Heat pump system state anomaly detection method based on depth auto-encoder

Abstract

The invention discloses a heat pump system state anomaly detection method based on a depth auto-encoder, and the method comprises the steps: collecting compressor data, and carrying out the standardization processing to construct a multi-dimensional time sequence; a spatial-temporal feature extraction depth auto-encoder with a thermodynamic coupling attention mechanism is constructed, coupling attention is utilized to calculate physical parameter coupling strength weights to extract spatial features, time features are extracted in combination with a long and short-term memory network, and normal state data are reconstructed and predicted through a decoder after fusion; residual vectors of predicted normal state data and original data are calculated, and a weighted mahalanobis distance is calculated by using a covariance matrix to generate an abnormal score; and constructing a sliding probability distribution model based on historical normal data, calculating a current score occurrence probability, and comparing the current score occurrence probability with a preset threshold to output an anomaly detection result. According to the method, a multi-physical parameter space coupling relationship and a time evolution rule are captured through a thermodynamic coupling attention mechanism, and the anomaly detection accuracy and robustness are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of anomaly detection, and particularly relates to a heat pump system state anomaly detection method based on a deep autoencoder. BACKGROUND

[0002] With the transformation of global energy structure and the promotion of the "double carbon" target, heat pump systems, as an efficient and energy-saving heating and refrigeration technology, have been widely used in the fields of building, industry and regional heating. In particular, heat pump systems using magnetic suspension compressor technology have gradually become the preferred solution for large commercial buildings and industrial parks due to their advantages of oil-free operation, low noise and high efficiency. However, during the long-term operation of heat pump systems, abnormal conditions such as refrigerant leakage, compressor performance degradation, and heat exchange efficiency decline may occur. If not detected and handled in a timely manner, it will not only lead to a significant reduction in system energy efficiency, but also may cause equipment damage and even safety accidents.

[0003] Traditional heat pump system anomaly detection methods mainly rely on rule-based threshold judgment. This kind of method sets upper and lower threshold values for key parameters such as pressure, temperature and current, and triggers an alarm when the measured value exceeds the pre-set range. However, the heat pump system is a highly coupled thermodynamic system, and there is a complex nonlinear correlation between various physical parameters. Simply relying on single-parameter threshold judgment can easily produce a large number of false positives and false negatives. At the same time, the normal parameter range under different operating conditions varies significantly, and fixed thresholds are difficult to adapt to dynamic operating conditions such as seasonal changes and load fluctuations.

[0004] Although statistical methods such as principal component analysis can perform multi-parameter analysis, such linear methods cannot effectively capture the nonlinear thermodynamic coupling relationships in heat pump systems, and only focus on the spatial correlation of data, ignoring the time evolution characteristics of operating states, making it difficult to identify gradual failure and performance degradation and other abnormal patterns. Supervised learning methods such as support vector machines face the problem of high cost of obtaining abnormal samples and unbalanced sample distribution, and lack the ability to generalize to new types of faults.

[0005] Deep learning methods such as long short-term memory networks have strong time series modeling capabilities, but simply using LSTM mainly focuses on time dimension features, and lacks effective modeling of the spatial coupling relationships between multiple physical parameters in heat pump systems. As an unsupervised learning model, autoencoders are suitable for application scenarios where normal data is abundant but abnormal samples are scarce, but traditional autoencoders are not specifically designed for the physical characteristics of heat pump systems, and the unified feature extraction mechanism for different types of parameters cannot take into account the characteristics of each parameter. In addition, existing autoencoder anomaly detection methods usually use Euclidean distance to measure reconstruction error, ignoring the contribution differences of different physical parameters and the correlation between parameters, which can easily lead to misjudgment.

[0006] In summary, the existing methods have the following shortcomings: first, the space-time coupling characteristics of the heat pump system operation data cannot be effectively modeled simultaneously; second, the abnormal evaluation index is simple in design and does not fully utilize the statistical correlation information in the multi-dimensional residual vector; third, there is a lack of adaptive mechanism for the dynamic changes of the normal state distribution under different working conditions. Therefore, it is urgent to develop an intelligent abnormality detection method that can deeply integrate space-time feature extraction, physical coupling relationship modeling and adaptive probability evaluation. SUMMARY

[0007] Therefore, the present application provides a heat pump system state abnormality detection method based on a deep autoencoder, which aims to build a space-time feature extraction deep autoencoder network with a thermodynamic coupling attention mechanism, which can simultaneously capture the spatial coupling relationship and time evolution law between multiple physical parameters, and realize accurate abnormal state recognition by combining weighted Mahalanobis distance and a sliding probability distribution model, thereby improving the accuracy, robustness and adaptability of heat pump system abnormality detection and ensuring safe and stable operation of the system.

[0008] To achieve the above-mentioned purpose, the present application provides a heat pump system state abnormality detection method based on a deep autoencoder, which comprises the following steps: B1: Collecting the compressor discharge pressure, compressor suction pressure, discharge temperature, suction temperature, magnetic suspension bearing vibration, motor current and electronic expansion valve opening degree data of the heat pump system, standardizing the data in chronological order, and constructing a multi-dimensional time series input data; B2: Building a space-time feature extraction deep autoencoder network with a thermodynamic coupling attention mechanism, calculating the coupling strength weight between physical parameters to extract spatial features by using the thermodynamic coupling attention mechanism, combining a long short-term memory network to extract time evolution features, fusing spatial features and time features to map into hidden variables, and reconstructing the predicted normal state data through a decoder network; B3: Calculating the residual vector between the predicted normal state data and the multi-dimensional time series input data, calculating the weighted Mahalanobis distance of the residual vector by using the covariance matrix obtained from the training data set, quantifying the degree of deviation of the current data point from the normal joint distribution, and generating a comprehensive abnormality score; B4: Building a sliding probability distribution model based on historical normal data, substituting the comprehensive abnormality score into the sliding probability distribution model to calculate the occurrence probability of the current score, comparing it with a preset confidence threshold, and outputting the final heat pump system abnormality detection result.

[0009] As a further improved method of the present application: Optionally, the B1 step further comprises: A data acquisition network for the heat pump system is established, connecting sensors of the compressor, magnetic levitation bearing, motor, and electronic expansion valve within the system. A data sampling frequency is set to acquire real-time operating data of the heat pump system. This real-time operating data includes compressor discharge pressure, compressor suction pressure, discharge temperature, suction temperature, magnetic levitation bearing vibration, motor current, and electronic expansion valve opening. The operating data is then standardized using the Z-score standardization method to eliminate dimensional differences between different physical quantities, resulting in standardized operating data. This standardized operating data is then arranged chronologically to construct a multi-dimensional time series input data. Represented as a vector , Standardized data includes seven dimensions: compressor discharge pressure, compressor suction pressure, discharge temperature, suction temperature, magnetic levitation bearing vibration, motor current, and electronic expansion valve opening.

[0010] Optionally, step B2 further includes: A deep autoencoder network for spatiotemporal feature extraction with a thermodynamically coupled attention mechanism is constructed, wherein the deep autoencoder network includes an encoder and a decoder; The encoder includes a feature embedding module, a thermodynamic coupling attention module, a long short-term memory network module, and a feature fusion mapping module. The feature embedding module is used to map multidimensional time series input data into a high-dimensional feature embedding representation. The thermodynamic coupling attention module is used to calculate the coupling strength weights between physical parameters and extract spatial features. The long short-term memory network module is used to extract temporal evolution features. The feature fusion mapping module is used to fuse spatial features and temporal features and map them into latent variables. The feature embedding module receives the time Multidimensional time series input data It is mapped to a feature embedding matrix through a fully connected layer. ; The thermodynamically coupled attention module receives the time Feature Embedding Matrix The coupling strength weights between each physical parameter dimension are calculated, and these coupling strength weights are represented by the attention score matrix. The calculation formula is: ; in, Indicates time The attention score matrix, where softmax represents the normalized exponential function. Indicates time The query matrix, Indicates time a key matrix, denotes the time , the transpose of the key matrix, denotes the dimension number of the key vector, used to scale the dot product result; the query matrix and the key matrix are obtained by performing linear transformation on the feature embedding matrix , the query matrix is calculated as , and the key matrix is calculated as wherein denotes the query weight matrix, denotes the key weight matrix; The input data is weighted according to the attention score matrix to obtain the spatial feature , and the calculation formula of the spatial feature is: ; wherein, denotes the spatial feature at the time , and denotes the value matrix at the time , the value matrix is obtained by performing linear transformation on the feature embedding matrix , and the calculation formula of the value matrix is wherein denotes the value weight matrix; The long short-term memory network module receives multi-dimensional time series input data at consecutive , …, , …, extracts time evolution features , and the time evolution features are represented by the hidden state of the long short-term memory network; The feature fusion mapping module concatenates the spatial feature and the time evolution feature , and maps them to the latent variable through a fully connected layer, and the calculation formula of the latent variable is: ; wherein, denotes the latent variable at the time , and denotes the fusion weight matrix, denotes the feature vector obtained by concatenating the spatial feature and the time evolution feature , and a fusion bias vector; The decoder receives the latent variable , reconstructs the predicted normal state data through a multi-layer fully connected network , the predicted normal state data has the same dimension as the multi-dimensional time series input data .

[0011] Optionally, the B3 step further comprises: calculating the residual vector between the predicted normal state data at time point and the multi-dimensional time series input data . The residual vector is calculated according to the following formula: ; Wherein, represents the residual vector at time point . The weighted Mahalanobis distance of the residual vector is calculated using the covariance matrix , which is obtained by calculating the covariance of the residual vectors of all normal state samples in the training data set, and the weighted Mahalanobis distance is used as the comprehensive anomaly score , and the calculation formula is as follows: ; Wherein, represents the comprehensive anomaly score at time point , represents the transpose of the residual vector at time point , represents the inverse matrix of the covariance matrix , and the weighted Mahalanobis distance quantifies the degree of deviation of the current data point from the normal joint distribution in the training phase, which can consider the correlation and scale difference between variables compared with the Euclidean distance. Optionally, the B4 step further comprises: Based on the historical normal operation data, a sliding probability distribution model is constructed, which fits the probability distribution using the comprehensive anomaly score samples in the sliding time window, specifically, the comprehensive anomaly scores of the last

[0012] consecutive time points before time point , , , , , , wherein represents the length of the sliding window, represents the comprehensive anomaly score confirmed as the normal state at time point , and the collected comprehensive anomaly scores are used to fit the probability distribution of the sliding probability distribution model.​​ The probability distribution model is obtained by fitting the chi-square distribution of the comprehensive abnormal scores of normal states. Time Comprehensive abnormal score Substitute the fitted probability distribution model into the equation to calculate the probability of the current score occurring. The probability of occurrence refers to the observed comprehensive anomaly score being greater than or equal to [a certain value] under normal operating conditions. The probability of; The probability of occurrence Compared with the preset confidence threshold When comparing, At that time, it is determined that the heat pump system is at time... In an abnormal state, when At that time, it is determined that the heat pump system is at time... It is in a normal state.

[0013] The anomaly detection mechanism based on dynamic distribution established in this step can adapt to the time-varying characteristics of the heat pump system's operating state. The normal state characteristics of a heat pump system change under different operating conditions. For example, the normal fluctuation range of parameters such as system pressure and temperature varies under different ambient temperatures and load conditions. Traditional fixed threshold methods cannot adapt to such dynamic changes and are prone to false alarms or missed alarms when operating conditions change. This step, by using a sliding time window to dynamically update the probability distribution model, can track the changes in the statistical characteristics of the system's normal operating state in real time. This allows the anomaly detection threshold to adaptively adjust with operating conditions, thereby improving the accuracy and robustness of anomaly detection and reducing false alarms and missed alarms caused by changes in operating conditions.

[0014] Compared with the prior art, the present invention has at least the following beneficial effects: This invention constructs a deep autoencoder network with a thermodynamic coupling attention mechanism, which can simultaneously and effectively extract the spatiotemporal coupling features of heat pump system operating data. The thermodynamic coupling attention module automatically learns the coupling strength weights between physical parameters such as compressor discharge pressure and discharge temperature, and suction pressure and suction temperature, fully capturing the inherent thermodynamic constraints of the heat pump system. The long short-term memory network module extracts the temporal evolution features of the system state. By deeply fusing spatial and temporal features, this invention significantly improves the ability to identify complex anomaly patterns compared to traditional methods that only focus on single-dimensional features.

[0015] This invention employs weighted Mahalanobis distance as the method for calculating the comprehensive anomaly score. By introducing the covariance matrix, it fully considers the correlation and scale differences between the residuals of various physical parameters, enabling a more accurate quantification of the degree to which the current data points deviate from the normal joint distribution. When faced with situations where some parameters have slight deviations but the overall distribution exhibits an abnormal correlation pattern, Mahalanobis distance can capture this abnormal correlation, while Euclidean distance may miss it. The invention uses singular value decomposition to calculate the inverse of the covariance matrix and performs regularization, effectively improving the stability of the numerical calculation and giving the anomaly detection system better robustness and reliability. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for detecting abnormal states in a heat pump system based on a deep autoencoder, according to an embodiment of the present invention. Figure 2 The following is a schematic diagram of thermodynamic coupling attention weights: (a) attention weight distribution under normal operating conditions, (b) attention weight distribution under abnormal operating conditions with refrigerant leakage. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0018] Example 1: A method for detecting abnormal states in a heat pump system based on a deep autoencoder, such as... Figure 1 As shown, it includes the following steps: B1: Collect data on compressor discharge pressure, compressor suction pressure, discharge temperature, suction temperature, magnetic levitation bearing vibration, motor current, and electronic expansion valve opening of the heat pump system. Standardize these data in chronological order to construct a multi-dimensional time-series input data set, including: A data acquisition network for the heat pump system is established, connecting sensors of the compressor, magnetic bearing, motor, and electronic expansion valve within the system. A data sampling frequency is set to acquire real-time operating data of the heat pump system. This real-time operating data includes compressor discharge pressure, compressor suction pressure, discharge temperature, suction temperature, magnetic bearing vibration, motor current, and electronic expansion valve opening. In this embodiment, the data sampling frequency is set to 1Hz. Optionally, during data acquisition, when a sensor malfunction or data loss is detected, an adaptive interpolation compensation method is used to repair the missing data: first, the time span of the missing data is determined; if the missing duration is less than 3 seconds, a linear interpolation method is used. Interpolation methods: If the missing duration is between 3 and 10 seconds, cubic spline interpolation is used, employing a cubic polynomial constructed with 5 normal data points before and after the missing point for smooth interpolation; if the missing duration exceeds 10 seconds, the data for that time period is marked as unusable and excluded from subsequent anomaly detection to avoid misjudgment due to accumulated interpolation errors; the running data is standardized using Z-score standardization to eliminate dimensional differences between different physical quantities, resulting in standardized running data; the standardized running data is then arranged chronologically to construct multidimensional time series input data, where the multidimensional time series input data is at time... Represented as a vector , Standardized data includes seven dimensions: compressor discharge pressure, compressor suction pressure, discharge temperature, suction temperature, magnetic levitation bearing vibration, motor current, and electronic expansion valve opening.

[0019] B2: Construct a deep autoencoder network for spatiotemporal feature extraction with a thermodynamic coupled attention mechanism. Utilize this mechanism to calculate the coupling strength weights between physical parameters to extract spatial features. Combine this with a long short-term memory network to extract temporal evolution features. Merge spatial and temporal features and map them as latent variables. Reconstruct the predicted normal state data through a decoder network, including: A deep autoencoder network for spatiotemporal feature extraction with a thermodynamically coupled attention mechanism is constructed, wherein the deep autoencoder network includes an encoder and a decoder; The encoder includes a feature embedding module, a thermodynamic coupling attention module, a long short-term memory network module, and a feature fusion mapping module. The feature embedding module is used to map multidimensional time series input data into a high-dimensional feature embedding representation. The thermodynamic coupling attention module is used to calculate the coupling strength weights between physical parameters and extract spatial features. The long short-term memory network module is used to extract temporal evolution features. The feature fusion mapping module is used to fuse spatial features and temporal features and map them into latent variables. The feature embedding module receives the time Multidimensional time series input data It is mapped to a feature embedding matrix through a fully connected layer. In this embodiment, the multidimensional time series input data after feature embedding The dimensions are 7×64; The thermodynamically coupled attention module receives the time Feature Embedding Matrix The coupling strength weights between each physical parameter dimension are calculated, and these coupling strength weights are represented by the attention score matrix. The calculation formula is: ; in, Indicates time The attention score matrix, such as Figure 2 As shown, softmax represents the normalized exponential function. Indicates time The query matrix, Indicates time The key matrix, Indicates time The transpose of the key matrix, The dimension of the key vector is used to scale the dot product result; the query matrix Bond matrix By analyzing the feature embedding matrix The query matrix is ​​obtained by performing a linear transformation. The calculation formula is Key matrix The calculation formula is ,in This represents the query weight matrix. This represents the key weight matrix; in this embodiment, and All dimensions are set to 64×64; Optionally, to enhance the model's ability to learn the physical constraints of the heat pump system, a physical prior constraint mechanism is introduced into the thermodynamic coupling attention module: a physical constraint mask matrix is ​​constructed by predefining the strong coupling relationships between physical parameters based on the laws of thermodynamics. For parameter pairs with strong physical coupling ,set up Otherwise set ,in In this embodiment, the weighting coefficients are weakly coupled. Set to 0.3; introduce physical constraints when calculating the attention score matrix, and the revised calculation formula is as follows: , where ⊙ denotes element-wise multiplication; Based on the attention score matrix The input data is weighted to obtain spatial features. The formula for calculating the spatial features is: in, Indicates time Spatial characteristics, Indicates time The value matrix, the value matrix By analyzing the feature embedding matrix The value matrix is ​​obtained by performing a linear transformation. The calculation formula is ,in This represents the value weight matrix; in this embodiment, The dimension is 64×64; to facilitate subsequent fusion with temporal features, Flattened, it becomes a one-dimensional vector with dimensions 448; The long short-term memory network module receives continuous Multidimensional time series input data at each time point , … Extracting temporal evolution features In this embodiment, The value is set to 20, and the temporal evolution features are represented by the hidden states of a Long Short-Term Memory network. In this embodiment, the number of hidden layer units is set to 128. Take the hidden state at the last moment; The feature fusion mapping module will integrate spatial features and temporal evolution characteristics The data is concatenated and mapped to latent variables through a fully connected layer. The formula for calculating the latent variable is: ; in, Indicates time Latent variables, Represents the fusion weight matrix. Indicates spatial features and temporal evolution characteristics The feature vector obtained by concatenation This represents the fusion bias vector, in this embodiment The dimensions are set to 32×576. The dimension is 32×1; The decoder receives hidden variables. The predicted normal state data is reconstructed through a multi-layer fully connected network. The predicted normal state data and multidimensional time series input data Having the same dimensions, in this embodiment, the decoder includes a three-layer fully connected network. The first fully connected layer maps the 32-dimensional latent variables to 64 dimensions; the second fully connected layer maps the 64 dimensions to 32 dimensions; and the third fully connected layer maps the 32 dimensions to 7 dimensions to obtain the predicted normal state data.

[0020] B3: Calculate the residual vector between the predicted normal state data and the multidimensional time series input data. Using the covariance matrix obtained from the training dataset, calculate the weighted Mahalanobis distance of the residual vector to quantify the degree to which the current data point deviates from the normal joint distribution, and generate a comprehensive anomaly score, including: Calculation time Predicted normal state data With multidimensional time series input data The residual vector between The formula for calculating the residual vector is: ; in, Indicates time The residual vector; Using the covariance matrix Calculate the residual vector The weighted Mahalanobis distance, the covariance matrix The weighted Mahalanobis distance is obtained by calculating the covariance of the residual vectors of all normal state samples in the training dataset, and is used as the comprehensive anomaly score. The calculation formula is: ; in, Indicates time The overall abnormal score, Indicates time The transpose of the residual vector, Represents the covariance matrix The inverse matrix of the covariance matrix is ​​calculated using the weighted Mahalanobis distance, which quantifies the degree to which the current data points deviate from the normal joint distribution during the training phase. Compared to Euclidean distance, it can consider the correlation and scale differences between variables. In this embodiment, the inverse matrix of the covariance matrix is ​​calculated using the singular value decomposition method to improve numerical stability. When the covariance matrix is ​​close to singular, values ​​less than a threshold are considered. The singular values ​​are regularized in this embodiment. Set as .

[0021] B4: Construct a sliding probability distribution model based on historical normal data, substitute the comprehensive anomaly score into the sliding probability distribution model to calculate the probability of the current score occurring, compare it with the preset confidence threshold, and output the final anomaly detection result of the heat pump system, including: A sliding probability distribution model is constructed based on historical normal operation data, and the sliding probability distribution model fits a probability distribution by using comprehensive abnormal score samples in a sliding time window. Specifically, comprehensive abnormal scores of previous continuous time points are collected , , …, , wherein denotes a sliding window length, and in the embodiment, the sliding window length is set to 500, denotes a comprehensive abnormal score confirmed as a normal state at time point ; the collected normal state comprehensive abnormal scores are subjected to chi-square distribution fitting to obtain a fitted probability distribution model; The comprehensive abnormal score at time point is substituted into the fitted probability distribution model to calculate an occurrence probability of the current score, and the occurrence probability represents a probability of observing a comprehensive abnormal score greater than or equal to under a normal operation state; The occurrence probability is compared with a preset confidence threshold , when , it is determined that the heat pump system is in an abnormal state at time point , and when , it is determined that the heat pump system is in a normal state at time point . In the embodiment, the confidence threshold is set to 0.05.

[0022] It should be noted that the above-mentioned serial numbers of the embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments. Moreover, the terms “include”, “contain” or any other variant thereof in the present text are intended to cover non-exclusive inclusion, so that the process, device, article or method including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such process, device, article or method. Without more limitations, the element defined by the statement “including a…” does not exclude the presence of another identical element in the process, device, article or method including the element.

[0023] Those skilled in the art can clearly understand the above-mentioned embodiment method can be realized by means of software and the necessary general hardware platform, of course, also can be through hardware, but in many cases the former is the better embodiment. Based on such understanding, the technical solutions of the present application essentially or say the part of the contribution to the prior art can be embodied in the form of software products, the computer software product is stored in a storage medium (such as ROM / RAM, magnetic disc, optical disc) as described above, including a number of instructions to make a terminal device (may be a mobile phone, computer, server, or network equipment, etc.) executes the method described in various embodiments of the present application.

[0024] The above is only the preferred embodiment of the present application, not therefore limit the patent scope of the present application, any equivalent structure or equivalent flow transformation made by using the content of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A deep autoencoder-based heat pump system state anomaly detection method, characterized by, The method comprises the following steps: B1: Collecting the compressor discharge pressure, compressor suction pressure, discharge temperature, suction temperature, magnetic bearing vibration, motor current and electronic expansion valve opening degree data of the heat pump system, standardizing the data in chronological order, and constructing a multi-dimensional time series input data; B2: Constructing a deep autoencoder network with thermodynamic coupling attention mechanism for spatio-temporal feature extraction, calculating the coupling strength weight between physical parameters to extract spatial features using the thermodynamic coupling attention mechanism, combining the long short-term memory network to extract time evolution features, fusing spatial features and time features to map into latent variables, and reconstructing the predicted normal state data through the decoder network; B3: Calculate the residual vector between the predicted normal state data and the multi-dimensional time series input data, calculate the weighted Mahalanobis distance of the residual vector using the covariance matrix obtained from the training data set, quantify the degree of deviation of the data points from the normal joint distribution, and generate a comprehensive anomaly score; B4: Based on the historical normal data, a sliding probability distribution model is constructed, the comprehensive anomaly score is substituted into the sliding probability distribution model to calculate the occurrence probability of the current score, compared with the preset confidence threshold, and the final heat pump system anomaly detection result is output.

2. The deep autoencoder-based heat pump system state anomaly detection method according to claim 1, characterized by, The step B1 comprises: A heat pump system data acquisition network is established, sensors of a compressor, a magnetic suspension bearing, a motor and an electronic expansion valve in the heat pump system are connected, a data sampling frequency is set to obtain real-time operation data of the heat pump system, the real-time operation data includes compressor discharge pressure, compressor suction pressure, discharge temperature, suction temperature, magnetic suspension bearing vibration, motor current and electronic expansion valve opening degree; the operation data is standardized, Z-score standardization method is used to standardize each dimension data, dimension difference between different physical quantities is eliminated, and standardized operation data is obtained; the standardized operation data is arranged in time sequence, multi-dimensional time sequence input data is constructed, the multi-dimensional time sequence input data at time is expressed as a vector , contains seven dimensions of standardized data of compressor discharge pressure, compressor suction pressure, discharge temperature, suction temperature, magnetic suspension bearing vibration, motor current and electronic expansion valve opening degree.

3. The deep autoencoder-based heat pump system state anomaly detection method according to claim 2, characterized by, The step B2 comprises: Constructing a deep autoencoder network with thermodynamic coupling attention mechanism for spatio-temporal feature extraction, the deep autoencoder network comprising an encoder and a decoder; The encoder comprises a feature embedding module, a thermodynamic coupling attention module, a long short-term memory network module and a feature fusion mapping module; the feature embedding module is used to map the multi-dimensional time series input data into high-dimensional feature embedding representation, the thermodynamic coupling attention module is used to calculate the coupling strength weight between physical parameters and extract spatial features, the long short-term memory network module is used to extract time evolution features, and the feature fusion mapping module is used to fuse spatial features and time features and map into latent variables; The feature embedding module receives a multi-dimensional time series input data at a time instant and maps it to a feature embedding matrix through a fully connected layer; The thermodynamic coupling attention module receives a feature embedding matrix at a time point , calculates coupling strength weights between each physical parameter dimension, and the coupling strength weights are represented by an attention score matrix The calculation formula of the attention score matrix is: ； wherein, denotes the attention score matrix at time , softmax denotes the normalized exponential function, denotes the query matrix at time , denotes the key matrix at time , denotes the transpose of the key matrix at time , denotes the dimension number of the key vector, used to scale the dot product result; the query matrix and the key matrix are obtained by linear transformation on the feature embedding matrix , the calculation formula of the query matrix is , and the calculation formula of the key matrix is , wherein denotes the query weight matrix, denotes the key weight matrix; According to the attention score matrix The input data is weighted to obtain a spatial feature The calculation formula of the spatial feature is ； wherein, denotes a spatial feature at time denotes a value matrix at time is obtained by linear transformation on the feature embedding matrix The calculation formula of the value matrix denotes a value weight matrix;​​​​ The long short-term memory network module receives continuous Multidimensional time series input data at each time point , … Extracting temporal evolution features The temporal evolution features are represented by the hidden states of a long short-term memory network; The feature fusion mapping module fuses the spatial features and the time evolution features to splice and map to the latent variable z through a full connection layer , and a calculation formula of the latent variable z is: ； wherein, denotes the time instant a hidden variable, denotes a fusion weight matrix, denotes a feature vector obtained by concatenating spatial features and temporal evolution features , denotes a fusion bias vector; The decoder receives the latent variable , reconstructs predicted normal state data through a multi-layer fully connected network , the predicted normal state data has the same dimension as the multi-dimensional time series input data .

4. The deep autoencoder-based heat pump system state anomaly detection method according to claim 3, characterized by, The step B3 comprises: a time instant of a predicted normal state data between the multi-dimensional time series input data a residual vector , the residual vector being calculated according to the formula: ； wherein represents the residual vector at time ; Utilizing a covariance matrix Computing a residual vector A weighted Mahalanobis distance of the residual vector The weighted Mahalanobis distance is obtained by computing the covariance of the residual vectors of all normal state samples in the training data set, as a comprehensive abnormality score The formula is: ； wherein represents the integrated abnormal score at time represents the transpose of the residual vector at time represents the inverse of the covariance matrix The weighted Mahalanobis distance quantifies the degree to which the current data point deviates from the normal joint distribution in the training phase, and can consider the correlation and scale difference between variables compared to the Euclidean distance.​​ 5. The deep autoencoder-based heat pump system state anomaly detection method according to claim 4, characterized by, The step B4 comprises: A sliding probability distribution model is constructed based on historical normal operation data, and the sliding probability distribution model fits a probability distribution using comprehensive abnormal score samples in a sliding time window. Specifically, comprehensive abnormal scores at a collection time point and a previous continuous time points , , are collected, where represents a sliding window length, represents a comprehensive abnormal score confirmed as a normal state at a time point ; and chi-square distribution fitting is performed on the collected normal state comprehensive abnormal scores to obtain a fitted probability distribution model. Time Comprehensive abnormal score Substitute the fitted probability distribution model into the equation to calculate the probability of the current score occurring. The probability of occurrence refers to the observed comprehensive anomaly score being greater than or equal to [a certain value] under normal operating conditions. The probability of; Probability of occurrence with a preset confidence threshold is compared, when the heat pump system is determined to be in an abnormal state at the time point in time, when the heat pump system is determined to be in a normal state at the time point in time.

Images