An intelligent heating system abnormal data monitoring method

Abstract

The application discloses an intelligent heating system abnormal data monitoring method, and belongs to the technical field of data processing. The method comprises the following steps: heating system data acquisition, heating system data preprocessing, construction of a heating system abnormal data monitoring model, model parameter optimization and intelligent monitoring. The scheme combines time sequence characteristics and statistical characteristics to construct a time-varying feature map, aggregates node features using graph convolution, combines graph-level feature similarity, node-level feature similarity and adjacency matrix similarity, obtains graph similarity by comprehensively considering multi-dimensional similarity, and determines abnormal data; the final initial population is obtained by combining static reverse positions and dynamic reverse positions, a new position is generated by combining the farthest neighbor position and the average neighbor position, the global optimal position is replaced, the individual position is updated by combining the Cauchy mutation, the optimal parameters are found, the accuracy and reliability of abnormal data monitoring are improved, and the false detection and missed detection conditions are reduced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data processing, and specifically refers to an intelligent heating system abnormal data monitoring method. BACKGROUND

[0002] The heating system abnormal data monitoring method is based on deep neural network technology, analyzes the potential law in the historical heating system data, realizes intelligent monitoring of abnormal data, helps to optimize the operation efficiency of the heating system, improves the heating service quality, and guarantees the safe and stable operation of the heating system. However, in the existing heating system abnormal data monitoring method, the data of the heating system has high dimensionality, time sequence dynamics and multivariate coupling, the traditional method is difficult to analyze the nonlinear relationship between variables, and it is difficult to comprehensively evaluate the data of the heating system, resulting in the problems of decline of abnormal detection accuracy and sensitivity; in the existing heating system abnormal data monitoring method, the abnormal data in the heating system has complex spatio-temporal variation characteristics and background noise, and the traditional method based on experience or fixed parameters is difficult to accurately capture the diversity and dynamics of the abnormality, resulting in high false detection and missed detection rate of the abnormal data. SUMMARY

[0003] In view of the above, in order to overcome the defects of the prior art, the present application provides an intelligent heating system abnormal data monitoring method, which is aimed at the problems that the data of the existing heating system abnormal data monitoring method has high dimensionality, time sequence dynamics and multivariate coupling, the traditional method is difficult to analyze the nonlinear relationship between variables, and it is difficult to comprehensively evaluate the data of the heating system, resulting in the problems of decline of accuracy and sensitivity of abnormal detection, the present application is based on deep neural network, time sequence characteristics and statistical characteristics are combined to construct time-varying feature map, which can capture short-term fluctuations and represent long-term distribution characteristics, and better understand the relationship between variables; through the graph convolution operation, the node features are aggregated, and the influence of noise on abnormal detection is reduced; through the attention mechanism, the graph level feature vector is obtained, the graph level feature similarity is calculated, and the comprehensiveness of abnormal detection is improved; the node level feature similarity is calculated through the histogram, which can sensitively capture local anomalies; the adjacency matrix similarity is calculated through the node degree, which is helpful to identify abnormal conditions caused by changes in the heating network topology; the graph similarity is obtained by comprehensively considering the multi-dimensional similarity, and the abnormal data is determined, which can more comprehensively and accurately evaluate the abnormal conditions of the data, and improve the accuracy and reliability of abnormal detection; in view of the problems that the existing heating system abnormal data monitoring method has complex spatio-temporal variation characteristics and background noise in the heating system, and the traditional method based on experience or fixed parameters is difficult to accurately capture the diversity and dynamics of the abnormality, resulting in high false detection and missed detection rate of abnormal data, the present application combines the static reverse position and the dynamic reverse position to obtain the final initial population, which avoids the local concentration problem of the initial population and increases the search diversity; the farthest neighbor position and the average neighbor position are combined to generate a new position, the global optimal position is replaced, and the adaptability of the model to different types of abnormal data is improved; then the individual position is updated by combining the Cauchy mutation, which can increase the diversity of the population and find the optimal parameters represented by the individual position, thereby improving the effect of heating system abnormal data monitoring and reducing the false detection and missed detection.

[0004] The technical scheme adopted by the present application is as follows: the present application provides an intelligent heating system abnormal data monitoring method, which comprises the following steps:

[0005] Step S1: heating system data acquisition;

[0006] Step S2: heating system data preprocessing;

[0007] Step S3: constructing a heating system abnormal data monitoring model;

[0008] Step S4: model parameter optimization;

[0009] Step S5: intelligent monitoring.

[0010] Further, in step S1, the heat supply system data collection is to collect historical normal heat supply system time series data and real-time heat supply system time series data;

[0011] The historical normal heat supply system time series data and the real-time heat supply system time series data both include timestamp, temperature data, pressure data, circulating water flow rate data, equipment state data and energy consumption data.

[0012] Further, in step S2, the heat supply system data preprocessing is to perform data cleaning, data smoothing, data encoding, data normalization and dimension reduction on the collected data; the dimension reduction specifically includes the following steps:

[0013] Step S21: preliminary dimension reduction; set mutual information threshold I th , calculate the mutual information and Granger causality between each two feature time series data in the historical normal data set, filter out the feature time series data satisfying the mutual information greater than I th or the Granger causality exists, and construct a preliminary dimension reduction data set X1 of the historical normal data set based on the filtering result;

[0014] Step S22: secondary dimension reduction; use the autoencoder AE to perform secondary dimension reduction processing on X1; first map X1 to a low-dimensional latent space representation Z1 through an encoder, then reconstruct Z1 through a decoder to obtain reconstructed data , and based on X1 and , combine the reconstruction error and the gradient to calculate the importance score of each feature time series data, and finally arrange all the feature time series data in descending order of the importance score, select the first Q feature time series data, and construct a secondary dimension reduction data set of the historical normal data set; and extract the corresponding feature time series data from the real-time data set to construct a secondary dimension reduction data set of the real-time data set.

[0015] Further, in step S3, the heat supply system abnormal data monitoring model is constructed; based on the deep neural network construction model, the real-time heat supply system time series data is determined to be abnormal; specifically including the following steps:

[0016] Step S31: construct a time-varying feature map; including the following steps:

[0017] Step S311: construct a time-varying feature map of ; when the time step t is greater than or equal to 10, calculate the time series feature and the statistical feature of each feature time series data at the time step t, and combine and obtaining the merged features of each feature time series data ; for creating a node for each feature time series data in the feature time series data set, the node feature being the merged feature of the corresponding feature time series data, and calculating the Granger causality between each two feature time series data in the feature time series data set, and adding an edge between the corresponding two nodes if the Granger causality exists between the two feature time series data, to obtain the time-varying feature map at time step t ; wherein, is a secondary dimension reduction data set of the historical normal data set, t is a time step index, , and are respectively the time series feature, the statistical feature and the merged feature of the i-th feature time series data in the feature time series data set at time step t, , and are respectively the values of the i-th feature time series data in the feature time series data set at time steps t-10, t-9 and t-1, is the statistical feature of calculating the statistical feature including the mean, the maximum, the minimum, the first quartile, the second quartile and the third quartile, is the time-varying feature map of at time step t;

[0018] Step S312: constructing the time-varying feature map of ; when the time step t is greater than or equal to 10, the same method as step S311 is adopted to obtain the time-varying feature map of at time step t; wherein, is a secondary dimension reduction data set of the real-time data set, is the time-varying feature map of at time step t;

[0019] Step S32: aggregation; by a graph convolution operation, the node features in the time-varying feature maps of and are aggregated to generate new node features;

[0020] Step S33: calculating the graph similarity; from the graph-level feature similarity, the node-level feature similarity and the adjacency matrix similarity, the graph similarities of and are obtained; including the following steps:

[0021] Step S331: Calculate the graph-level feature similarity; respectively calculate the time-varying feature graph and graph-level feature vectors and , calculate the interaction score vector between and by the neural tensor network NTN, and take the mean value of the normalized interaction score vector to obtain the graph-level feature similarity and ; The calculation of the graph-level feature vector is to first calculate the average value of all node features in the time-varying feature graph, and then perform a nonlinear transformation to obtain a global average vector. Then, for each node in the time-varying feature graph, the inner product of its node feature and the global average vector is calculated, and the Sigmoid function is used to map the inner product result to the range of (0, 1) to obtain the attention weight. Then, the sum of all node features weighted by the attention weight is obtained to obtain the graph-level feature vector.

[0022] Step S332: Calculate the node-level feature similarity; calculate the interaction score between each two corresponding nodes in the time-varying feature graph and by the Sigmoid function, construct an interaction score matrix, evenly divide all values of the interaction score matrix into 5 intervals, count the frequency of scores in each interval to generate a histogram feature vector, and take the mean value of the normalized histogram feature vector to obtain the node-level feature similarity and ;

[0023] Step S333: Calculate the adjacency matrix similarity; based on the adjacency matrices and and , calculate the degree of each node in and , and then calculate the difference value between the degrees of each two corresponding nodes in and to obtain a difference vector, and take the mean value of the normalized difference vector to obtain the adjacency matrix similarity and ;

[0024] Step S334: Calculate the graph similarity; combine the graph-level feature similarity, the node-level feature similarity, and the adjacency matrix similarity by weighted combination to obtain the graph similarity and ;

[0025] ​​​​​Step S34: Abnormal data determination; set the graph similarity threshold D th , if the mutual information of the three consecutive time steps in the real-time data set is less than D th , the data is determined as abnormal data; otherwise, it is determined as normal data.

[0026] Further, in step S4, the model parameter optimization is based on a swarm intelligence algorithm to optimize the parameters in the heating system abnormal data monitoring model to find the optimal parameters; specifically including the following steps:

[0027] Step S41: Initial population; the mutual information threshold I th , the graph similarity threshold D th , the first weight ω1, the second weight ω2 and the third weight ω3 in the heating system abnormal data monitoring model establish a parameter search space, randomly initialize R individual positions in the parameter search space, the R individual positions as a population, and each individual position represents a group of parameters. The cross-entropy loss of the heating system abnormal data monitoring model based on the parameters is used as the fitness value of the corresponding individual position; for the randomly initialized R individual positions , based on the maximum value and the minimum value of the individual positions in the population, the static reverse position of each is calculated, and R static reverse positions are obtained. Based on the boundary range of the parameter search space and introducing random disturbance, the static reverse position is dynamically adjusted, the dynamic reverse position of each is calculated, and R dynamic reverse positions are obtained. The R static reverse positions and the R dynamic reverse positions are arranged in ascending order according to the fitness value, and the first R individual positions with the minimum fitness value are selected as the final initial population; wherein, is the randomly initialized rth individual position, and r is the individual position index;

[0028] Step S42: Global optimal position replacement; when the iteration number t is greater than or equal to 5, check whether the fitness value of the global optimal position in the last five iterations has changed, if not, replace the current global optimal position and the historical global optimal position ; the replacement of the current global optimal position is first to select S individual positions closest to as the neighbor positions of , and the farthest neighbor position as the farthest neighbor position of , and calculate the average neighbor position of , based on and ​generate corresponding new position , if the fitness value of is less than the fitness value of , replace with , otherwise, replace with ; replace the historical global optimal position with a new position generated based on the farthest neighbor position and the average neighbor position of , if the fitness value of is less than the fitness value of , replace with , otherwise, replace with ; wherein, and are the current global optimal position and the historical global optimal position at the hth iteration, h is the iteration number index, and are the new positions corresponding to and , respectively, and are the average neighbor positions of and , respectively, and are the farthest neighbor positions of and , respectively; Step S43: update the population; update all individual positions of the population according to the replaced current global optimal position and historical global optimal position, and combine Cauchy mutation;

[0029] Step S44: determine the optimal parameter; pre-set a fitness threshold and a maximum iteration number, when there is an individual position whose fitness value is less than the fitness threshold, the parameter represented by the individual position is taken as the optimal parameter, and a heating system abnormal data monitoring model is established based on the optimal parameter; otherwise, if the maximum iteration number is reached, return to step S41 to re-initialize the population; otherwise, increase the iteration number by 1 and return to step S42 to continue iteration.

[0030] Step S44: determine the optimal parameter; pre-set a fitness threshold and a maximum iteration number, when there is an individual position whose fitness value is less than the fitness threshold, the parameter represented by the individual position is taken as the optimal parameter, and a heating system abnormal data monitoring model is established based on the optimal parameter; otherwise, if the maximum iteration number is reached, return to step S41 to re-initialize the population; otherwise, increase the iteration number by 1 and return to step S42 to continue iteration.

[0031] ​Further, in step S5, the intelligent monitoring is to input the preprocessed historical normal heating system time series data and real-time heating system time series data into the heating system abnormal data monitoring model based on the optimal parameters for analysis, to realize intelligent heating system abnormal data monitoring; when it is determined as abnormal data, a warning is given.

[0032] The present application has the following beneficial effects by using the above scheme:

[0033] (1) In view of the problem that in the existing heating system abnormal data monitoring method, the data of the heating system has high dimensionality, time sequence dynamics and multivariate coupling, the traditional method is difficult to analyze the nonlinear relationship between variables, and it is difficult to comprehensively evaluate the data of the heating system, resulting in the decline of the accuracy and sensitivity of the abnormal detection, the present scheme based on deep neural network combines time sequence characteristics and statistical characteristics to construct time-varying feature map, which can capture short-term fluctuations and represent long-term distribution characteristics, better understand the relationship between variables; through graph convolution operation, the node features are aggregated to reduce the influence of noise on abnormal detection, so that the abnormal features are more obvious; through the attention mechanism, the graph level feature vector is obtained, the graph level feature similarity is calculated, the contribution of each node is considered comprehensively, and the comprehensiveness of the abnormal detection is improved; the node level feature similarity is calculated through the histogram, which can sensitively capture local anomalies; the adjacency matrix similarity is calculated through the node degree, which is helpful to identify abnormal conditions caused by changes in the heating network topology; the graph similarity is obtained by comprehensively considering the multi-dimensional similarity, and the abnormal data is determined, which can more comprehensively and accurately evaluate the abnormality of the data, and improve the accuracy and reliability of the abnormal detection.

[0034] (2) In view of the problem that in the existing heating system abnormal data monitoring method, the abnormal data in the heating system has complex spatio-temporal variation characteristics and background noise, and the traditional method based on experience or fixed parameters is difficult to accurately capture the diversity and dynamics of the abnormality, resulting in high false detection and missed detection rate of the abnormal data, the present scheme uses each individual position to represent a group of parameters, combines the static reverse position and the dynamic reverse position to obtain the final initial population, which avoids the local concentration problem of the initial population and increases the search diversity, and finds the parameter region suitable for heating system abnormal data monitoring faster, laying a good foundation for subsequent accurate monitoring of abnormal data; the farthest neighbor position and the average neighbor position are combined to generate a new position, and the global optimal position is replaced, which can continue to search in a wider parameter space, improve the adaptability of the model to different types of abnormal data; and the individual position is updated by combining the Cauchy mutation, which can increase the diversity of the population and help to maintain the exploration ability of the population in the parameter space, find the optimal parameters represented by the individual position, and thus improve the effect of heating system abnormal data monitoring and reduce the false detection and missed detection. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 A flowchart of an intelligent heating system abnormal data monitoring method provided by the present application is shown in the figure;

[0036] Figure 2 A flowchart of step S3 is shown in the figure;

[0037] Figure 3 A flowchart of step S4 is shown in the figure.

[0038] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, and are used to explain the present application together with embodiments of the present application, and do not constitute a limitation on the present application. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0040] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0041] Embodiment one, refer to Figure 1 The present application provides an intelligent heating system abnormal data monitoring method, which comprises the following steps:

[0042] Step S1: heating system data acquisition; collecting historical normal heating system time series data and real-time heating system time series data;

[0043] Step S2: heating system data preprocessing; data cleaning, data smoothing, data encoding, data normalization and dimension reduction are performed on the collected data;

[0044] Step S3: constructing a heating system abnormal data monitoring model; time series features and statistical features are combined to construct a time-varying feature map, graph convolution is used to aggregate node features, a graph-level feature vector is obtained based on an attention mechanism, a graph-level feature similarity is calculated, a node-level feature similarity is calculated based on a histogram, an adjacency matrix similarity is calculated based on node degree, a graph similarity is obtained by comprehensively considering multi-dimensional similarities, and abnormal data is determined;

[0045] Step S4: model parameter tuning; combining the static reverse position and the dynamic reverse position to obtain the final initial population, combining the farthest neighbor position and the average neighbor position to generate a new position, replacing the global optimal position, and updating the individual position by combining the Cauchy mutation to find the optimal parameters represented by the individual position;

[0046] Step S5: intelligent monitoring; using the abnormal data monitoring model of the heating system established based on the optimal parameters for analysis to realize intelligent abnormal data monitoring of the heating system.

[0047] Example two, see Figure 1 This example is based on the above example, in step S1, the heating system data collection is to collect historical normal heating system time series data and real-time heating system time series data;

[0048] The historical normal heating system time series data and real-time heating system time series data both include timestamp, temperature data, pressure data, circulating water flow rate data, equipment state data and energy consumption data;

[0049] The temperature data includes water supply temperature, return water temperature, water supply and return water temperature difference and environment temperature;

[0050] The pressure data includes water supply pressure, return water pressure and pressure difference;

[0051] The equipment state data includes water pump power, water pump speed and valve opening degree;

[0052] The energy consumption data includes instantaneous heat load, thermal efficiency and electric energy consumption.

[0053] Example three, see Figure 1 This example is based on the above example, in step S2, the heating system data preprocessing is to clean, smooth, encode, normalize and reduce the dimension of the collected data; the data cleaning is to remove error values, missing values and abnormal values in the data; the data smoothing is to use the KPSS test method to judge whether the data is smooth, if the KPSS test rejects the smoothness hypothesis, then first-order difference is performed; the data encoding is to use One-Hot encoding to convert category type data into numerical type data; the data normalization is to use the maximum-minimum scaling method to unify the numerical type data into the same range; the dimension reduction is based on the historical normal heating system time series data and real-time heating system time series data processed by data cleaning, data smoothing, data encoding and data normalization to construct historical normal data set and real-time data set respectively, and to reduce the dimension of the historical normal data set and real-time data set; the dimension reduction specifically includes the following steps:

[0054] Step S21: Preliminary dimensionality reduction; setting the mutual information threshold I th Calculate the mutual information and Granger causality between every two feature time series data in the historical normal dataset, and filter out those that satisfy the condition that the mutual information is greater than I. th For time series data that may exhibit Granger causality, a preliminary dimensionality-reduced dataset X1 is constructed based on the screening results of the historical normal dataset.

[0055] Step S22: Secondary dimensionality reduction; Perform secondary dimensionality reduction on X1 using an autoencoder AE; First, map X1 to a low-dimensional latent space representation Z1 using the encoder, and then reconstruct Z1 using the decoder to obtain the reconstructed data. Based on X1 and By combining reconstruction error and gradient, an importance score is calculated for each feature time series data. Finally, all feature time series data are sorted in descending order of importance score, and the top Q feature time series data are selected to construct a secondary dimensionality-reduced dataset of the historical normal dataset. ; and extract from real-time datasets Based on the corresponding feature time series data, construct a second-dimensionality-reduced dataset of the real-time dataset. .

[0056] Example 4, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S3, the construction of an abnormal data monitoring model for the heating system and the anomaly determination of real-time heating system time series data based on a deep neural network are specifically included in the following steps:

[0057] Step S31: Construct a time-varying feature graph; Heating system data has strong temporal sequence and multivariate coupling. Traditional static features cannot reflect the dynamic changes of the system, and there are physical correlations between different feature data in the heating system. It is necessary to quantify the causal dependencies between variables to improve the accuracy of anomaly detection. By extracting the time series data of the most recent 10 time steps and combining it with statistical features, it is possible to capture both short-term fluctuations and characterize long-term distribution characteristics. Based on the Granger causality graph structure, the implicit dependencies of the physical system are made explicit, enhancing the model's sensitivity to system linkage anomalies; including the following steps:

[0058] Step S311: Build The time-varying feature map; when time step t is greater than or equal to 10, calculate the time series features of each feature time series data at time step t. and statistical characteristics ,merge and This yields the merged features of each feature's time series data. ;for A node is created for each feature time series data, and the node feature is the merged feature of the corresponding feature time series data, and the calculation is The Granger causality between each two feature time series data is calculated, and if the Granger causality exists between the two feature time series data, an edge is added between the corresponding two nodes, and the graph is obtained The time-varying feature map at time step t ; wherein, is the secondary dimension reduction data set of the historical normal data set, t is the time step index, , and are respectively the time series feature, the statistical feature and the merged feature of the i-th feature time series data in at time step t, , and are respectively the values of the i-th feature time series data at time steps t-10, t-9 and t-1, is the statistical feature calculated The statistical features include mean, maximum, minimum, first quartile, second quartile and third quartile, is the time-varying feature map of at time step t;

[0059] Step S312: Constructing the time-varying feature map of When the time step t is greater than or equal to 10, the same method as step S311 is used to obtain The time-varying feature map of at time step t; wherein, is the secondary dimension reduction data set of the real-time data set, is the time-varying feature map of at time step t;

[0060] Step S32: Aggregation; single feature data cannot reflect the overall state of the heating system, the neighbor node features are weighted by graph convolution to identify local abnormal diffusion, and ReLU is used to filter noise interference and highlight significant abnormal features; through graph convolution operation, the node features in the time-varying feature maps of and are aggregated to generate new node features; the formula used is as follows:

[0061] ;

[0062] In the formula, is the new node feature of the v-th node in is the new node feature of the v-th node in is the new node feature of the v-th node in The original node feature of the u-th node in the graph, v and u are node indexes, and ReLU(•) is a ReLU activation function, The first-order neighbor set of the v-th node in the graph, including the v-th node itself, The degree of the v-th and u-th nodes in the graph plus 1, The weight matrix and bias term of the graph convolution network GCN at the m-th layer, respectively, and m is the layer index.

[0063] Step S33: Calculate the graph similarity; from the graph-level feature similarity, the node-level feature similarity, and the adjacency matrix similarity, obtain the graph similarity of The graph similarity includes the following steps:

[0064] Step S331: Calculate the graph-level feature similarity; the abnormal data of the heating system needs to be evaluated as a whole, and the global feature is weighted through the attention mechanism to identify global anomalies; calculate the graph-level feature vectors of the time-varying feature graphs and , calculate the interaction score vector between through the neural tensor network NTN, and take the mean value after normalizing the interaction score vector to obtain the graph-level feature similarity of ; To calculate the graph-level feature vector, first calculate the average value of all node features in the time-varying feature graph, and perform a nonlinear transformation to obtain the global average vector. Then, for each node in the time-varying feature graph, calculate the inner product of its node feature and the global average vector, and use the Sigmoid function to map the inner product result to the range of (0, 1) to obtain the attention weight. Then, sum all node features weighted by the attention weight to obtain the graph-level feature vector. The formula used to calculate the graph-level feature vector is as follows:

[0065] ;

[0066] ;

[0067] In the formula, are the global average vector and the graph-level feature vector of , tanh(•) is a tanh activation function, V is the number of nodes in , W2 is the feature weight, and T is the transpose operation.​​​​​​​​​​

[0068] Step S332: Calculate node-level feature similarity; minor anomalies in a single feature may indicate serious faults. Histograms are used to statistically analyze node feature differences and sensitively capture local anomalies; time-varying feature maps are calculated using the Sigmoid function. and The interaction scores between every two corresponding nodes are used to construct an interaction score matrix. All values ​​in the interaction score matrix are evenly divided into five intervals. The frequency of scores within each interval is counted, generating a histogram feature vector. The histogram feature vector is then normalized and its mean is taken to obtain the final value. and Node-level feature similarity The formula used to calculate the interaction score is as follows:

[0069] ;

[0070] In the formula, f v yes and The interaction score between the v-th nodes. yes The new node feature of the v-th node;

[0071] Step S333: Calculate the adjacency matrix similarity; changes in the heating network topology may cause false anomalies, so topology anomalies are detected through node degree differences; based on and adjacency matrix and ,calculate and Calculate the degree of each node in the algorithm. and The difference in degree between any two corresponding nodes is used to obtain a difference vector. This difference vector is then normalized and its mean is taken to obtain... and Adjacency matrix similarity The formula used to calculate the degree is as follows:

[0072] ;

[0073] In the formula, yes The degree of the v-th node in the middle. yes The v-th and u-th nodes in The corresponding value in, if If there is an edge between the v-th and u-th nodes, then ,otherwise ;

[0074] Step S334: Calculate graph similarity; integrate multi-dimensional similarity to avoid misjudgment by a single indicator; weight and combine graph-level feature similarity, node-level feature similarity, and adjacency matrix similarity to obtain... and Graph similarity The formula used is as follows:

[0075] ;

[0076] In the formula, ω1, ω2 and ω3 are the first weight, the second weight and the third weight, respectively;

[0077] Step S34: Abnormal data identification; Set graph similarity threshold D th If the real-time dataset contains three consecutive time steps Less than D th If the condition is met, the data will be classified as abnormal; otherwise, it will be classified as normal data.

[0078] By performing the above operations, this solution addresses the problems of existing methods for monitoring abnormal data in heating systems. These problems stem from the high dimensionality, temporal dynamics, and multivariate coupling of heating system data. Traditional methods struggle to analyze the nonlinear relationships between variables and comprehensively evaluate the data, leading to decreased accuracy and sensitivity in anomaly detection. This solution, based on a deep neural network, combines temporal and statistical features to construct a time-varying feature graph. This graph captures both short-term fluctuations and long-term distribution characteristics, contributing to a more comprehensive description of the heating system data and a better understanding of the relationships between variables. Graph convolution operations aggregate node features, reducing the impact of noise on anomaly detection and making anomaly features more prominent. An attention mechanism is used to obtain graph-level feature vectors, and graph-level feature similarity is calculated, comprehensively considering the contribution of each node to improve the comprehensiveness of anomaly detection. Histogram-based calculation of node-level feature similarity enables sensitive capture of local anomalies. Adjacency matrix similarity is calculated using node degree, helping to identify anomalies caused by changes in the heating network topology. Finally, a graph similarity is obtained by combining multi-dimensional similarities for anomaly judgment, enabling a more comprehensive and accurate assessment of data anomalies and improving the accuracy and reliability of anomaly detection.

[0079] Example 5, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S4, the model parameters are optimized; the parameters in the heating system abnormal data monitoring model are optimized based on a swarm intelligence algorithm to find the optimal parameters; specifically, it includes the following steps:

[0080] Step S41: initial population; each parameter has coupling, and random initialization may miss key parameter combination, static reverse position generates symmetric solution through population extreme value, avoids initial population concentrating in local area, dynamic reverse position introduces random disturbance and boundary constraint, enhances search diversity, helps to converge to better parameter region more quickly in complex heating system abnormal data monitoring environment, thereby improving model accuracy of heating system abnormal data monitoring, reducing false detection and missed detection rate; mutual information threshold I th , figure similarity threshold D th , first weight ω1, second weight ω2 and third weight ω3 establish parameter search space, R individual positions are randomly initialized in the parameter search space, the R individual positions are taken as a population, each individual position represents a group of parameters, and the cross-entropy loss of the heating system abnormal data monitoring model based on the parameters is taken as the fitness value of the corresponding individual position; for the R randomly initialized individual positions , based on the maximum value and the minimum value of the individual positions in the population, the static reverse position of each is calculated, R static reverse positions are obtained, the boundary range of the parameter search space is introduced and random disturbance is introduced, the dynamic reverse position of each is calculated, R dynamic reverse positions are obtained, the R static reverse positions and the R dynamic reverse positions are arranged in ascending order according to the size of the fitness value, and the first R individual positions with the minimum fitness value are selected as the final initial population; the formula is as follows:

[0081] ;

[0082] ;

[0083] In the formula, is the rth individual position of random initialization, r is the individual position index, and are the static reverse position and the dynamic reverse position of , lb and ub are the lower limit and the upper limit of the parameter search space, s1 and s2 are two random numbers in the range of (0, 1) without interference;

[0084] Step S42: Global Optimal Position Replacement; Due to the diversity and dynamism of anomalous data, local optima may fail to accurately reflect the anomalies in the entire heating system, and existing methods ignore the guiding value of past optimal solutions. Replacing the global optimal position helps to escape local optima and continue searching in a wider parameter space, thus better adapting to the complex spatiotemporal variations of anomalous data in the heating system. This improves the model's adaptability to different types of anomalous data and further reduces false positives and false negatives. When the iteration count t is greater than or equal to 5, check whether the fitness value of the global optimal position has changed in the last five iterations. If it has not changed, then replace the current global optimal position... and historical global optimal position Perform the replacement; find the current globally optimal position. To perform a replacement, first select with The locations of the S nearest individuals are used as The neighboring locations are selected, with the farthest neighbor's location being used as the neighboring location. farthest neighbor location and calculate Average neighbor location ,based on and generate Corresponding new position ,like The fitness value is less than The fitness value is then used. replace Otherwise, use replace ; the optimal position in the historical global context Replace with The same method, based on farthest neighbor location and average neighbor location generate Corresponding new position ,like The fitness value is less than The fitness value is then used. replace Otherwise, use replace The formula used is as follows:

[0085] ;

[0086] ;

[0087] In the formula, and is the current global optimal position and the historical global optimal position at the hth iteration, respectively, the current global optimal position is the global optimal position in the population at the hth iteration, the historical global optimal position is the global optimal position since the hth iteration, the global optimal position is the individual position with the minimum fitness value, and h is an iteration number index, and are the current global optimal position and the historical global optimal position at the hth iteration, respectively, and are the corresponding new positions, is a U-dimensional random vector generated, U being the dimension of the individual position, and are the average neighbor positions of and , respectively, and are the farthest neighbor positions of and , respectively;

[0088] Step S43: updating the population; as the iteration proceeds, the individual positions in the population need to be constantly updated to adapt to the complexity of the heating system abnormal data. The traditional method may not fully utilize the existing global optimal information during the updating process, resulting in low search efficiency. According to the global optimal position and Cauchy mutation, the global optimal information is fully utilized, and the Cauchy mutation can increase the diversity of the population, helping to maintain the exploration ability of the population in the parameter space, so that the model can better capture the diversity and dynamics of the abnormal data, thereby improving the effect of monitoring the heating system abnormal data and reducing the false detection and missed detection; according to the current global optimal position and the historical global optimal position after replacement, all individual positions of the population are updated in combination with the Cauchy mutation; the formula used is as follows:

[0089]

[0090] In the formula, and are the rth individual position at the h+1th and hth iteration, respectively, is the Cauchy mutation, and are the current global optimal position and the historical global optimal position after replacement at the hth iteration, respectively, and are the fitness values of and , respectively, is the average individual position of the population at the hth iteration, and s3 is a random number in the range of (0, 1) that does not interfere with s1 and s2;

[0091] ​Step S44: determining the optimal parameter; presetting the fitness threshold and the maximum number of iterations, when the fitness value of the individual position is less than the fitness threshold, the parameter represented by the individual position is taken as the optimal parameter, and the heating system abnormal data monitoring model is established based on the optimal parameter; otherwise, if the maximum number of iterations is reached, return to step S41 to re-initialize the population; otherwise, add 1 to the number of iterations and return to step S42 to continue iteration.

[0092] By performing the above operation, for the existing heating system abnormal data monitoring method, the abnormal data in the heating system has complex spatio-temporal variation characteristics and background noise, the traditional method based on experience or fixed parameters is difficult to accurately capture the diversity and dynamics of the abnormality, resulting in high false detection and missed detection rate of abnormal data. In the present scheme, each individual position represents a group of parameters, the static reverse position and the dynamic reverse position are combined to obtain the final initial population, which avoids the local concentration problem of the initial population and increases the search diversity, and the parameter region suitable for heating system abnormal data monitoring is found faster, laying a good foundation for subsequent accurate monitoring of abnormal data. The farthest neighbor position and the average neighbor position are combined to generate a new position, and the global optimal position is replaced, which can continue to search in a wider parameter space and improve the adaptability of the model to different types of abnormal data. Then, the individual position is updated by combining the Cauchy mutation, which can increase the diversity of the population and help to maintain the exploration ability of the population in the parameter space, find the optimal parameter represented by the individual position, and thus improve the effect of heating system abnormal data monitoring and reduce false detection and missed detection.

[0093] Embodiment six, see Figure 1 This embodiment is based on the above-mentioned embodiments. In step S5, the intelligent monitoring is performed on the preprocessed historical normal heating system time series data and real-time heating system time series data, which are input into the heating system abnormal data monitoring model established based on the optimal parameter for analysis, to realize intelligent heating system abnormal data monitoring. When it is determined that the data is abnormal, a warning is given.

[0094] It should be noted that, in this document, relational terms such as first and second and the like can only be used to distinguish one entity or action from another entity or action, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0095] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary and that changes can be made in the embodiments without departing from the spirit and scope of the application.

[0096] The above description of the application and its embodiments is not restrictive, and the embodiments shown in the drawings are only one of the embodiments of the application, and the actual structure is not limited thereto. In general, if a person skilled in the art is inspired by it, without departing from the purpose of the application, without creative design, similar structure and embodiments of the technical solution should belong to the protection scope of the application.

Claims

1. An intelligent heating system abnormal data monitoring method, characterized in that: The method comprises the following steps: Step S1: heating system data collection; Step S2: heat supply system data preprocessing; set mutual information threshold I th , perform preliminary dimension reduction and secondary dimension reduction; Step S3: constructing an abnormal data monitoring model of the heating system; combining time sequence features and statistical features to construct a time-varying feature graph, using graph convolution to aggregate node features, obtaining a graph-level feature vector based on an attention mechanism, calculating a graph-level feature similarity, calculating a node-level feature similarity based on a histogram, calculating an adjacency matrix similarity based on a node degree, weighting and combining the graph-level feature similarity, the node-level feature similarity and the adjacency matrix similarity based on a first weight ω1, a second weight ω2 and a third weight ω3, obtaining a graph similarity by comprehensively considering multi-dimensional similarities, and setting a graph similarity threshold D th to determine abnormal data; Step S4: model parameter tuning; mutual information threshold I in the heating system abnormal data monitoring model th , image similarity threshold D th , the first weight ω1, the second weight ω2 and the third weight ω3 establish a parameter search space, combine the static reverse position and the dynamic reverse position to obtain the final initial population, combine the farthest neighbor position and the average neighbor position to generate a new position, replace the global optimal position, and update the individual position by combining the Cauchy mutation to find the optimal parameter represented by the individual position; Step S5: intelligent monitoring; using the heating system abnormal data monitoring model established based on the optimal parameters for analysis to realize intelligent heating system abnormal data monitoring. 2.The abnormal data monitoring method of the intelligent heating system according to claim 1, characterized in that: In step S3, the heating system abnormal data monitoring model is constructed; based on the deep neural network model, the real-time heating system time series data is determined for abnormality; Specifically comprising the following steps: Step S31: constructing a time-varying feature map; comprising the following steps: Step S311: Build The time-varying feature map; when time step t is greater than or equal to 10, calculate the time series features of each feature time series data at time step t. and statistical characteristics ,merge and This yields the merged features of each feature's time series data. ;for For each feature time series data point, a node is created. The node's feature is the merged feature of the corresponding feature time series data. Calculation To determine the Granger causality between any two feature time series data points, if a Granger causality exists between two feature time series data points, an edge is added between the corresponding two nodes. Time-varying feature map at time step t ;in, It is a quadratic dimensionality reduction dataset of a historical normal dataset, where t is the time step index. , and They are The time series characteristics, statistical characteristics, and merged characteristics of the i-th feature time series data at time step t. , and They are The values ​​of the i-th feature time series data at time steps t-10, t-9, and t-1. It is a calculation Statistical characteristics yes Time-varying feature map at time step t; Step S312: constructing the time-varying feature map at time step t; when time step t is greater than or equal to 10, the same method as step S311 is adopted to obtain the time-varying feature map at time step t ; wherein, is a second dimension reduction data set of the real-time data set, is the time-varying feature map at time step t; Step S32: Aggregation; aggregating the node features in the time-varying feature map of and through a graph convolution operation to generate new node features; Step S33: Calculate the graph similarity; from the graph level feature similarity, the node level feature similarity and the adjacency matrix similarity, obtain the graph similarity of and ​ Step S34: Abnormal data identification; Set graph similarity threshold D th If the real-time dataset contains three consecutive time steps Less than D th If the condition is met, the data is considered abnormal; otherwise, it is considered normal. yes and The similarity between the graphs.

3. The method according to claim 2, characterized in that: In step S33, the calculation of the graph similarity specifically comprises the following steps: Step S331: Calculate the graph-level feature similarity; respectively calculate the time-varying feature graph and graph-level feature vectors and , calculate the interaction score vector between and by the neural tensor network NTN, and take the mean value after normalizing the interaction score vector to obtain the graph-level feature similarity and ; The calculation of the graph-level feature vector is to first calculate the average value of all node features in the time-varying feature graph, and then perform a nonlinear transformation to obtain a global average vector. Then, for each node in the time-varying feature graph, the inner product of its node feature and the global average vector is calculated, and the inner product result is mapped to the range of (0, 1) using the Sigmoid function to obtain the attention weight. Then, the graph-level feature vector is obtained by weighting and summing all node features according to the attention weight; wherein, and are the graph-level feature vectors of and ​​ Step S332: Calculate the node-level feature similarity; calculate the time-varying feature map through the Sigmoid function and The interaction score between each two corresponding nodes in the interaction score matrix is divided into 5 intervals, the frequency of the score in each interval is counted, a histogram feature vector is generated, and the histogram feature vector is normalized to obtain the and Node-level feature similarity ; Step S333: calculating the adjacency matrix similarity; based on and adjacency matrix and , calculating the degree of each node in and , and then calculating the difference value between the degrees of each two corresponding nodes in and , obtaining a difference value vector, and taking the mean value after normalizing the difference value vector, obtaining the adjacency matrix similarity of and ; wherein, and are the adjacency matrices of and , respectively.​ Step S334: calculating the graph similarity; combining the graph-level feature similarity, the node-level feature similarity and the adjacency matrix similarity to obtain the graph similarity and .​ 4. The method according to claim 1, characterized in that: In step S4, the model parameter optimization; based on the group intelligence algorithm, the parameters in the heating system abnormal data monitoring model are optimized to find the optimal parameters; specifically comprising the following steps: Step S41: initial population; R individual positions are randomly initialized in the parameter search space, the R individual positions are taken as a population, each individual position represents a set of parameters, and the cross-entropy loss of the heating system abnormal data monitoring model established based on the parameters is taken as the fitness value of the corresponding individual position; for the R randomly initialized individual positions , based on the maximum value and the minimum value of the individual positions in the population, the static reverse position of each is calculated, R static reverse positions are obtained, the boundary range of the parameter search space is introduced and random disturbance is introduced to dynamically adjust the static reverse positions, the dynamic reverse position of each is calculated, R dynamic reverse positions are obtained, the R static reverse positions and the R dynamic reverse positions are arranged in ascending order according to the size of the fitness value, and the first R individual positions with the minimum fitness value are selected as the final initial population; wherein, is the rth individual position randomly initialized, and r is an individual position index; Step S42: global optimal position replacement; Step S43: updating the population; according to the replaced current global optimal position and the historical global optimal position, and combining the Cauchy mutation to update the position of all individuals in the population; Step S44: determining the optimal parameters; the fitness threshold and the maximum iteration number are set in advance, when the fitness value of the individual position is less than the fitness threshold, the parameter represented by the individual position is taken as the optimal parameter, and the heating system abnormal data monitoring model is established based on the optimal parameters; otherwise, if the maximum iteration number is reached, return to step S41 to re-initialize the population; otherwise, increase the iteration number by 1 and return to step S42 for iteration.

5. The method according to claim 4, characterized in that: In step S42, the global optimal position replacement involves checking whether the fitness value of the global optimal position has changed in the last five iterations when the iteration number t is greater than or equal to 5. If there is no change, then the current global optimal position is replaced. and historical global optimal position Perform the replacement; find the current globally optimal position. To perform a replacement, first select with The locations of the S nearest individuals are used as The neighboring locations are selected, with the farthest neighbor's location being used as the neighboring location. farthest neighbor location and calculate Average neighbor location ,based on and generate Corresponding new position ,like The fitness value is less than The fitness value is then used. replace Otherwise, use replace ; the optimal position in the historical global context Replace with The same method, based on farthest neighbor location and average neighbor location generate Corresponding new position ,like The fitness value is less than The fitness value is then used. replace Otherwise, use replace ;in, and These are the current global best position and the historical global best position at the h-th iteration, respectively, where h is the iteration number index. and They are and The corresponding new position, and They are and Average neighbor locations and They are and The location of the farthest neighbor.

6. The method of claim 1, wherein the method further comprises: In step S2, the heating system data preprocessing is data cleaning, data smoothing, data encoding, data normalization and dimensionality reduction on the collected data; the dimensionality reduction specifically comprises the following steps: Step S21: preliminary dimension reduction; set mutual information threshold I th , calculate the mutual information and Granger causality between each two feature time series data in the historical normal data set, and screen out the feature time series data satisfying mutual information greater than I th or Granger causality exists, based on the screening result, construct the preliminary dimension reduction data set X1 of the historical normal data set; Step S22: secondary dimension reduction; using the autoencoder AE to perform secondary dimension reduction on X1; first mapping X1 to a low-dimensional latent space representation Z1 through the encoder, and then reconstructing Z1 through the decoder to obtain the reconstructed data , and combining the reconstruction error and the gradient to calculate the importance score of each feature time series data, and finally arranging all the feature time series data in descending order of the importance score, selecting the top Q feature time series data, and constructing a secondary dimension reduction data set of the historical normal data set ; and extracting the corresponding feature time series data from the real-time data set , and constructing a secondary dimension reduction data set of the real-time data set .

7. The method of claim 1, wherein the method further comprises: In step S1, the heating system data collection is to collect historical normal heating system time series data and real-time heating system time series data; the historical normal heating system time series data and real-time heating system time series data both include timestamp, temperature data, pressure data, circulating water flow rate data, equipment state data and energy consumption data. 8.The method of claim 1, wherein the method further comprises: In step S5, the intelligent monitoring is to input the preprocessed collected historical normal heating system time series data and real-time heating system time series data into the heating system abnormal data monitoring model established based on the optimal parameters for analysis to realize intelligent heating system abnormal data monitoring; When it is determined that the data is abnormal, a warning is given.

Images