Building heating, ventilation and air conditioning system fault diagnosis system and method and medium

Abstract

The invention relates to the technical field of intelligent building fault diagnosis, and discloses a building heating, ventilation and air conditioning system fault diagnosis system and method and a medium. The data construction module is used for calculating Pearson's correlation coefficients among the multi-source sensors and constructing a correlation graph; the feature extraction module is used for extracting spatial features in the association graph and extracting spatial-temporal features in the spatial features; the anomaly detection module calculates the anomaly score of each sensor node in the association graph and judges potential fault points; and the fault diagnosis module generates a predicted value of the potential fault point at the next time point, calculates an error between the multi-source sensor data of the potential fault point and the predicted value of the potential fault point at the next time point, and judges an abnormal point. According to the invention, the weighted graph structure is constructed, and the graph attention network and the space-time graph neural network are utilized to realize high-precision anomaly detection and fault diagnosis under the background of multi-source data.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis of intelligent buildings, and particularly to a fault diagnosis system, method and medium for a building heating, ventilation and air conditioning system. Background Art

[0002] With the continuous improvement of industrial automation and intelligence levels, heating, ventilation and air conditioning systems are more widely used in modern buildings. However, the complexity of heating, ventilation and air conditioning systems makes them face various potential fault risks during long-term operation, such as compressor failure, insufficient cooling water flow and sensor failure. These faults not only lead to a decline in equipment performance, but also may cause unstable system operation and increase maintenance costs. Therefore, developing effective fault diagnosis methods is of great significance for ensuring the reliable operation of heating, ventilation and air conditioning systems and reducing energy consumption.

[0003] Traditional fault diagnosis methods mainly rely on experience and regular maintenance. However, this method has significant limitations and is difficult to adapt to rapidly changing operating environments and real-time monitoring requirements. In recent years, data-driven methods have gradually emerged. In particular, the application of machine learning and deep learning technologies has provided new ideas for fault diagnosis. These methods can utilize real-time data obtained from sensors to identify potential fault patterns. However, many existing data-driven methods often rely on a large amount of labeled data and are difficult to play a role in scenarios without labeled data. Summary of the Invention

[0004] The purpose of the present invention is to overcome the problems existing in the prior art that the sensor data in heating, ventilation and air conditioning systems are diverse in types, have significant spatio-temporal correlation characteristics, a large amount of data is unlabeled, and faults cannot be effectively diagnosed. The present invention provides a fault diagnosis system, method and medium for a building heating, ventilation and air conditioning system. The present invention constructs a weighted graph structure by calculating the Pearson correlation coefficient of sensor data, and uses a graph attention network and a spatio-temporal graph neural network to extract the spatial and temporal features of the data respectively. Combining with the principal component analysis method, this solution calculates the anomaly score in an unsupervised environment, so as to achieve high-precision anomaly detection and fault diagnosis in the context of multi-source data.

[0005] To achieve the above object, on the one hand, the present invention provides a fault diagnosis system for a building heating, ventilation and air conditioning system, including:

[0006] A data acquisition and processing module is used to collect multi-source sensor data from multi-source sensors at different positions and components of the heating, ventilation and air conditioning system, and preprocess the collected multi-source sensor data to obtain preprocessed multi-source sensor data;

[0007] The data construction module is used to calculate the Pearson correlation coefficient between multi-source sensors based on the preprocessed multi-source sensor data, construct a correlation coefficient matrix according to the calculated Pearson correlation coefficient between multi-source sensors, and construct an association graph using the correlation coefficient matrix;

[0008] The feature extraction module is used to extract the spatial features in the association graph using a multi-layer graph attention network, and then extract the spatio-temporal features in the spatial features using a spatio-temporal graph neural network;

[0009] The anomaly detection module is used to perform principal component analysis dimensionality reduction on the extracted spatio-temporal features using principal component analysis technology to obtain the dimensionality-reduced spatio-temporal features. Based on the dimensionality-reduced spatio-temporal features, use principal component analysis technology to calculate the anomaly score of each sensor node in the association graph, and compare the calculated anomaly score with the set anomaly score threshold to judge potential fault points;

[0010] The fault diagnosis module is used to generate the predicted value of the potential fault point at the next time point using a spatio-temporal graph neural network according to the spatio-temporal features corresponding to the potential fault point, calculate the error between the multi-source sensor data of the potential fault point and the predicted value of the potential fault point at the next time point, and compare the calculated error with the normal range to judge the anomaly point.

[0011] Preferably, the preprocessing of the collected multi-source sensor data includes standardizing and noise filtering the collected multi-source sensor data.

[0012] Preferably, the expression for calculating the Pearson correlation coefficient between multi-source sensors based on the preprocessed multi-source sensor data and constructing a correlation coefficient matrix according to the calculated Pearson correlation coefficient between multi-source sensors is:

[0013]

[0014] In the formula, r ij represents the Pearson correlation coefficient between the i-th multi-source sensor and the j-th multi-source sensor; x ti represents the preprocessed multi-source sensor data of the i-th multi-source sensor at time t; x tj represents the preprocessed multi-source sensor data of the j-th multi-source sensor at time t; represents the mean of the preprocessed multi-source sensor data of the i-th multi-source sensor during the acquisition time; represents the mean of the preprocessed multi-source sensor data of the j-th multi-source sensor during the acquisition time; N represents the length of the time series; A ij represents the element in the correlation coefficient matrix; θ represents the threshold of the Pearson correlation coefficient;

[0015] The specific operations for constructing the association graph using the correlation coefficient matrix include: connecting sensor nodes with weighted edges, where the weight value is determined by the Pearson correlation coefficient between multi-source sensors.

[0016] Preferably, the expression for extracting spatial features from the association graph using a multi-layer graph attention network is:

[0017] e ij = LeakyReLU(a T [Wh i ||Wh j

[0018]

[0019] In the formula, e ij represents the calculated attention score; a T represents the learning parameter; LeakyReLU represents the activation function; α ij is the normalization coefficient between sensor node i and sensor node j; W represents the linear transformation weight matrix; σ represents the sigmoid activation function; N(i) represents the set of neighbor sensor nodes of sensor node i; represents the updated feature of sensor node i at time t; h j represents the spatial feature vector provided by sensor node j; h i represents the spatial feature vector provided by sensor node i; K represents the number of attention heads; represents the attention coefficient of the k-th attention head; W k represents the weight matrix of the k-th attention head; e il represents the attention score between sensor node i and neighbor sensor node l of sensor node i;

[0020] The expression for extracting spatio-temporal features from spatial features using a spatio-temporal graph neural network is::

[0021]

[0022] In the formula, represents the feature vector of sensor node j at time t; y t represents the output at time t; X t-k+1 represents the input feature at time t-k+1; w k represents the convolution kernel; represents the updated feature of sensor node i at time t+1; represents the final fusion feature of sensor node i; α represents the weight coefficient controlling the fusion ratio of spatial features and temporal features; represents the spatial feature of sensor node i. ​

[0023] Preferably, principal component analysis technology is used to perform principal component analysis and dimensionality reduction on the extracted spatio-temporal features, and the expression of the spatio-temporal features after dimensionality reduction is:

[0024]

[0025] X r = X c W

[0026] In the formula, represents the sample mean; X c is the data matrix after centering processing; X r represents the spatio-temporal feature data after dimensionality reduction; m represents the number of samples; x i represents the sample data; X represents the spatio-temporal feature matrix; represents the transpose matrix of X c .

[0027] Preferably, based on the spatio-temporal features after dimensionality reduction, the calculation formula for the anomaly score of each sensor node in the correlation graph using principal component analysis technology is:

[0028]

[0029] In the formula, P represents the anomaly score; represents the sample reconstructed based on the spatio-temporal features after dimensionality reduction;

[0030] The calculated anomaly score is compared with the set anomaly score threshold. When the calculated anomaly score is higher than the set anomaly score threshold, the sensor node is a potential fault point.

[0031] Preferably, the calculation formula for the error between the multi-source sensor data of the potential fault point and the predicted value of the potential fault point at the next time point is:

[0032] e t = V a - V p

[0033] In the formula, e t represents the error between the multi-source sensor data of the potential fault point at time t and the predicted value of the potential fault point at the next time point; V a represents the multi-source sensor data of the potential fault point at time t; V p represents the predicted value of the potential fault point at the next time point;

[0034] The calculation formula for the normal range is:

[0035]

[0036] Wherein, μ e represents the mean error; σ e represents the standard deviation of the error; A represents the normal range;

[0037] Compare the calculated error with the normal range. When the calculated error exceeds the normal range, the potential fault point is the abnormal point.

[0038] The second aspect of the present invention provides a method for fault diagnosis of a building heating, ventilation and air conditioning system, including:

[0039] Collect multi-source sensor data from multi-source sensors at different positions and components of the heating, ventilation and air conditioning system, and preprocess the collected multi-source sensor data to obtain preprocessed multi-source sensor data;

[0040] Calculate the Pearson correlation coefficient between multi-source sensors according to the preprocessed multi-source sensor data, construct a correlation coefficient matrix based on the calculated Pearson correlation coefficient between multi-source sensors, and construct an association graph using the correlation coefficient matrix;

[0041] Adopt a multi-layer graph attention network to extract the spatial features in the association graph, and then adopt a spatio-temporal graph neural network to extract the spatio-temporal features in the spatial features;

[0042] Use the principal component analysis technique to perform principal component analysis and dimensionality reduction on the extracted spatio-temporal features to obtain the dimensionality-reduced spatio-temporal features. Based on the dimensionality-reduced spatio-temporal features, use the principal component analysis technique to calculate the anomaly score of each sensor node in the association graph, compare the calculated anomaly score with the set anomaly score threshold, and judge the potential fault point;

[0043] According to the spatio-temporal features corresponding to the potential fault point, use the spatio-temporal graph neural network to generate the predicted value of the potential fault point at the next time point, calculate the error between the multi-source sensor data of the potential fault point and the predicted value of the potential fault point at the next time point, and compare the calculated error with the normal range to judge the abnormal point.

[0044] Preferably, preprocessing the collected multi-source sensor data includes standardizing and noise filtering the collected multi-source sensor data.

[0045] Preferably, the expression for calculating the Pearson correlation coefficient between multi-source sensors according to the preprocessed multi-source sensor data and constructing a correlation coefficient matrix is:

[0046]

[0047] Wherein, r ijRepresents the Pearson correlation coefficient between the i-th multi-source sensor and the j-th multi-source sensor; x ti Represents the preprocessed multi-source sensor data of the i-th multi-source sensor at time t; x tj Represents the preprocessed multi-source sensor data of the j-th multi-source sensor at time t; Represents the mean of the preprocessed multi-source sensor data of the i-th multi-source sensor during the acquisition time; Represents the mean of the preprocessed multi-source sensor data of the j-th multi-source sensor during the acquisition time; N represents the length of the time series; A ij Represents an element in the correlation coefficient matrix; θ represents the threshold of the Pearson correlation coefficient;

[0048] The specific operation of constructing an association graph using the correlation coefficient matrix includes: connecting sensor nodes with weighted edges, and the weight value is determined by the Pearson correlation coefficient between multi-source sensors.

[0049] Preferably, the expression for extracting spatial features in the association graph using a multi-layer graph attention network is:

[0050] e ij =LeakyReLU(a T [Wh i ||Wh j

[0051]

[0052] In the formula, e ij Represents the calculated attention score; a T Represents the learning parameter; LeakyReLU represents the activation function; α ij Is the normalization coefficient between sensor node i and sensor node j; W represents the linear transformation weight matrix; σ represents the sigmoid activation function; N(i) represents the set of neighbor sensor nodes of sensor node i; Represents the updated feature of sensor node i at time t; h j Represents the spatial feature vector provided by sensor node j; h i Represents the spatial feature vector provided by sensor node i; K represents the number of attention heads; Represents the attention coefficient of the k-th attention head; W k Represents the weight matrix of the k-th attention head; e ij Represents the attention score between sensor node i and the neighbor sensor node l of sensor node i;

[0053] The expression for extracting spatio-temporal features in the spatial features using a spatio-temporal graph neural network is:: ​

[0054]

[0055] Wherein, represents the feature vector of sensor node j at time t; y t represents the output at time t; X T-k+1 represents the input feature at time t-k+1; w k represents the convolution kernel; represents the updated feature of sensor node i at time t+1; represents the final fusion feature of sensor node i; α represents the weight coefficient for controlling the fusion ratio of spatial features and temporal features; represents the spatial feature of sensor node i.

[0056] Preferably, principal component analysis technology is used to perform principal component analysis dimensionality reduction on the extracted spatio-temporal features, and the expression of the spatio-temporal features after dimensionality reduction is:

[0057]

[0058] X r =X c W

[0059] Wherein, represents the sample mean; X c is the data matrix after centering processing; X r represents the spatio-temporal feature data after dimensionality reduction; m represents the number of samples; x i represents the sample data; X represents the spatio-temporal feature matrix; represents the transpose matrix of X c .

[0060] Preferably, based on the spatio-temporal features after dimensionality reduction, the calculation formula for the anomaly score of each sensor node in the association graph using principal component analysis technology is:

[0061]

[0062] Wherein, P represents the anomaly score; represents the sample reconstructed based on the spatio-temporal features after dimensionality reduction;

[0063] The calculated anomaly score is compared with the set anomaly score threshold. When the calculated anomaly score is higher than the set anomaly score threshold, the sensor node is a potential fault point.

[0064] Preferably, the calculation formula for the error between the multi-source sensor data of the potential fault point and the predicted value of the potential fault point at the next time point is:

[0065] e t= V a -V p

[0066] Wherein, e t represents the error between the multi-source sensor data of the potential fault point at time t and the predicted value of the potential fault point at the next time point; V a represents the multi-source sensor data of the potential fault point at time t; V p represents the predicted value of the potential fault point at the next time point;

[0067] The calculation formula for the normal range is:

[0068]

[0069] A = μ e ± 3σ e

[0070] Wherein, μ e represents the mean error; σ e represents the standard deviation of the error; A represents the normal range;

[0071] Compare the calculated error with the normal range. When the calculated error exceeds the normal range, the potential fault point is an abnormal point.

[0072] The third aspect of the present invention provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the above method are implemented.

[0073] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0074] Aiming at the problem of multi-source data fusion in the fault diagnosis of heating, ventilation and air conditioning systems, the present invention proposes an unsupervised method based on graph neural network to solve the complexity of multi-sensor data and the problem of real-time diagnosis. The fault diagnosis method of heating, ventilation and air conditioning systems in intelligent buildings based on graph neural network of the present invention significantly improves the accuracy, real-time and robustness of fault detection of heating, ventilation and air conditioning systems by constructing a fault diagnosis system based on graph neural network. In this method, the Pearson correlation coefficient between each multi-source sensor data is first calculated, and a graph structure reflecting the correlation of multi-source sensors is constructed to capture the spatial correlation characteristics of multi-source data in the system. Combined with graph attention network and spatiotemporal graph neural network, this method can extract the spatial and temporal characteristics of multi-sensor data, so as to comprehensively identify the dynamic changes of the system. In addition, principal component analysis is used to reduce the dimension of data, and potential faults are judged according to the abnormal scores after dimension reduction, without relying on a large amount of labeled data, and fault detection in an unlabeled data environment is successfully realized. This design not only solves the challenges brought by the diversity, spatial heterogeneity and dynamics of heating, ventilation and air conditioning system data, but also has strong real-time processing capability and noise resistance. By accurately calculating anomaly scores in unsupervised scenarios, accurate identification of heating, ventilation and air-conditioning system faults is achieved, reducing maintenance costs and manual labeling costs, while improving the reliability and stability of system operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 It is a schematic diagram of the structure of the building heating, ventilation and air conditioning system fault diagnosis system of the present invention;

[0076] Figure 2 It is a flow chart of the building heating, ventilation and air conditioning system fault diagnosis method of the present invention. DETAILED DESCRIPTION

[0077] The specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation described herein is only used to illustrate and explain the present invention, and is not used to limit the present invention.

[0078] Example 1

[0079] like Figure 1 The building heating, ventilation and air conditioning system fault diagnosis system shown includes:

[0080] The data acquisition and processing module is used to collect multi-source sensor data from multi-source sensors at different locations and components of the heating, ventilation and air-conditioning system, and pre-process the collected multi-source sensor data to obtain pre-processed multi-source sensor data;

[0081] The data construction module is used to calculate the Pearson correlation coefficient between multi-source sensors based on the preprocessed multi-source sensor data, construct a correlation coefficient matrix according to the calculated Pearson correlation coefficient between multi-source sensors, and construct an association graph using the correlation coefficient matrix;

[0082] The feature extraction module is used to extract the spatial features in the association graph using a multi-layer graph attention network, and then extract the spatio-temporal features in the spatial features using a spatio-temporal graph neural network;

[0083] The anomaly detection module is used to perform principal component analysis dimensionality reduction on the extracted spatio-temporal features using principal component analysis technology to obtain the dimensionality-reduced spatio-temporal features. Based on the dimensionality-reduced spatio-temporal features, use principal component analysis technology to calculate the anomaly score of each sensor node in the association graph, compare the calculated anomaly score with the set anomaly score threshold, and judge potential fault points;

[0084] The fault diagnosis module is used to generate a predicted value of the potential fault point at the next time point using a spatio-temporal graph neural network according to the spatio-temporal features corresponding to the potential fault point, calculate the error between the multi-source sensor data of the potential fault point and the predicted value of the potential fault point at the next time point, and compare the calculated error with the normal range to judge the anomaly point.

[0085] In the present invention, the intelligent controller in the heating, ventilation, and air conditioning system (HVAC system) obtains data in real time by communicating with the sensor network for data acquisition and processing.

[0086] The multi-source sensors in the present invention include working state sensors (such as temperature, pressure, flow, current sensors, etc.) and environmental monitoring sensors (such as humidity, air quality, outdoor meteorological sensors, etc.), from which multi-dimensional data such as equipment operating status, refrigerant flow characteristics, and environmental conditions can be collected, providing comprehensive data support for fault diagnosis and operation optimization

[0087] Multi-source sensor data refers to various data obtained from multiple different sources and different types of multi-source sensors. These data relate to the working status and environmental conditions of each component of the HVAC system. The HVAC system is composed of multiple functional components and monitoring points, and each component has different working environments and working parameters. Therefore, corresponding data can be collected from multi-source sensors at different positions and components of the system.

[0088] In the HVAC system, the multi-source sensors at different positions and components of the heating, ventilation, and air conditioning system refer to the multi-source sensors installed at different positions and components of the heating, ventilation, and air conditioning system, which are used to collect data such as temperature, humidity, air flow, pressure, and energy consumption in real time.

[0089] In specific embodiments, multi-source sensor data can be collected in real time from air handling units, variable air volume systems, chiller systems, variable refrigerant flow systems, etc.

[0090] The application of multi-source data fusion technology (the application of multi-source data fusion technology is mainly reflected in the multi-dimensional and multi-type fusion of sensor data) in fault diagnosis has significant advantages. Especially in complex systems such as heating, ventilation, and air conditioning systems, it can effectively improve the intelligent level and diagnostic ability of the system. First of all, multi-source data fusion can integrate diverse information from different sensors, giving full play to the complementarity of various types of data and providing a more comprehensive view of the system state. The processing of this comprehensive information not only helps to reveal the complex relationships between sensors but also identify potential fault patterns, enhancing the accuracy and reliability of fault diagnosis. Secondly, by fusing data from different sources, the noise and bias that may exist in a single data source can be overcome, thereby improving the robustness of the system. In practical applications, sensor data is often interfered with by external environmental factors, and multi-source data fusion can effectively reduce the impact of these interferences on the diagnostic results.

[0091] In addition, multi-source data fusion also supports real-time monitoring and dynamic analysis, and can quickly respond to changes in the system state. By integrating and analyzing data in real time, potential problems can be identified at an early stage of fault detection, thereby reducing downtime and maintenance costs and improving the overall operating efficiency of the system. Finally, with the development of intelligent technologies, multi-source data fusion methods can be combined with advanced machine learning and deep learning algorithms to achieve self-learning and optimization, enabling the system to continuously improve its fault diagnosis ability during long-term operation.

[0092] Taking the chiller system, which is the core of heating, ventilation, and air conditioning systems, as an example, the data of the chiller comes from multiple modules (such as chilled water pumps, condensers, compressors), and the information is complex, making it suitable for precise diagnosis through multi-source data fusion: the combination of water temperature and flow data can effectively monitor the normal operation and energy efficiency of the water circuit. The operating state of the compressor and the refrigerant pressure can help identify refrigeration faults and compressor efficiency problems. Through the present invention, early warnings and fault diagnoses can be carried out on abnormal states of the chiller (such as compressor failure, refrigerant leakage).

[0093] In a preferred case, the preprocessing described in the present invention includes standardization and noise filtering, that is, preprocessing the collected multi-source sensor data includes standardizing and noise filtering the collected multi-source sensor data; standardization and noise filtering are part of data preprocessing and are conventional operations in the art; standardization is to normalize the data by the unit vector according to its norm, so that the size of the data is not affected; noise filtering is to remove high-frequency noise in the data with a filter. The purpose of standardization is to unify the scales of multi-source sensor data, avoid the imbalance impact between different features, and improve the algorithm stability to prevent some large-scale data from dominating the model. The purpose of noise filtering is to eliminate random errors, reduce noise interference, and improve the robustness of the model.

[0094] In the present invention, the purpose of preprocessing the collected multi-source sensor data is to ensure data consistency and stability; calculating the Pearson correlation coefficient between multi-source sensors using the preprocessed multi-source sensor data aims to eliminate the dimension difference and ensure that the data between different multi-source sensors can be compared on the same scale.

[0095] Further, calculate the Pearson correlation coefficient between multi-source sensors based on the preprocessed multi-source sensor data (the process of calculating the Pearson correlation coefficient depends on classical statistical methods and can determine the association strength between each multi-source sensor), and construct an expression for the correlation coefficient matrix based on the calculated Pearson correlation coefficient between multi-source sensors (the correlation coefficient matrix is obtained by calculating the Pearson correlation coefficient between all multi-source sensors. It is an N×N matrix that shows the linear correlation between each pair of multi-source sensors) as follows:

[0096]

[0097] In the formula, r ij represents the Pearson correlation coefficient between the i-th multi-source sensor and the j-th multi-source sensor. The closer the absolute value of r ij is to 1, the stronger the association between the multi-source sensors; x ti represents the preprocessed multi-source sensor data of the i-th multi-source sensor at time t (that is, the multi-source sensor data collected by the i-th multi-source sensor at time t is preprocessed to obtain the preprocessed multi-source sensor data); x tj represents the preprocessed multi-source sensor data of the j-th multi-source sensor at time t (that is, the multi-source sensor data collected by the j-th multi-source sensor at time t is preprocessed to obtain the preprocessed multi-source sensor data). denotes the mean of the pre - processed multi - source sensor data of the \(i\) - th multi - source sensor during the acquisition time (that is, within a certain acquisition time, multiple multi - source sensor data collected by the \(i\) - th multi - source sensor are all pre - processed, and the mean of the pre - processed multi - source sensor data is obtained). The acquisition time is usually set during system design. By collecting multi - source sensor data within a reasonable time window, the spatio - temporal change characteristics of the system can be captured, which helps subsequent analysis and fault diagnosis. For example, if the change pattern of the historical data of the multi - source sensor is not significant, it is difficult to accurately diagnose potential faults through only the data at one moment; denotes the mean of the pre - processed multi - source sensor data of the \(j\) - th multi - source sensor during the acquisition time (that is, within a certain acquisition time, multiple multi - source sensor data collected by the \(j\) - th multi - source sensor are all pre - processed, and the mean of the pre - processed multi - source sensor data is obtained); \(N\) represents the length of the time series, that is, the total number of data points (the total number of multi - source sensor data collected); \(A\) ij represents an element in the correlation coefficient matrix, indicating whether an edge connection is established between the \(i\) - th multi - source sensor and the \(j\) - th multi - source sensor; \(\theta\) represents the threshold of the Pearson correlation coefficient. \(\theta\) is used to perform thresholding on these Pearson correlation coefficients. That is, if the absolute value \(|r|\) of the Pearson correlation coefficient between the \(i\) - th multi - source sensor and the \(j\) - th multi - source sensor ij is greater than the set threshold \(\theta\), it is considered that there is a strong correlation between them, and a connection is established in the association graph; if \(|r|\) ij is less than or equal to \(\theta\), it is considered that the correlation between them is weak, and no connection is established.

[0098] During the above process of calculating the Pearson correlation coefficient, it is usually necessary to centralize the data of each multi - source sensor by subtracting the mean of the data to eliminate the data offset, so that the calculation result can reflect the relative change of the data.

[0099] In the present invention, the specific operation of constructing an association graph using the correlation coefficient matrix includes: connecting sensor nodes with weighted edges, and the weight value is determined by the Pearson correlation coefficient between multi - source sensors. A large weight represents a strong correlation.

[0100] In the present invention, the construction of the association graph uses the Pearson correlation coefficient as the connection weight between each multi-source sensor, and constructs a graph structure reflecting the mutual relationship between multi-source sensors. Each multi-source sensor can be regarded as a sensor node in the association graph. The connection between multi-source sensors is weighted according to the Pearson correlation coefficient between them. A higher weight indicates a stronger linear correlation between two multi-source sensors, while a lower weight indicates a weaker correlation or almost no linear relationship. The construction of the association graph is mainly through constructing the association graph between multi-source sensors, analyzing the mutual relationship between different multi-source sensors, representing the spatial relationship of multi-source sensor data, and being used for spatial feature learning in the graph neural network.

[0101] The construction of the association graph is a key step in fault diagnosis using spatio-temporal graph neural networks. It can accurately capture and present the complex mutual relationship between multi-source sensors and its dynamic changes over time. This not only lays a foundation for in-depth fault analysis, but also provides the necessary structured information for subsequent diagnostic decision-making and interpretability analysis. By analyzing the data after sliding window processing, the correlation between various variables is calculated to reveal the potential correlation patterns in multi-source sensor data. Subsequently, an association graph reflecting the relationship between multi-source sensors will be constructed based on the results of these correlation analyses.

[0102] In the present invention, the constructed association graph structure provides accurate association information for the multi-layer graph attention network. The multi-layer graph attention network further uses this information to dynamically allocate weights through the multi-head attention mechanism, thereby strengthening the spatial feature learning of important sensor nodes.

[0103] In the preferred embodiment of the present invention, the multi-layer graph attention network is based on the graph attention network, with the addition of multiple graph convolutional layers or graph attention layers (the operation of adding multiple graph convolutional layers or graph attention layers is a conventional operation in the art, and the number and order of addition depend on the specific required operation). The multi-layer graph attention network is used to process the association graph of multi-source sensors, automatically learning the weighted relationship between multi-source sensors, and assigning higher weights to important multi-source sensor relationships. The multi-layer graph attention network processes the association graph layer by layer by stacking multiple graph convolutional layers or graph attention layers, gradually aggregating and enhancing the features of each sensor node. The graph attention mechanism of each layer can learn the complex relationship between sensor nodes through adaptive weighted aggregation of the information of neighboring sensor nodes, and finally generate the embedding representation of each sensor node, capturing the spatial association features between different multi-source sensors, where the sensor node embedding is used to represent the spatial features of multi-source sensors while retaining the characteristics of multi-source sensors and their associated sensor nodes.

[0104] In the preferred embodiment, the expression for extracting the spatial features from the association graph using the multi-layer graph attention network is:

[0105] e ij = LeakyReLU(a T [Wh i ||Wh j

[0106]

[0107] where, e ij represents the calculated attention score; a T represents a learning parameter (during training, it is gradually adjusted as the model learns) to learn the association between sensor nodes; the "||" in the above formula represents the concatenation of feature vectors (i.e., the "||" in the expression for extracting spatial features in the multi-layer graph attention network represents the concatenation of feature vectors); LeakyReLU represents an activation function to ensure that the calculated attention score is positive; α ij is the normalization coefficient between sensor node i and sensor node j (after constructing the association graph, multi-source sensors can be regarded as sensor nodes in the association graph, so sensor nodes are equal to multi-source sensors. At this time, sensor node i and sensor node j are the i-th multi-source sensor and the j-th multi-source sensor); W represents the linear transformation weight matrix; σ represents the sigmoid activation function; N(i) represents the set of neighbor sensor nodes of sensor node i, representing all sensor nodes directly connected to sensor node i; represents the updated feature of sensor node i at time t; h j represents the spatial feature vector provided by sensor node j, reflecting the data features of this sensor node at a specific spatial position; h i represents the spatial feature vector provided by sensor node i, reflecting the data features of this sensor node at a specific spatial position; K represents the number of attention heads; represents the attention coefficient of the k-th attention head; W k represents the weight matrix of the k-th attention head; e il represents the attention score between sensor node i and neighbor sensor node l of sensor node i.

[0108] ​In the present invention, on the basis of using a multi-layer graph attention network to extract spatial features in the association graph, a spatio-temporal graph neural network is further used to extract spatio-temporal features in the spatial features. The spatio-temporal graph neural network is obtained by stacking a graph convolutional layer and a temporal convolutional layer, which is a conventional operation in the art. The specific stacking operation is manually constructed and designed according to specific problem requirements. The spatio-temporal graph neural network can simultaneously model the spatial structure and time series characteristics of data. The graph convolutional layer processes the association graph structure to extract the spatial features of each sensor node (the reason for using a multi-layer graph attention network to extract spatial features in the association graph in the previous text is that it can more accurately capture the spatial dependence between multi-source sensors. Through the graph attention mechanism, it can accurately capture which sensor nodes have strong correlations and their spatial relationships. This step helps to extract more meaningful spatial features, especially when there are complex multi-source sensor associations in the data, and the attention mechanism can effectively help the model learn this complex spatial dependence). The temporal convolutional layer models the temporal features of each sensor node and learns the temporal dependence of the system. By using the temporal convolutional layer to process the temporal dependence of each multi-source sensor data, the periodic and mutation characteristics during the operation of the device can be captured. The spatio-temporal graph neural network can not only understand the data in the time dimension but also combine the spatial dependence obtained by the graph attention network to more comprehensively evaluate the anomalies in the heating, ventilation, and air conditioning system.

[0109] Preferably, the expression for using a spatio-temporal graph neural network to extract spatio-temporal features in the spatial features is:

[0110]

[0111] In the formula, represents the feature vector of sensor node j at time t; y t represents the output at time t; X t-k+1 represents the input feature at time t - k + 1; w k represents the convolutional kernel used to process time series data; the "*" in the above formula is the convolution operation used to capture the dynamic features in the time dimension; represents the updated feature of sensor node i at time t + 1; represents the final fused feature of sensor node i, integrating spatial and temporal information; α represents the weight coefficient controlling the fusion ratio of spatial and temporal features; represents the spatial feature of sensor node i, reflecting the relationship between sensor nodes.

[0112] Furthermore, in the present invention, the principal component analysis technique (PCA) is used to perform principal component analysis dimensionality reduction on the extracted spatio-temporal features, obtaining the spatio-temporal features after dimensionality reduction (i.e., obtaining the data matrix after dimensionality reduction, the eigenvectors and eigenvalues corresponding to each principal component). PCA extracts the most representative features by finding the direction of the largest variance in the data to reduce noise and redundant information. After receiving the preliminary anomaly score data from the spatio-temporal graph neural network, the principal component analysis removes the noise and unstable factors in the score data through principal component decomposition, thereby generating more robust anomaly scores, enabling subsequent tasks to be carried out more efficiently.

[0113] Among them, the expression for performing principal component analysis dimensionality reduction on the extracted spatio-temporal features using the principal component analysis technique to obtain the spatio-temporal features after dimensionality reduction is:

[0114]

[0115] X r = X c W

[0116] In the formula, represents the sample mean (i.e., the sample mean after being processed by the spatio-temporal graph neural network); X c is the data matrix after centering processing; X r represents the spatio-temporal feature data after dimensionality reduction; the "∑" in the above formula is a covariance calculation operation (i.e., the "∑" in the expression for performing principal component analysis dimensionality reduction on the extracted spatio-temporal features using the principal component analysis technique to obtain the spatio-temporal features after dimensionality reduction is a covariance calculation operation); m represents the number of samples; x i represents the sample data (i.e., the sample data after being processed by the spatio-temporal graph neural network); X represents the spatio-temporal feature matrix; represents the transpose matrix of X c .

[0117] In the present invention, based on the spatio-temporal features after dimensionality reduction, using the principal component analysis technique to calculate the anomaly score of each sensor node in the correlation graph is used to identify potential abnormal situations. In the low-dimensional space of the principal component analysis, anomalies are detected by calculating the degree of deviation of the sensor nodes from the normal mode. Among them, PCA is mainly used to identify abnormal data points, and the anomaly score is calculated by calculating the reconstruction error or projection error.

[0118] Among them, the calculation formula for using the principal component analysis technique to calculate the anomaly score of each sensor node in the correlation graph based on the spatio-temporal features after dimensionality reduction is:

[0119]

[0120] In the formula, P represents the anomaly score; Denote the samples reconstructed based on the dimensionality-reduced spatio-temporal features (i.e., the reconstructed samples obtained by inverse transformation of the dimensionality-reduced spatio-temporal features).

[0121] In the present invention, the calculation formula of the above-mentioned anomaly score calculates the difference between a data point and its reconstruction result. In anomaly detection, a larger reconstruction error indicates that the data point is abnormal, and a sensor node with a higher anomaly score will be marked as a potential fault point.

[0122] Further, compare the calculated anomaly score with a set anomaly score threshold. When the calculated anomaly score is higher than the set anomaly score threshold, the sensor node is a potential fault point.

[0123] The anomaly score threshold is set by calculating the mean and standard deviation of the anomaly scores of all data points, and is usually set as the mean plus or minus several times the standard deviation. The specific threshold setting can be adjusted according to actual requirements or experimental results.

[0124] In the present invention, the spatio-temporal graph neural network first extracts spatial and temporal patterns from the original data, and then uses these patterns to make predictions for the next time point. If there are observed data inconsistent with the historical patterns, the spatio-temporal graph neural network will reflect it in the form of a large error, which can thus be used to detect anomalies.

[0125] Further, the calculation formula for the error between the multi-source sensor data of the potential fault point and the predicted value of the potential fault point at the next time point is:

[0126] e t =V a -V p

[0127] In the formula, e t represents the error between the multi-source sensor data of the potential fault point at time t and the predicted value of the potential fault point at the next time point; V a represents the multi-source sensor data of the potential fault point at time t, that is, the multi-source sensor data of the potential fault point collected at time t; V p represents the predicted value of the potential fault point at the next time point (i.e., the predicted value at time t + 1), and the predicted value of the potential fault point at the next time point refers to the predicted multi-source sensor data of the potential fault point at the next time point.

[0128] Among them, predicting the multi-source sensor data at the next time point is different from directly collecting the multi-source sensor data at the next time point. The prediction method can identify potential anomalies in advance, while directly collecting the multi-source sensor data at the next time point is for post-event diagnosis. Therefore, in the present invention, predicting the multi-source sensor data at the next time point can provide more timely feedback.

[0129] The calculation formula for the normal range is as follows:

[0130]

[0131] A = μ e ± 3σ e

[0132] In the formula, μ e represents the mean error; σ e represents the standard deviation of the error; A represents the normal range.

[0133] In the specific implementation, the calculated error is compared with the normal range. When the calculated error exceeds the normal range, the potential fault point is the abnormal point.

[0134] Among them, the abnormal points indicate that at a certain time point and on a certain sensor node (multi-source sensor), the data has a significant deviation from the normal mode. However, these points do not necessarily directly point to the fault source. There may be multiple factors behind the abnormal points, and they may be closely related to the states of other sensor nodes, the environment in the system, or the operating state. Therefore, it is necessary to further analyze the adjacent sensor nodes and the correlation of the abnormal points to determine the main fault sources behind these abnormal points.

[0135] In the preferred case, the weight information of the neighbor sensor nodes in the correlation graph learned by the multi-layer graph attention network is used to analyze the adjacent variables of the abnormal points (high-abnormal-score sensor nodes) to identify the main fault sources (that is, by analyzing the influence of its neighbor sensor nodes through the multi-layer graph attention network to help locate the possible fault sources), and the fault diagnosis is completed.

[0136] In the training stage of the model in the present invention, only normal data is used. By learning the normal mode, a baseline is established. In the detection stage, the data deviating from the normal mode is identified through the abnormal score, and automatic and real-time fault detection is achieved.

[0137] In the present invention, the graph neural network can effectively process graph-structured data, extract the complex relationships between sensor nodes, and provide a new solution for multi-source data fusion. In particular, the multi-layer graph attention network and the spatio-temporal graph neural network perform excellently in capturing spatio-temporal dynamic features and can effectively combine the spatial associations and temporal changes between different multi-source sensors. This characteristic provides a good basis for the fault diagnosis of heating, ventilation, and air conditioning systems.

[0138] In addition, as a classic dimensionality reduction technique, principal component analysis is widely used in feature extraction and anomaly detection. By reducing the dimensionality of data, principal component analysis can reveal the underlying structure in the data, reduce noise interference, and improve the training efficiency of subsequent models. In an unsupervised learning environment, the reconstruction error of principal component analysis can be used to judge the abnormality of data points, thereby realizing effective fault detection.

[0139] In summary, by combining graph neural networks, unsupervised learning, and principal component analysis techniques, an innovative multi-source data fusion fault diagnosis method can be constructed to meet the fault monitoring requirements of heating, ventilation, and air conditioning systems in an environment of unlabeled data. This research not only helps to improve the fault diagnosis ability of heating, ventilation, and air conditioning systems but also provides a theoretical basis and technical support for the intelligent operation and maintenance of other complex systems.

[0140] The unsupervised fault diagnosis method for multi-source data fusion based on graph neural networks in the present invention shows significant advantages compared with traditional fault detection techniques. First of all, traditional methods usually rely on a large amount of labeled data, while the present invention can operate effectively in an unlabeled environment, reducing the cost and complexity of manual labeling, especially suitable for scenarios where faults occur relatively less frequently in heating, ventilation, and air conditioning systems. Secondly, by using graph neural networks, especially graph attention networks and spatio-temporal graph neural networks, this method can deeply mine the spatial and temporal features of multi-source data, accurately capture the complex correlations between multi-source sensors, and improve the accuracy and reliability of fault diagnosis. Compared with traditional single feature extraction methods, the present invention shows stronger adaptability and effectiveness when dealing with diverse sensor data.

[0141] In addition, by combining principal component analysis for anomaly score calculation, the performance of this method in real-time fault detection is further enhanced. Traditional methods often cannot effectively handle the noise and interference in data, while the present invention has strong robustness, ensuring accurate fault judgment even in an uncertain operating environment. Finally, due to the flexibility and scalability of its design, the present invention is not only applicable to heating, ventilation, and air conditioning systems but also can be extended to the intelligent fault diagnosis of other complex systems, providing new ideas and solutions for broader industrial applications.

[0142] Embodiment 2

[0143] As Figure 2 shown, the fault diagnosis method for a building heating, ventilation, and air conditioning system includes:

[0144] Collect multi-source sensor data from multi-source sensors at different positions and components of the heating, ventilation, and air conditioning system, and preprocess the collected multi-source sensor data to obtain preprocessed multi-source sensor data;

[0145] Calculate the Pearson correlation coefficients between multi-source sensors based on the preprocessed multi-source sensor data, construct a correlation coefficient matrix based on the calculated Pearson correlation coefficients between multi-source sensors, and construct an association graph using the correlation coefficient matrix;

[0146] Use a multi-layer graph attention network to extract spatial features from the association graph, and then use a spatio-temporal graph neural network to extract spatio-temporal features from the spatial features;

[0147] Use principal component analysis technology to perform principal component analysis and dimensionality reduction on the extracted spatio-temporal features to obtain the dimensionality-reduced spatio-temporal features. Based on the dimensionality-reduced spatio-temporal features, use principal component analysis technology to calculate the anomaly scores of each sensor node in the association graph, compare the calculated anomaly scores with the set anomaly score threshold, and judge potential fault points;

[0148] According to the spatio-temporal features corresponding to the potential fault points, use a spatio-temporal graph neural network to generate predicted values of the potential fault points at the next time point, calculate the error between the multi-source sensor data of the potential fault points and the predicted values of the potential fault points at the next time point, and compare the calculated error with the normal range to judge anomaly points.

[0149] In the present invention, an intelligent controller in a heating, ventilation, and air conditioning (HVAC) system communicates with a sensor network to obtain data in real time for data acquisition and processing.

[0150] The multi-source sensors in the present invention include working state sensors (such as temperature, pressure, flow rate, current sensors, etc.) and environmental monitoring sensors (such as humidity, air quality, outdoor meteorological sensors, etc.), from which multi-dimensional data such as equipment operation status, refrigerant flow characteristics, and environmental conditions can be collected, providing comprehensive data support for fault diagnosis and operation optimization

[0151] Multi-source sensor data refers to various data obtained from multi-source sensors of different sources and types. These data relate to the working states and environmental conditions of various components of the HVAC system. The HVAC system is composed of multiple functional components and monitoring points, and each component has different working environments and working parameters. Therefore, corresponding data can be collected from multi-source sensors at different positions and components of the system.

[0152] In the HVAC system, multi-source sensors at different positions and components of the heating, ventilation, and air conditioning system refer to multi-source sensors installed at different positions and components of the heating, ventilation, and air conditioning system for real-time collection of data such as temperature, humidity, air flow, pressure, and energy consumption.

[0153] In the specific implementation, multi-source sensor data can be collected in real time from air handling units, variable air volume systems, chiller systems, variable refrigerant flow systems, etc.

[0154] The application of multi-source data fusion technology (the application of multi-source data fusion technology is mainly reflected in the multi-dimensional and multi-type fusion of sensor data) in fault diagnosis has significant advantages. Especially in complex systems such as heating, ventilation, and air conditioning systems, it can effectively improve the intelligent level and diagnostic ability of the system. First of all, multi-source data fusion can integrate diverse information from different sensors, giving full play to the complementarity of various types of data and providing a more comprehensive view of the system state. The processing of this comprehensive information not only helps to reveal the complex relationships between sensors but also enables the identification of potential fault patterns, enhancing the accuracy and reliability of fault diagnosis. Secondly, by fusing data from different sources, the noise and bias that may exist in a single data source can be overcome, thereby improving the robustness of the system. In practical applications, sensor data is often affected by external environmental factors, and multi-source data fusion can effectively reduce the impact of these interferences on the diagnostic results.

[0155] In addition, multi-source data fusion also supports real-time monitoring and dynamic analysis, and can quickly respond to changes in the system state. By integrating and analyzing data in real time, potential problems can be identified at an early stage of fault detection, thereby reducing downtime and maintenance costs and improving the overall operating efficiency of the system. Finally, with the development of intelligent technologies, multi-source data fusion methods can be combined with advanced machine learning and deep learning algorithms to achieve self-learning and optimization, enabling the system to continuously improve its fault diagnosis ability during long-term operation.

[0156] Taking the chiller system, which is the core of the heating, ventilation, and air conditioning system, as an example, the data of the chiller comes from multiple modules (such as chilled water pumps, condensers, compressors), and the information is complex, making it suitable for accurate diagnosis through multi-source data fusion: the combination of water temperature and flow rate data can effectively monitor the normal operation and energy efficiency of the water circuit. The operating state of the compressor and the refrigerant pressure can help identify refrigeration faults and compressor efficiency problems. Through the present invention, early warnings and fault diagnoses can be carried out for abnormal states of the chiller (such as compressor failure, refrigerant leakage).

[0157] In the preferred case, the preprocessing described in the present invention includes standardization and noise filtering, that is, the preprocessing of the multi-source sensor data collected includes standardizing and filtering the noise of the multi-source sensor data collected; standardization and noise filtering are part of data preprocessing and are conventional operations in the art; standardization is to normalize the data according to its norm by unit vector normalization, so that the magnitude of the data is not affected; noise filtering is to remove high-frequency noise in the data using a filter. The purpose of standardization is to unify the scales of multi-source sensor data, avoid the imbalance between different features from affecting, and improve the algorithm stability to prevent some large-scale data from dominating the model. The purpose of noise filtering is to eliminate random errors, reduce noise interference, and improve the robustness of the model.

[0158] In the present invention, the purpose of preprocessing the collected multi-source sensor data is to ensure data consistency and stability; calculating the Pearson correlation coefficient between multi-source sensors using the preprocessed multi-source sensor data aims to eliminate the dimension difference and ensure that the data between different multi-source sensors can be compared on the same scale.

[0159] Furthermore, calculate the Pearson correlation coefficient between multi-source sensors based on the preprocessed multi-source sensor data (the process of calculating the Pearson correlation coefficient relies on classical statistical methods and can determine the association strength between each multi-source sensor), and construct an expression for the correlation coefficient matrix based on the calculated Pearson correlation coefficient between multi-source sensors (the correlation coefficient matrix is obtained by calculating the Pearson correlation coefficient between all multi-source sensors, and it is an N×N matrix that shows the linear correlation between each pair of multi-source sensors) as follows:

[0160]

[0161] In the formula, r ij represents the Pearson correlation coefficient between the i-th multi-source sensor and the j-th multi-source sensor. The closer the absolute value of r ij is to 1, the stronger the association between the multi-source sensors; x ti represents the preprocessed multi-source sensor data of the i-th multi-source sensor at time t (that is, the multi-source sensor data collected by the i-th multi-source sensor at time t is preprocessed to obtain the preprocessed multi-source sensor data); x tj represents the preprocessed multi-source sensor data of the j-th multi-source sensor at time t (that is, the multi-source sensor data collected by the j-th multi-source sensor at time t is preprocessed to obtain the preprocessed multi-source sensor data). represents the mean of the preprocessed multi-source sensor data of the i-th multi-source sensor during the acquisition time (that is, within a certain acquisition time, the multi-source sensor data collected by the i-th multi-source sensor are all preprocessed to obtain the mean of the preprocessed multi-source sensor data). The acquisition time is usually set during system design. By collecting multi-source sensor data within a reasonable time window, the spatio-temporal variation characteristics of the system can be captured, which helps subsequent analysis and fault diagnosis. For example, if the historical data change pattern of the multi-source sensor is not significant, it is difficult to accurately diagnose potential faults based on only the data at one moment. denotes the mean of the pre - processed multi - source sensor data of the j - th multi - source sensor within the acquisition time (i.e., within a certain acquisition time, the multiple multi - source sensor data collected by the j - th multi - source sensor are all pre - processed, and the mean of the pre - processed multi - source sensor data is obtained); N represents the length of the time series, that is, the total number of data points (the total number of the collected multi - source sensor data); A ij represents an element in the correlation coefficient matrix, indicating whether an edge connection is established between the i - th multi - source sensor and the j - th multi - source sensor; θ represents the threshold of the Pearson correlation coefficient, and θ is used to perform thresholding on these Pearson correlation coefficients, that is, if the absolute value |r ij | of the Pearson correlation coefficient between the i - th multi - source sensor and the j - th multi - source sensor is greater than the set threshold θ, it is considered that there is a strong correlation between them, and a connection is established in the correlation graph; if |r ij | is less than or equal to θ, it is considered that the correlation between them is weak, and no connection is established.

[0162] In the above process of calculating the Pearson correlation coefficient, it is usually necessary to centralize the data of each multi - source sensor, and eliminate the data offset by subtracting the mean of the data, so that the calculation result can reflect the relative change of the data.

[0163] In the present invention, the specific operation of constructing a correlation graph using the correlation coefficient matrix includes: connecting sensor nodes with weighted edges, and the weight value is determined by the Pearson correlation coefficient between multi - source sensors, and a large weight represents a strong correlation.

[0164] In the present invention, the correlation graph is constructed by using the Pearson correlation coefficient as the connection weight between each multi - source sensor to construct a graph structure reflecting the mutual relationship between multi - source sensors. Each multi - source sensor can be regarded as a sensor node in the correlation graph. The connection between multi - source sensors sets the weight according to the Pearson correlation coefficient between them. A higher weight indicates a stronger linear correlation between two multi - source sensors, and a lower weight indicates a weaker correlation or almost no linear relationship. Constructing the correlation graph is mainly to analyze the mutual relationship between different multi - source sensors by constructing the correlation graph between multi - source sensors, which is used to represent the spatial relationship of multi - source sensor data and is used for spatial feature learning in the graph neural network.

[0165] The construction of the correlation graph is a key step in fault diagnosis using spatio-temporal graph neural networks. It can accurately capture and present the complex interrelationships among multi-source sensors and their dynamic changes over time. This not only lays a foundation for in-depth fault analysis but also provides the necessary structured information for subsequent diagnostic decision-making and interpretability analysis. By analyzing the data processed by the sliding window, the correlations between various variables are calculated to reveal the potential correlation patterns in the multi-source sensor data. Subsequently, a correlation graph reflecting the relationships among multi-source sensors will be constructed based on the results of these correlation analyses.

[0166] In the present invention, the constructed correlation graph structure provides accurate correlation information for the multi-layer graph attention network, and the multi-layer graph attention network further utilizes this information to dynamically allocate weights through the multi-head attention mechanism, thereby strengthening the learning of the spatial features of important sensor nodes.

[0167] In a preferred embodiment of the present invention, the multi-layer graph attention network is based on the graph attention network and adds multiple graph convolutional layers or graph attention layers (the operation of adding multiple graph convolutional layers or graph attention layers is a conventional operation in the art, and the number and order added depend on the specific operation required). The multi-layer graph attention network is used to process the correlation graph of multi-source sensors, automatically learn the weighted relationships among multi-source sensors, and assign higher weights to important multi-source sensor relationships. The multi-layer graph attention network processes the correlation graph layer by layer by stacking multiple graph convolutional layers or graph attention layers, gradually aggregating and enhancing the features of each sensor node. The graph attention mechanism of each layer can learn the complex relationships between sensor nodes by adaptively weighted aggregation of the information of neighboring sensor nodes, and finally generate the embedding representation of each sensor node, capturing the spatial correlation features among different multi-source sensors, where the sensor node embedding is used to represent the spatial features of multi-source sensors while retaining the characteristics of multi-source sensors and their associated sensor nodes.

[0168] In a preferred embodiment, the expression for extracting the spatial features from the correlation graph using the multi-layer graph attention network is:

[0169] e ij =LeakyReLU(a T [Wh i ||Wh j

[0170]

[0171] In the formula, e ij represents the calculated attention score; a T ​denotes the learning parameter (which will be gradually adjusted during the training process as the model learns) to learn the associations between sensor nodes; the "||" in the above formula represents the concatenation of feature vectors (i.e., the "||" in the expression for extracting spatial features in the association graph using the multi-layer graph attention network represents the concatenation of feature vectors); LeakyReLU represents the activation function to ensure that the calculated attention score is positive; α ij is the normalization coefficient between sensor node i and sensor node j (after constructing the association graph, multi-source sensors can be regarded as sensor nodes in the association graph, so sensor nodes are equivalent to multi-source sensors. At this time, sensor node i and sensor node j are the i-th multi-source sensor and the j-th multi-source sensor); W represents the linear transformation weight matrix; σ represents the sigmoid activation function; N(i) represents the set of neighbor sensor nodes of sensor node i, representing all sensor nodes directly connected to sensor node i; represents the updated feature of sensor node i at time t; h j represents the spatial feature vector provided by sensor node j, reflecting the data features of this sensor node at a specific spatial position; h i represents the spatial feature vector provided by sensor node i, reflecting the data features of this sensor node at a specific spatial position; K represents the number of attention heads; represents the attention coefficient of the k-th attention head; W k represents the weight matrix of the k-th attention head; e il represents the attention score between sensor node i and neighbor sensor node l of sensor node i.

[0172] In the present invention, on the basis of using a multi-layer graph attention network to extract spatial features in the association graph, a spatio-temporal graph neural network is further used to extract spatio-temporal features in the spatial features. The spatio-temporal graph neural network is obtained by stacking a graph convolutional layer and a temporal convolutional layer, which is a conventional operation in the art. The specific stacking operation is manually constructed and designed according to specific problem requirements. The spatio-temporal graph neural network can simultaneously model the spatial structure and time series characteristics of data. The graph convolutional layer processes the association graph structure to extract the spatial features of each sensor node (the reason for using a multi-layer graph attention network to extract spatial features in the association graph in the previous text is that it can more accurately capture the spatial dependence between multi-source sensors. Through the graph attention mechanism, it can accurately capture which sensor nodes have strong correlations and their spatial relationships. This step helps to extract more meaningful spatial features, especially when there are complex multi-source sensor associations in the data, and the attention mechanism can effectively help the model learn this complex spatial dependence). The temporal convolutional layer models the temporal features of each sensor node and learns the temporal dependence of the system. By using the temporal convolutional layer to process the temporal dependence of multi-source sensor data, the periodic and mutation characteristics during device operation can be captured. The spatio-temporal graph neural network can not only understand the data in the time dimension but also combine the spatial dependence obtained by the graph attention network to more comprehensively evaluate the anomalies in the heating, ventilation, and air conditioning system.

[0173] Preferably, the expression for using a spatio-temporal graph neural network to extract spatio-temporal features in the spatial features is:

[0174]

[0175] In the formula, represents the feature vector of sensor node j at time t; y t represents the output at time t; X t-k+1 represents the input feature at time t - k + 1; w k represents the convolution kernel used to process time series data; the "*" in the above formula is the convolution operation used to capture the dynamic features in the time dimension; represents the updated feature of sensor node i at time t + 1; represents the final fused feature of sensor node i, integrating spatial and temporal information; α represents the weight coefficient controlling the fusion ratio of spatial and temporal features; represents the spatial feature of sensor node i, reflecting the relationship between sensor nodes.

[0176] Furthermore, in the present invention, the principal component analysis technique (PCA) is used to perform principal component analysis dimensionality reduction on the extracted spatio-temporal features, obtaining the spatio-temporal features after dimensionality reduction (i.e., obtaining the data matrix after dimensionality reduction, the eigenvectors and eigenvalues corresponding to each principal component). PCA extracts the most representative features by finding the direction of the maximum variance in the data to reduce noise and redundant information. After receiving the preliminary anomaly score data from the spatio-temporal graph neural network, principal component analysis removes the noise and unstable factors in the score data by using principal component decomposition, thereby generating a more robust anomaly score, enabling the subsequent tasks to be carried out more efficiently.

[0177] Among them, the expression for performing principal component analysis dimensionality reduction on the extracted spatio-temporal features using the principal component analysis technique to obtain the spatio-temporal features after dimensionality reduction is:

[0178]

[0179] X r =X c W

[0180] In the formula, represents the sample mean (i.e., the sample mean after being processed by the spatio-temporal graph neural network); X c is the data matrix after centering processing; X r represents the spatio-temporal feature data after dimensionality reduction; the "∑" in the above formula is the covariance calculation operation (i.e., the "∑" in the expression for performing principal component analysis dimensionality reduction on the extracted spatio-temporal features using the principal component analysis technique to obtain the spatio-temporal features after dimensionality reduction is the covariance calculation operation); m represents the number of samples; x i represents the sample data (i.e., the sample data after being processed by the spatio-temporal graph neural network); X represents the spatio-temporal feature matrix; represents the transpose matrix of X c .

[0181] In the present invention, based on the spatio-temporal features after dimensionality reduction, using the principal component analysis technique to calculate the anomaly score of each sensor node in the correlation graph is used to identify potential abnormal situations. In the low-dimensional space of principal component analysis, anomalies are detected by calculating the degree of deviation of the sensor nodes from the normal mode. Among them, PCA is mainly used to identify abnormal data points, and the anomaly score is calculated by calculating the reconstruction error or projection error.

[0182] Among them, the calculation formula for using the principal component analysis technique to calculate the anomaly score of each sensor node in the correlation graph based on the spatio-temporal features after dimensionality reduction is:

[0183]

[0184] In the formula, P represents the anomaly score; Denote the samples reconstructed based on the reduced - dimensional spatio - temporal features (i.e., the reconstructed samples obtained by inverse transformation of the reduced - dimensional spatio - temporal features).

[0185] In the present invention, the calculation formula of the above - mentioned anomaly score calculates the difference between the data point and its reconstruction result. In anomaly detection, a larger reconstruction error indicates that the data point is abnormal, and the sensor nodes with higher anomaly scores will be marked as potential fault points.

[0186] Furthermore, compare the calculated anomaly score with the set anomaly score threshold. When the calculated anomaly score is higher than the set anomaly score threshold, the sensor node is a potential fault point.

[0187] Among them, the anomaly score threshold is set by calculating the mean and standard deviation of the anomaly scores of all data points, usually set as the mean plus or minus several times the standard deviation. The specific threshold setting can be adjusted according to actual needs or experimental results.

[0188] In the present invention, the spatio - temporal graph neural network first extracts spatial and temporal patterns from the original data, and then uses these patterns to make predictions for the next time point. If there is observational data inconsistent with the historical patterns, the spatio - temporal graph neural network will reflect it in the form of a large error, which can thus be used to detect anomalies.

[0189] Furthermore, the calculation formula for the error between the multi - source sensor data of the potential fault point and the predicted value of the potential fault point at the next time point is as follows:

[0190] e t =V a -V p

[0191] In the formula, e t represents the error between the multi - source sensor data of the potential fault point at time t and the predicted value of the potential fault point at the next time point; V a represents the multi - source sensor data of the potential fault point at time t, that is, the multi - source sensor data of the potential fault point collected at time t; V p represents the predicted value of the potential fault point at the next time point (i.e., the predicted value at time t + 1), and the predicted value of the potential fault point at the next time point refers to the predicted multi - source sensor data of the potential fault point at the next time point.

[0192] Among them, predicting the multi - source sensor data at the next time point is different from directly collecting the multi - source sensor data at the next time point. The prediction method can identify potential anomalies in advance, while directly collecting the multi - source sensor data at the next time point is for post - event diagnosis. Therefore, in the present invention, predicting the multi - source sensor data at the next time point can provide more timely feedback.

[0193] The calculation formula for the normal range is as follows:

[0194]

[0195] A = μ e ± 3σ e

[0196] In the formula, μ e represents the mean error; σ e represents the standard deviation of the error; A represents the normal range.

[0197] In the specific implementation, the calculated error is compared with the normal range. When the calculated error exceeds the normal range, the potential fault point is the abnormal point.

[0198] Among them, the abnormal points indicate that at a certain time point and on a certain sensor node (multi-source sensor), the data has a significant deviation from the normal mode. However, these points do not necessarily directly point to the fault source. There may be multiple factors behind the abnormal points, and they may be closely related to the states of other sensor nodes, the environment in the system, or the operating state. Therefore, it is necessary to further analyze the adjacent sensor nodes and the relevance of the abnormal points to determine the main fault source behind these abnormal points.

[0199] In the preferred case, the weight information of the neighbor sensor nodes in the correlation graph learned by the multi-layer graph attention network is used to analyze the adjacent variables of the abnormal points (high-abnormal-score sensor nodes) to identify the main fault source (that is, by analyzing the influence of its neighbor sensor nodes through the multi-layer graph attention network to help locate the possible fault source), and the fault diagnosis is completed.

[0200] In the training stage of the model in the present invention, only normal data is used. By learning the normal mode, a baseline is established. In the detection stage, the data deviating from the normal mode is identified through the abnormal score, and automatic and real-time fault detection is realized.

[0201] In the present invention, the graph neural network can effectively process graph-structured data, extract the complex relationships between sensor nodes, and provide a new solution for multi-source data fusion. In particular, the multi-layer graph attention network and the spatio-temporal graph neural network perform excellently in capturing spatio-temporal dynamic features and can effectively combine the spatial correlations and time variations between different multi-source sensors. This feature provides a good basis for the fault diagnosis of heating, ventilation, and air conditioning systems.

[0202] In addition, principal component analysis, as a classic dimensionality reduction technique, is widely used in feature extraction and anomaly detection. By reducing the dimensionality of data, principal component analysis can reveal the underlying structure in the data, reduce noise interference, and improve the training efficiency of subsequent models. In an unsupervised learning environment, the reconstruction error of principal component analysis can be used to determine the abnormality of data points, thereby achieving effective fault detection.

[0203] In summary, by combining graph neural networks, unsupervised learning and principal component analysis techniques, an innovative multi-source data fusion fault diagnosis method can be constructed to meet the fault monitoring needs of heating, ventilation and air conditioning systems in an unlabeled data environment. This research not only helps to improve the fault diagnosis capabilities of heating, ventilation and air conditioning systems, but also provides a theoretical basis and technical support for the intelligent operation and maintenance of other complex systems.

[0204] The multi-source data fusion unsupervised fault diagnosis method based on graph neural network of the present invention shows significant advantages over traditional fault detection technology. First, traditional methods usually rely on a large amount of labeled data, while the present invention can operate effectively in an unlabeled environment, reducing the cost and complexity of manual labeling, and is particularly suitable for scenarios where relatively few faults occur in heating, ventilation and air-conditioning systems. Secondly, by utilizing graph neural networks, especially graph attention networks and spatiotemporal graph neural networks, this method can deeply mine the spatial and temporal characteristics of multi-source data, accurately capture the complex correlations between multi-source sensors, and improve the accuracy and reliability of fault diagnosis. Compared with traditional single feature extraction methods, the present invention shows stronger adaptability and effectiveness when processing diversified sensor data.

[0205] In addition, the anomaly score calculation combined with principal component analysis further enhances the performance of this method in real-time fault detection. Traditional methods are often unable to effectively handle noise and interference in data, while the present invention has strong robustness, ensuring that accurate fault judgments can be made even in uncertain operating environments. Finally, thanks to the flexibility and scalability of its design, the present invention is not only suitable for heating, ventilation and air-conditioning systems, but can also be extended to intelligent fault diagnosis of other complex systems, providing new ideas and solutions for achieving wider industrial applications.

[0206] Example 3

[0207] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of Example 2 are implemented.

[0208] It should be understood that parts not elaborated in detail in this specification belong to the prior art.

[0209] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited thereto. Within the scope of the technical concept of the present invention, various simple modifications can be made to the technical solutions of the present invention, including the combination of each technical feature in any other suitable manner. These simple modifications and combinations should also be regarded as the content disclosed by the present invention and fall within the protection scope of the present invention.

Claims

1. A building heating, ventilation and air conditioning system fault diagnosis system, characterized in that: include: The data acquisition and processing module is used to collect multi-source sensor data from multi-source sensors at different locations and components of the heating, ventilation and air-conditioning system, and pre-process the collected multi-source sensor data to obtain pre-processed multi-source sensor data; The data construction module is used to calculate the Pearson correlation coefficient between the multi-source sensors according to the preprocessed multi-source sensor data, and to construct a correlation coefficient matrix according to the calculated Pearson correlation coefficient between the multi-source sensors, and to construct an association graph using the correlation coefficient matrix; The feature extraction module is used to extract spatial features in the association graph using a multi-layer graph attention network, and then use a spatiotemporal graph neural network to extract spatiotemporal features in the spatial features; The anomaly detection module is used to perform principal component analysis and dimensionality reduction on the extracted spatiotemporal features using principal component analysis technology to obtain the spatiotemporal features after dimensionality reduction. Based on the spatiotemporal features after dimensionality reduction, the principal component analysis technology is used to calculate the anomaly score of each sensor node in the association graph, and the calculated anomaly score is compared with the set anomaly score threshold to determine the potential fault point. The fault diagnosis module is used to generate the predicted value of the potential fault point at the next time point based on the spatiotemporal characteristics corresponding to the potential fault point using the spatiotemporal graph neural network, calculate the error between the multi-source sensor data of the potential fault point and the predicted value of the potential fault point at the next time point, and compare the calculated error with the normal range to determine the abnormal point.

2. The building heating, ventilation and air conditioning system fault diagnosis system according to claim 1, characterized in that: Preprocessing the collected multi-source sensor data includes standardizing and noise filtering the collected multi-source sensor data.

3. The building heating, ventilation and air conditioning system fault diagnosis system according to claim 1 or 2, characterized in that: The Pearson correlation coefficient between the multi-source sensors is calculated based on the preprocessed multi-source sensor data, and the expression of the correlation coefficient matrix is ​​constructed based on the calculated Pearson correlation coefficient between the multi-source sensors: In the formula, r ij represents the Pearson correlation coefficient between the i-th multi-source sensor and the j-th multi-source sensor; x ti represents the preprocessed multi-source sensor data of the i-th multi-source sensor at time t; x tj represents the preprocessed multi-source sensor data of the j-th multi-source sensor at time t; represents the mean value of the preprocessed multi-source sensor data of the i-th multi-source sensor within the acquisition time; represents the mean value of the preprocessed multi-source sensor data of the j-th multi-source sensor within the acquisition time; N represents the length of the time series; A ij represents the elements in the correlation coefficient matrix; θ represents the threshold of Pearson correlation coefficient; The specific operation of constructing the association graph using the correlation coefficient matrix includes: connecting the sensor nodes with weighted edges, and the weight value is determined by the Pearson correlation coefficient between the multi-source sensors.

4. The building heating, ventilation and air conditioning system fault diagnosis system according to claim 1, characterized in that: The expression of extracting spatial features in the association graph using a multi-layer graph attention network is: e ij =LeakyReLU(a T [Wh i ||Wh j ] In the formula, e ij represents the calculated attention score; a T represents learning parameters; LeakyReLU represents activation function; α ij is the normalization coefficient between sensor node i and sensor node j; W represents the linear transformation weight matrix; σ represents the sigmoid activation function; N(i) represents the set of neighboring sensor nodes of sensor node i; represents the update characteristics of sensor node i at time t; h j represents the spatial feature vector provided by sensor node j; h i represents the spatial feature vector provided by sensor node i; K represents the number of attention heads; represents the attention coefficient of the kth attention head; W k represents the weight matrix of the kth attention head; e il represents the attention score between sensor node i and its neighboring sensor node l; The expression of the spatiotemporal features extracted from the spatial features using the spatiotemporal graph neural network is: In the formula, represents the feature vector of sensor node j at time t; y t represents the output at time t; X t-k+1 Represents the input features at time t-k+1; w k represents the convolution kernel; represents the updated features of sensor node i at time t+1; represents the final fusion feature of sensor node i; α represents the weight coefficient that controls the fusion ratio of spatial features and temporal features; represents the spatial characteristics of sensor node i.

5. The building heating, ventilation and air conditioning system fault diagnosis system according to claim 1, characterized in that: The principal component analysis technology is used to perform principal component analysis and dimension reduction on the extracted spatiotemporal features, and the expression of the spatiotemporal features after dimension reduction is obtained as follows: X r =X c W In the formula, represents the sample mean; X c is the data matrix after centralization; X r Represents the spatiotemporal feature data after dimensionality reduction; m represents the number of samples; x i represents sample data; X represents the spatiotemporal feature matrix; Represents X c The transposed matrix of .

6. The building heating, ventilation and air conditioning system fault diagnosis system according to claim 1 or 5, characterized in that: Based on the spatiotemporal features after dimensionality reduction, the calculation formula for calculating the anomaly score of each sensor node in the association graph using principal component analysis technology is: Where P represents the anomaly score; Represents the samples reconstructed based on the spatiotemporal features after dimensionality reduction; The calculated anomaly score is compared with the set anomaly score threshold. When the calculated anomaly score is higher than the set anomaly score threshold, the sensor node is a potential fault point.

7. The building heating, ventilation and air conditioning system fault diagnosis system according to claim 1, characterized in that: The calculation formula for the error between the multi-source sensor data of the potential fault point and the predicted value of the potential fault point at the next time point is: yes t =V a -V p In the formula, e t V represents the error between the multi-source sensor data of the potential fault point at time t and the predicted value of the potential fault point at the next time point; a V represents the multi-source sensor data of the potential fault point at time t; p Represents the predicted value of the potential failure point at the next time point; The calculation formula for the normal range is: A=μ e ±3σ e In the formula, μ e represents the mean error; σ e represents the standard deviation of error; A represents the normal range; The calculated error is compared with the normal range. When the calculated error exceeds the normal range, the potential fault point is an abnormal point.

8. A method for diagnosing faults in a building heating, ventilation and air conditioning system, characterized in that: include: Collecting multi-source sensor data from multi-source sensors at different locations and components of the heating, ventilation and air conditioning system, and preprocessing the collected multi-source sensor data to obtain preprocessed multi-source sensor data; The Pearson correlation coefficients between the multi-source sensors are calculated according to the preprocessed multi-source sensor data, and a correlation coefficient matrix is ​​constructed according to the calculated Pearson correlation coefficients between the multi-source sensors, and a correlation graph is constructed using the correlation coefficient matrix; A multi-layer graph attention network is used to extract spatial features in the correlation graph, and then a spatiotemporal graph neural network is used to extract spatiotemporal features from the spatial features; The principal component analysis technology is used to reduce the dimension of the extracted spatiotemporal features to obtain the spatiotemporal features after dimensionality reduction. Based on the spatiotemporal features after dimensionality reduction, the principal component analysis technology is used to calculate the anomaly score of each sensor node in the association graph, and the calculated anomaly score is compared with the set anomaly score threshold to determine the potential fault point. According to the spatiotemporal characteristics corresponding to the potential fault point, the spatiotemporal graph neural network is used to generate the predicted value of the potential fault point at the next time point, and the error between the multi-source sensor data of the potential fault point and the predicted value of the potential fault point at the next time point is calculated. The calculated error is compared with the normal range to determine the abnormal point.

9. The method for diagnosing faults in a building heating, ventilation and air conditioning system according to claim 8, characterized in that: Preprocessing the collected multi-source sensor data includes standardizing and noise filtering the collected multi-source sensor data.

10. The method for diagnosing faults in a building heating, ventilation and air conditioning system according to claim 8 or 9, characterized in that: The Pearson correlation coefficient between the multi-source sensors is calculated based on the preprocessed multi-source sensor data, and the expression of the correlation coefficient matrix is ​​constructed based on the calculated Pearson correlation coefficient between the multi-source sensors: In the formula, r ij represents the Pearson correlation coefficient between the i-th multi-source sensor and the j-th multi-source sensor; x ti represents the preprocessed multi-source sensor data of the i-th multi-source sensor at time t; x tj represents the preprocessed multi-source sensor data of the j-th multi-source sensor at time t; represents the mean value of the preprocessed multi-source sensor data of the i-th multi-source sensor within the acquisition time; represents the mean value of the preprocessed multi-source sensor data of the j-th multi-source sensor within the acquisition time; N represents the length of the time series; A ij represents the elements in the correlation coefficient matrix; θ represents the threshold of Pearson correlation coefficient; The specific operation of constructing the association graph using the correlation coefficient matrix includes: connecting the sensor nodes with weighted edges, and the weight value is determined by the Pearson correlation coefficient between the multi-source sensors.

11. The method for diagnosing faults in a building heating, ventilation and air conditioning system according to claim 8, characterized in that: The expression of extracting spatial features in the association graph using a multi-layer graph attention network is: e ij =LeakyReLU(a T [Wh i ||Wh j ] In the formula, e ij represents the calculated attention score; a T represents learning parameters; LeakyReLU represents activation function; α ij is the normalization coefficient between sensor node i and sensor node j; W represents the linear transformation weight matrix; σ represents the sigmoid activation function; N(i) represents the set of neighboring sensor nodes of sensor node i; represents the update characteristics of sensor node i at time t; h j represents the spatial feature vector provided by sensor node j; h i represents the spatial feature vector provided by sensor node i; K represents the number of attention heads; represents the attention coefficient of the kth attention head; W k represents the weight matrix of the kth attention head; e il represents the attention score between sensor node i and its neighboring sensor node l; The expression of the spatiotemporal features extracted from the spatial features using the spatiotemporal graph neural network is: In the formula, represents the feature vector of sensor node j at time t; y t represents the output at time t; X t-k+1 Represents the input features at time t-k+1; w k represents the convolution kernel; represents the updated features of sensor node i at time t+1; represents the final fusion feature of sensor node i; α represents the weight coefficient that controls the fusion ratio of spatial features and temporal features; represents the spatial characteristics of sensor node i.

12. The method for diagnosing faults in a building heating, ventilation and air conditioning system according to claim 8, characterized in that: The principal component analysis technology is used to perform principal component analysis and dimension reduction on the extracted spatiotemporal features, and the expression of the spatiotemporal features after dimension reduction is obtained as follows: X r =X c W In the formula, represents the sample mean; X c is the data matrix after centralization; X r Represents the spatiotemporal feature data after dimensionality reduction; m represents the number of samples; x i represents sample data; X represents the spatiotemporal feature matrix; Represents X c The transposed matrix of .

13. The method for diagnosing faults in a building heating, ventilation and air conditioning system according to claim 8 or 12, characterized in that: Based on the spatiotemporal features after dimensionality reduction, the calculation formula for calculating the anomaly score of each sensor node in the association graph using principal component analysis technology is: Where P represents the anomaly score; Represents the samples reconstructed based on the spatiotemporal features after dimensionality reduction; The calculated anomaly score is compared with the set anomaly score threshold. When the calculated anomaly score is higher than the set anomaly score threshold, the sensor node is a potential fault point.

14. The building heating, ventilation and air conditioning system fault diagnosis system according to claim 8, characterized in that: The calculation formula for the error between the multi-source sensor data of the potential fault point and the predicted value of the potential fault point at the next time point is: yes t =V a -V p In the formula, e t V represents the error between the multi-source sensor data of the potential fault point at time t and the predicted value of the potential fault point at the next time point; a V represents the multi-source sensor data of the potential fault point at time t; p Represents the predicted value of the potential failure point at the next time point; The calculation formula for the normal range is: A=μ e ±3σ e In the formula, μ e represents the mean error; σ e represents the standard deviation of error; A represents the normal range; The calculated error is compared with the normal range. When the calculated error exceeds the normal range, the potential fault point is an abnormal point.

15. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 8 to 14 are implemented.

Images