameter distribution under healthy conditions more accurately without the need for fault labels, thereby improving the robustness and universality of anomaly detection.

[0143] In one possible implementation, historical operating parameters and historical equipment parameters of the multi-unit system are obtained; based on the historical operating parameters and historical equipment parameters, historical dynamic feature vectors and historical static condition vectors of the multi-unit system are generated; a loss function based on an attention mechanism is constructed, and based on the historical dynamic feature vectors and historical static condition vectors, the autoencoder is pre-trained with minimizing the loss function as the optimization objective to obtain a feature reconstruction model.

[0144] Understandably, the loss function is a mathematical function that balances the reconstruction accuracy of the corresponding model with the parameter regularization constraints. It is a core indicator guiding model training and optimization in machine learning / deep learning. The essence of model training is to continuously adjust the network parameters through backpropagation to minimize the output value of the loss function, thereby making the model's prediction results as close as possible to the true values. Therefore, the training process of the feature reconstruction model aims to minimize the loss function.

[0145] In this embodiment, for the historical operating data of the multi-unit system under normal operating conditions (healthy operating state), the corresponding historical operating parameters and historical equipment parameters are determined, and a sample set (including historical operating parameters and historical equipment parameters) that can be used for model training is generated. By standardizing and normalizing the data in the sample set, the differences in the dimensions of different parameters are eliminated, ensuring the uniformity and comparability of the model input data. Then, the joint probability distribution of normal data is learned through variational autoencoder, and the importance of features is dynamically allocated by combining attention mechanism, so that the model prioritizes the optimization of the reconstruction accuracy of key parameters, thereby improving the sensitivity to early faults.

[0146] The decoding branch explicitly does not reconstruct the static conditional vector, indicating that the feature reconstruction model learns the conditional probability distribution. That is, "what should the normal dynamic operating parameters of a multi-unit system with a given configuration look like", which greatly enhances the feature reconstruction model's ability to generalize to different types of equipment.

[0147] For example, the sample set is used as pre-training data to pre-train CAVAE and obtain a feature reconstruction model. In other words, the feature reconstruction model uses CAVAE to learn the latent space distribution of normal data and quantifies the importance of features through the attention mechanism to ensure that the reconstruction error of key parameters (such as current and exhaust temperature) is optimized first.

[0148] In one possible implementation, the training objective of the feature reconstruction model is to minimize the loss function. The loss function As shown in Formula 2 below:

[0149]

[0150] in, The total loss of the model, The reconstruction loss is used to measure the difference between the model output and the true target. This is an adjustable regularization weight coefficient. KL divergence loss is used to measure the difference between two probability distributions.

[0151] By setting This can strengthen the regularization constraint on the potential spatial distribution, prompting the feature reconstruction model to learn a smoother and more generalized normal data manifold, thereby enhancing the sensitivity of the feature reconstruction model to abnormal states that deviate from the manifold.

[0152] In actual training, fixed values ​​can be used (e.g., ) or use a step-by-step increment from 0 - Optimize using a gradual strategy; KL divergence term It is used to constrain the distribution of the latent space to approach the standard normal distribution, thereby playing a regularization role, preventing overfitting, and ensuring the continuity and structure of the latent space, which facilitates generation and interpolation.

[0153] Among them, the attention-weighted dynamic feature reconstruction term As shown in Formula 3 below:

[0154]

[0155] in, This refers to the weighted feature reconstruction loss during model training, which is also the reconstruction error during model training. For the dynamic feature vector during model training Each feature dimension The first attention weight vector during model training One element, These are dynamic feature vectors generated during model training. This refers to the reconstructed feature vector output by CAVAE after training during the model training process.

[0156] This loss function calculates the loss only for dynamic features, with zero error for conditional features. This forces the trained feature model to focus all its capabilities on learning healthy patterns of core operating parameters; the reconstruction error of each dynamic feature is reduced by its corresponding attention weight. Weighting; during backpropagation, the model receives a signal that the reconstruction error of high-weight features has a greater impact on the total loss; therefore, the model will prioritize optimizing the reconstruction accuracy of high-weight features, while tolerating larger errors of low-weight features; in this way, the attention branch learns to assign high weights to key features (such as current and exhaust temperature) that are indispensable for describing the health of the system and should have small fluctuations, thus achieving a fully data-driven importance classification.

[0157] In one possible implementation, a benchmark knowledge base for the multi-unit system is generated based on the training data of the feature reconstruction model.

[0158] The training data represents a sample set of historical normal operating conditions of the multi-split air conditioning system under healthy operating conditions. The benchmark knowledge base includes the dynamic feature statistical benchmark, attention weight benchmark distribution, and normal reconstruction error distribution of the multi-split air conditioning system under healthy operating conditions.

[0159] The dynamic feature statistical benchmark includes the benchmark mean and benchmark standard deviation for each feature dimension; the weight benchmark distribution includes the mean vector and standard deviation vector of historical weight vectors; and the normal reconstruction error distribution includes the set distribution of historical reconstruction errors corresponding to any historical dynamic feature vector in the training data.

[0160] Understandably, the training process of the feature reconstruction model uses massive amounts of validated normal operating condition samples. (That is, the concatenated vector of historical dynamic feature vector and historical static condition vector) trains the model; after training is completed, the model parameters obtained in this training are saved, and a benchmark knowledge base corresponding to the multi-unit system is established to store the dynamic feature statistical benchmark, attention weight benchmark distribution and normal reconstruction error distribution under the healthy operation state of the multi-unit system.

[0161] In this embodiment, the dynamic feature statistical benchmark is used when standardizing each dynamic feature. and Its dynamic feature statistical benchmark is used to standardize the operating parameters, providing a unified scale for data normalization in online detection, ensuring that real-time data and training data are in the same metric space, and serving as the basis for all subsequent quantitative comparisons; the attention weight benchmark distribution is used to calculate the attention weight vector of all samples in the sample set. mean vector and standard deviation vector This serves as a reference standard for "model cognition" in online diagnosis, quantifying the deviation of real-time attention weights from the normal pattern. It is a core indicator driving subsequent intelligent and focused diagnostic processes; the normal reconstruction error distribution records all training samples. Values, forming a set Its function is to characterize the intrinsic fluctuation level of a multi-unit system under healthy conditions and to determine an adaptive and robust dynamic threshold for anomaly detection, so as to effectively distinguish between normal fluctuations and real anomalies and improve the reliability of detection.

[0162] S207: Determine the difference vector between the reconstructed feature vector and the dynamic feature vector.

[0163] S208: Determine the reconstruction error of the running parameters based on the weight vector and the difference vector.

[0164] S209: If the reconstruction error exceeds the preset error threshold, it is determined that there is an anomaly in the multi-unit system.

[0165] The difference vector refers to the element-wise difference between the dynamic feature vector and the reconstructed feature vector, which is used to characterize the degree to which the parameters deviate from the healthy baseline; the reconstruction error refers to the difference vector after being weighted by the attention weight vector, which is used to quantify the degree of abnormality of each dynamic feature.

[0166] By calculating the difference vector between the reconstructed feature vector and the dynamic feature vector, and combining it with the attention weight vector, the reconstruction error of the operating parameters is calculated. If the reconstruction error exceeds the preset error threshold, it is determined that there is an anomaly in the multi-unit system, thereby realizing the localization of the abnormal dynamic features and further achieving the accurate localization of the core dynamic features of the anomaly.

[0167] In one possible implementation, the preset error threshold can be determined based on a dynamic threshold determination using EVT, specifically: invoking the normal reconstruction error distribution established during the training phase. The POT method is used to model the problem, fit the GPD, and estimate the error value corresponding to a given extremely low exceedance probability (e.g., 0.1%), which is then used as a preset error threshold. .

[0168] In one possible implementation, after determining that an anomaly exists in the multi-split air conditioning system, for each feature dimension, based on operating parameters and dynamic feature statistical benchmarks, the statistical deviation of the corresponding feature dimension of the multi-split air conditioning system is determined; based on the weight vector and weight benchmark distribution, the weight deviation of the corresponding feature dimension of the multi-split air conditioning system is determined; and based on the statistical deviation and weight deviation, the anomaly level of the corresponding feature dimension of the multi-split air conditioning system is determined.

[0169] Among them, statistical bias refers to the degree of deviation between the original data of the dynamic feature of any feature dimension and the historical statistical benchmark; weight bias refers to the degree of deviation between the attention weight of the dynamic feature of any feature dimension and the historical benchmark distribution; anomaly level is used to quantify the degree of anomaly of different feature dimensions of the multi-unit system.

[0170] For example, the statistical bias is calculated using the following formula 4:

[0171]

[0172] in, For the first Statistical bias of each feature dimension At the current sampling time, the first The original running parameters for each feature dimension were not standardized. and The dynamic feature statistical benchmark is retrieved from the knowledge base to reflect the degree to which the original value of the feature deviates from the historical normal value benchmark.

[0173] The weighted deviation is calculated using the following formula 5:

[0174]

[0175] in, For the first Weight bias of each feature dimension At the current sampling time, the first Attention weights for each feature dimension; and The attention weight baseline distribution, which is called from the knowledge base, is used to reflect the change in the feature reconstruction model's assessment of the importance of the current feature compared to historical norms.

[0176] according to and The dynamic features of multiple dimensions are classified into levels, specifically: in or In the case of determining the corresponding number The anomaly level of the feature in this dimension is 1. At this point, there is a slight deviation between the output of the feature reconstruction model and the input dynamic features, indicating a slight anomaly. An early warning can be issued for this dimension's anomaly to alert the user. or In the case of determining the corresponding number The anomaly level of the feature dimension is 2. At this point, there is a significant deviation between the output of the feature reconstruction model and the input dynamic features, indicating a significant anomaly. An alert can be issued for this anomaly to remind the user that maintenance is needed to address it. or In the case of determining the corresponding number The anomaly level of the dimension's features is level 3. At this point, there is a serious deviation between the output of the feature reconstruction model and the input dynamic features, which means there is a serious anomaly. For the anomaly of this dimension, an emergency alarm can be issued first to remind the user that there is a fault.

[0177] In one possible implementation, the statistical bias and weight bias determined in this instance are weighted and fused to obtain a comprehensive anomaly index for each feature dimension. The features are sorted in descending order; the features ranked first are the core suspicious parameters that caused this anomaly.

[0178] Based on the identified anomaly level, as well as the reconstruction error, statistical bias, and weighting bias, an actionable diagnostic report can be generated. This report is output in a structured format and includes a comprehensive anomaly index. The proportion exceeding the threshold, by A sorted list of Top-K dynamic features of anomalies (each item includes feature name, measured value, normal range, or...). , and The system includes the anomaly level of each feature, as well as intelligent maintenance guidance suggestions that combine abnormal feature combinations with attention weighting patterns.

[0179] Understandably, by using multi-dimensional assessment of statistical bias and weighted bias, the severity of anomalies can be graded, further improving the operability of diagnosis, providing operation and maintenance personnel with intuitive guidance on fault priority, and optimizing the efficiency of maintenance decision-making.

[0180] In this embodiment of the application, the calculation formula for the reconstruction error is the same as that for the training loss. The format is the same, but the calculation object is the new sample acquired at the current sampling time, where, Traverse all dynamic features, Assign the current sample to the model Attention weights for each feature and These are the standardized input value and the model reconstruction value for the feature, respectively; the resulting reconstruction error in the form of a score integrates the feature importance assessed based on the current state and the degree of deviation before and after reconstruction.

[0181] The anomaly detection method for multi-split air conditioning systems provided in this embodiment acquires the operating parameters and equipment parameters of the multi-split air conditioning system; preprocesses the operating parameters and standardizes them to obtain corresponding dynamic feature vectors; preprocesses the equipment parameters and normalizes them to obtain corresponding static condition vectors; inputs the dynamic feature vectors and static condition vectors into the encoding branch to obtain fused encoded latent vectors; inputs the latent vectors into the attention branch to obtain weight vectors corresponding to the dynamic feature vectors; inputs the latent vectors and static condition vectors into the decoding branch to obtain reconstructed feature vectors corresponding to the dynamic feature vectors; determines the difference vector between the reconstructed feature vectors and the dynamic feature vectors; and determines the reconstruction error of the operating parameters based on the weight vectors and the difference vectors; if the reconstruction error exceeds a preset error threshold, it is determined that there is an anomaly in the multi-split air conditioning system. This method constructs a healthy baseline model through unsupervised learning, combines the dynamic feature importance allocation of the attention mechanism with a multi-dimensional anomaly evaluation system, and achieves fault detection and accurate diagnosis of multi-split air conditioning systems. It significantly reduces data acquisition costs, effectively improves the ability to detect early faults, achieves high-precision detection of anomalies in multi-split air conditioning systems, and reduces the possibility of false alarms and missed alarms.

[0182] Figure 3 This is a schematic diagram of the anomaly detection device for the multi-split air conditioning system provided in this application. Figure 3 As shown, this application provides an anomaly detection device for a multi-split air conditioning system. The anomaly detection device 300 for the multi-split air conditioning system includes:

[0183] The acquisition module 301 is used to acquire the operating parameters and equipment parameters of the multi-unit system.

[0184] The processing module 302 is used to generate dynamic feature vectors and static condition vectors of the multi-split system based on operating parameters and equipment parameters. The dynamic feature vectors are used to reflect the operating status of the multi-split system under different feature dimensions, and the static condition vectors are used to reflect the inherent attributes of the multi-split system under different feature dimensions.

[0185] The processing module 302 is also used to input the dynamic feature vector and the static condition vector into the feature reconstruction model to obtain the reconstructed feature vector corresponding to the dynamic feature vector. The feature reconstruction model is used to learn the feature distribution law of the operating parameters corresponding to different equipment parameters in the multi-unit system under healthy operating conditions.

[0186] The processing module 302 is also used to determine the reconstruction error of the operating parameters based on the reconstructed feature vector and the dynamic feature vector, and to determine that there is an anomaly in the multi-unit system when the reconstruction error exceeds a preset error threshold. The reconstruction error is used to quantify the degree of deviation between the dynamic feature vector and the reconstructed feature vector.

[0187] In one possible implementation, the processing module 302 is further configured to preprocess the operating parameters and standardize the preprocessed operating parameters to obtain the corresponding dynamic feature vector.

[0188] The processing module 302 is also used to preprocess the device parameters and normalize the preprocessed device parameters to obtain the corresponding static condition vector.

[0189] In one possible implementation, the processing module 302 is further configured to input the dynamic feature vector and the static condition vector into the encoding branch to obtain the fused encoded latent vector, wherein the latent vector is a low-dimensional fused feature representation vector generated after fusing and dimensionality-reducing encoding of multi-source high-dimensional features.

[0190] The processing module 302 is also used to input the latent vector into the attention branch to obtain the weight vector corresponding to the dynamic feature vector, and the weight vector has the same number of dimensions as the dynamic feature vector.

[0191] The processing module 302 is also used to input the latent vector and the static conditional vector into the decoding branch to obtain the reconstructed feature vector corresponding to the dynamic feature vector. The reconstructed feature vector has the same number of dimensions as the dynamic feature vector.

[0192] In one possible implementation, the processing module 302 is further configured to determine the difference vector between the reconstructed feature vector and the dynamic feature vector.

[0193] The processing module 302 is also used to determine the reconstruction error of the running parameters based on the weight vector and the difference vector.

[0194] In one possible implementation, the acquisition module 301 is also used to acquire historical operating parameters and historical equipment parameters of the multi-unit system.

[0195] The processing module 302 is also used to generate historical dynamic feature vectors and historical static condition vectors of the multi-unit system based on historical operating parameters and historical equipment parameters.

[0196] The processing module 302 is also used to construct a loss function based on the attention mechanism, and to pre-train the autoencoder based on historical dynamic feature vectors and historical static condition vectors with the optimization objective of minimizing the loss function, so as to obtain a feature reconstruction model.

[0197] In one possible implementation, the processing module 302 is further configured to generate a benchmark knowledge base for the multi-unit system based on the training data of the feature reconstruction model. The training data is used to represent a sample set of historical normal operating conditions of the multi-unit system under healthy operating conditions. The benchmark knowledge base includes dynamic feature statistical benchmarks, attention weight benchmark distributions, and normal reconstruction error distributions of the multi-unit system under healthy operating conditions.

[0198] Among them, the dynamic feature statistical benchmark includes the benchmark mean and benchmark standard deviation for each feature dimension;

[0199] The baseline weight distribution includes the mean vector and standard deviation vector of the historical weight vectors;

[0200] The normal reconstruction error distribution includes the set distribution of historical reconstruction errors corresponding to any historical dynamic feature vector in the training data.

[0201] In one possible implementation, the processing module 302 is further configured to determine the statistical deviation of the corresponding feature dimension of the multi-unit system for each feature dimension, based on the operating parameters and dynamic feature statistical benchmark.

[0202] The processing module 302 is also used to determine the weight deviation of the corresponding feature dimension of the multi-unit system based on the weight vector and the weight benchmark distribution.

[0203] The processing module 302 is also used to determine the anomaly level of the corresponding feature dimension of the multi-split system based on statistical bias and weight bias. The anomaly level is used to quantify the degree of anomaly of different feature dimensions of the multi-split system.

[0204] Figure 4 A schematic diagram of the structure of the electronic device provided in this application. Figure 4 As shown, this application provides an electronic device 400, which includes a receiver 401, a transmitter 402, a processor 403, and a memory 404.

[0205] Receiver 401 is used to receive instructions and data;

[0206] Transmitter 402 is used to send commands and data;

[0207] Memory 404 is used to store instructions executed by the computer;

[0208] Processor 403 is used to execute computer execution instructions stored in memory 404 to implement the various steps performed by the anomaly detection method for the multi-unit system in the above embodiments. For details, please refer to the relevant descriptions in the foregoing embodiments of the anomaly detection method for the multi-unit system.

[0209] Optionally, the memory 404 can be either standalone or integrated with the processor 403.

[0210] When the memory 404 is set up independently, the electronic device also includes a bus for connecting the memory 404 and the processor 403.

[0211] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement an anomaly detection method for a multi-connected system as described above in the electronic device.

[0212] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the anomaly detection method for a multi-unit system according to any of the foregoing embodiments.

[0213] It will be understood by those skilled in the art that all or some of the steps, systems, or apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0214] The technical solutions of this application have been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it is readily understood by those skilled in the art that the scope of protection of this application is obviously not limited to these specific embodiments. 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 or all of the technical features therein. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for detecting anomalies in a multi-split air conditioning system, characterized in that, include: Obtain the operating parameters and equipment parameters of the multi-unit system; Based on the operating parameters and the equipment parameters, a dynamic feature vector and a static condition vector of the multi-split air conditioning system are generated. The dynamic feature vector is used to reflect the operating status of the multi-split air conditioning system under different feature dimensions, and the static condition vector is used to reflect the inherent attributes of the multi-split air conditioning system under different feature dimensions. The dynamic feature vector and the static condition vector are input into the feature reconstruction model to obtain the reconstructed feature vector corresponding to the dynamic feature vector. The feature reconstruction model is used to learn the feature distribution law of the operating parameters corresponding to different equipment parameters of the multi-unit system under healthy operating conditions. Based on the reconstructed feature vector and the dynamic feature vector, the reconstruction error of the operating parameters is determined, and if the reconstruction error exceeds a preset error threshold, it is determined that the multi-unit system has an anomaly. The reconstruction error is used to quantify the degree of deviation between the dynamic feature vector and the reconstructed feature vector.

2. The method according to claim 1, characterized in that, Based on the operating parameters and the device parameters, corresponding dynamic feature vectors and static condition vectors are generated; The operating parameters are preprocessed and then standardized to obtain the corresponding dynamic feature vector. The device parameters are preprocessed and then normalized to obtain the corresponding static condition vector.

3. The method according to claim 2, characterized in that, The feature reconstruction model includes an encoding branch, a decoding branch, and an attention branch; the encoding branch is used to extract the latent space representation of the dynamic feature vector, the decoding branch is used to reconstruct the features of the dynamic feature vector, and the attention branch is used to generate the attention weights of the dynamic feature vector. The step of inputting the dynamic feature vector and the static condition vector into the feature reconstruction model to obtain the reconstructed feature vector corresponding to the dynamic feature vector includes: The dynamic feature vector and the static condition vector are input into the encoding branch to obtain the fused encoding latent vector. The latent vector is a low-dimensional fused feature representation vector generated by fusing multi-source high-dimensional features and performing dimensionality reduction encoding. The latent vector is input into the attention branch to obtain the weight vector corresponding to the dynamic feature vector, and the weight vector has the same number of dimensions as the dynamic feature vector. The latent vector and the static conditional vector are input into the decoding branch to obtain the reconstructed feature vector corresponding to the dynamic feature vector. The reconstructed feature vector has the same number of dimensions as the dynamic feature vector.

4. The method according to claim 3, characterized in that, The step of determining the reconstruction error of the running parameters based on the reconstructed feature vector and the dynamic feature vector includes: Determine the difference vector between the reconstructed feature vector and the dynamic feature vector; Based on the weight vector and the difference vector, the reconstruction error of the running parameters is determined.

5. The method according to claim 4, characterized in that, The method further includes: Obtain the historical operating parameters and historical equipment parameters of the multi-unit system; Based on the historical operating parameters and the historical equipment parameters, the historical dynamic feature vector and historical static condition vector of the multi-unit system are generated. An attention-based loss function is constructed, and the autoencoder is pre-trained based on the historical dynamic feature vector and the historical static condition vector, with minimizing the loss function as the optimization objective, to obtain the feature reconstruction model.

6. The method according to claim 5, characterized in that, The method further includes: Based on the training data of the feature reconstruction model, a benchmark knowledge base for the multi-unit air conditioning system is generated. The training data is used to represent a sample set of historical normal operating conditions of the multi-unit air conditioning system under healthy operating conditions. The benchmark knowledge base includes dynamic feature statistical benchmarks, attention weight benchmark distributions, and normal reconstruction error distributions of the multi-unit air conditioning system under healthy operating conditions. The dynamic feature statistical benchmark includes the benchmark mean and benchmark standard deviation for each feature dimension. The weighted benchmark distribution includes the mean vector and standard deviation vector of the historical weight vectors; The normal reconstruction error distribution includes the set distribution of historical reconstruction errors corresponding to any historical dynamic feature vector in the training data.

7. The method according to claim 6, characterized in that, After determining that the multi-split air conditioning system has an anomaly, the method further includes: For each feature dimension, based on the operating parameters and the dynamic feature statistical benchmark, the statistical deviation of the corresponding feature dimension of the multi-unit system is determined; Based on the weight vector and the weight baseline distribution, the weight deviation of the corresponding feature dimension of the multi-unit system is determined; Based on the statistical bias and the weight bias, the anomaly level of the corresponding feature dimension of the multi-unit system is determined. The anomaly level is used to quantify the degree of anomaly of different feature dimensions of the multi-unit system.

8. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-7.

Images