LSTM-based thermal storage tank health state online early warning method and system

Abstract

The application relates to the technical field of industrial equipment health monitoring and fault early warning, in particular to a heat storage tank health state online early warning method and system based on an LSTM. The method comprises the following steps: obtaining a corresponding feature vector according to each sample in a historical sample set, constructing a fault feature initial graph according to the feature vector, establishing a dynamic association rule of the local density of nodes and the Gaussian kernel width for dynamically updating the node similarity weight to obtain a fault feature graph; updating the label of the historical sample set according to the fault feature graph, training a weighted LSTM model according to the updated historical sample set to obtain a fault diagnosis model; collecting real-time operation data of the heat storage tank, inputting the processed real-time operation data into the fault diagnosis model, and outputting a heat storage tank health state diagnosis result. The application can realize accurate and real-time heat storage tank health state early warning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial equipment health monitoring and fault early warning, and particularly relates to a heat storage tank health state online early warning method and system based on LSTM. BACKGROUND

[0002] As the core equipment of heat and power cogeneration and new energy storage systems, the heat storage tank undertakes the key functions of heat storage and load regulation, and its operation state directly affects the energy utilization efficiency and system stability. However, the heat storage tank faces various faults such as tank corrosion, temperature stratification damage and pipeline fouling during operation, and the fault evolution has the characteristics of concealment and time sequence accumulation, so the health state cannot be monitored in real time.

[0003] The traditional method has two defects: first, the heat storage tank in the industrial scene has a long normal operation time, and the fault cases are few and scattered in type. Although the weighted LSTM model is good at learning time sequence features, it has high dependence on sample size, and its performance is limited in the scene with few fault samples. Second, although the existing label propagation algorithm can expand data using a small number of known label samples, the kernel width parameter is fixed, which is difficult to adapt to the difference in sample distribution, resulting in insufficient label propagation accuracy. Therefore, the existing technology has the technical problems of weak model generalization ability and low early warning accuracy caused by the lack of fault samples and insufficient label propagation accuracy in the health state early warning of the heat storage tank. SUMMARY

[0004] In order to solve the above technical problems, the present application provides a heat storage tank health state online early warning method and system based on LSTM, which expands fault samples through a label propagation algorithm with adaptive kernel width, and combines a weighted LSTM model to capture time sequence correlation features, so as to realize accurate and real-time health state early warning.

[0005] In a first aspect, the application provides an LSTM-based online warning method for the health state of a heat storage tank, which adopts the following technical scheme: the LSTM-based online warning method for the health state of a heat storage tank comprises the following steps: collecting multi-dimensional time series data of the operation of the heat storage tank and performing preprocessing to obtain a historical sample set containing known label samples and label-free samples; constructing a fault feature initial graph according to the feature vectors of each sample in the historical sample set, establishing a dynamic association rule of the local density of the node and the Gaussian kernel width, dynamically updating the similarity weight of the node, and obtaining a fault feature graph; the nodes in the fault feature graph represent samples, and the edges represent the neighbor relationship between the samples; according to the fault feature graph, the label propagation algorithm is used to update the label of the label-free sample in the historical sample set to obtain an updated historical sample set; based on the updated historical sample set, a sample type label and a propagation probability for distinguishing whether a sample is a real label sample or a propagated label sample are introduced, a weighted cross-entropy loss function is constructed, a weighted LSTM model is trained, and a fault diagnosis model is obtained; wherein the propagation probability is the probability value of a propagated label sample obtaining its final label in the label propagation process; the loss function gives each sample in the historical sample set a weight, and the weight is determined according to the type label and the propagation probability of the sample: the weight of the sample as a real label sample is 1; the weight of the sample as a propagated label sample is the corresponding propagation probability; the real-time operation data of the heat storage tank is collected, the same preprocessing as the historical sample is performed, and then the data is input into the fault diagnosis model to output the diagnosis result of the health state of the heat storage tank.

[0006] Optionally, the dynamic association rule is: calculating the local density of the node; taking the average of the local densities of all nodes as the overall distribution density; obtaining the best Gaussian kernel width corresponding to the node by comparing the local density of the node with the overall distribution density; and calculating the similarity weight of the node using the Gaussian kernel function according to the best Gaussian kernel width and the distance between nodes.

[0007] Optionally, the method of obtaining the best Gaussian kernel width corresponding to the node by comparing the local density of the node with the overall distribution density comprises: comparing the local density of each node with the overall distribution density, if the local density of the current node is smaller, calculating the corresponding best Gaussian kernel width by multiplying the ratio of the overall distribution density and the local density of the current node with the initial Gaussian kernel width of the node; if the local density of the current node is larger, calculating the corresponding best Gaussian kernel width by multiplying the inverse of the above ratio with the initial Gaussian kernel width of the node; and the initial Gaussian kernel width of the node is the average distance between the node and its neighbor nodes.

[0008] Optionally, the local density of the node is the inverse of the average distance between the node and its neighbor nodes.

[0009] Optionally, the average distance between the node and its neighbor nodes is calculated using a Euclidean distance measurement method, and the Euclidean distance measurement method comprises: for each node, selecting K neighbors to connect, calculating the average distance between the node and the neighbor nodes; K is a hyperparameter.

[0010] Optionally, the calculation of the node similarity weight using the Gaussian kernel function comprises: mapping the distance between the node and its neighbor nodes into a weight value between 0 and 1 using the Gaussian kernel function for the K neighbors of each fault feature initial graph node, and the closer the distance, the higher the weight.

[0011] Optionally, according to the fault feature graph, the label of the unlabeled sample in the historical sample set is updated using a label propagation algorithm, and an updated historical sample set is obtained, and the label propagation algorithm is based on the node similarity weight, the node similarity weight is calculated by the Gaussian kernel function, and the Gaussian kernel width is dynamically adjusted according to the comparison result of the local density and the overall distribution density of the node.

[0012] Optionally, according to the fault feature graph, the label of the unlabeled sample in the historical sample set is updated using a label propagation algorithm, and an updated historical sample set is obtained, and the label propagation algorithm is based on the node similarity weight, the node similarity weight is calculated by the Gaussian kernel function, and the Gaussian kernel width is dynamically adjusted according to the comparison result of the local density and the overall distribution density of the node.

[0013] Optionally, the training of the weighted LSTM model according to the updated historical sample set to obtain the fault diagnosis model comprises: converting the labels of all samples in the updated historical sample set into an encoded format to obtain a model sample set, proportionally stratified sampling the model sample set, training the weighted LSTM model, and obtaining the fault diagnosis model.

[0014] In a second aspect, the application provides an online warning system for the health state of a heat storage tank based on an LSTM, comprising: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned online warning method for the health state of the heat storage tank based on the LSTM is realized.

[0015] The application has the following technical effects:

[0016] The application realizes effective expansion of fault samples by designing a label propagation algorithm with adaptive kernel width; and realizes accurate and real-time online early warning of the health state of the heat accumulator by combining the capturing ability of the weighted LSTM model for long-term time sequence correlation characteristics. An adaptive Gaussian kernel width adjustment strategy based on local density is adopted to dynamically optimize the edge weight calculation, so that the fault feature map is more consistent with the data distribution, the overfitting or underfitting problem caused by the fixed kernel width is solved, and the accuracy of the edge weight is improved; a label propagation algorithm with adaptive kernel width is used to give reliable labels to a large number of unlabeled samples with a small number of known labeled samples, expand the training data, and improve the generalization ability of the model; the adaptive kernel width strategy adapts to the sample distribution difference, improves the label propagation accuracy, and reduces the false alarm and missed alarm rates by combining the time sequence learning ability of the weighted LSTM model, thereby realizing online early warning of the health state of the heat accumulator. BRIEF DESCRIPTION OF DRAWINGS

[0017] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings in which several embodiments of the application are shown by way of illustration. In the drawings identical or corresponding elements are provided with the same reference numerals and descriptions are provided only for the first time. In the drawings:

[0018] Figure 1 is a flow chart of the online early warning method for the health state of the heat accumulator based on the LSTM provided by the embodiments of the application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only a part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the application.

[0020] The embodiments of the application disclose an online early warning method for the health state of a heat accumulator based on an LSTM, referring to Figure 1 , comprising:

[0021] S1: Collecting multi-dimensional time sequence data of the operation of the heat accumulator and performing preprocessing to obtain a historical sample set containing known labeled samples and unlabeled samples.

[0022] Specifically, the embodiments preset a sampling frequency (such as 10 seconds / time) and a sampling window (such as 5 minutes / time, containing 30 time points), and a plurality of temperature sensors arranged vertically in layers in the heat accumulator respectively collect temperature data at different heights (denoted as , , is the total number of temperature sensors), and a pressure sensor installed at the top gas space of the tank collects the pressure in the tank (denoted as ), the liquid level meter installed in the tank body collects the water level data in the tank - liquid level (denoted as ), each window corresponds to a continuous time sequence of multi-dimensional parameters, and the data at a single time can be represented as .

[0023] Specifically, data preprocessing and feature extraction are performed for each window, one window corresponds to one sample, the standardized time sequence suitable for LSTM (Long Short-Term Memory, Long Short-Term Memory Network) input is obtained, the historical sample set is obtained, and the health status label of the known label sample in the historical sample set is obtained based on the historical label information , wherein the sample with completed label annotation is a known label sample, and the unannotated sample is a no-label sample.

[0024] Specifically, taking one window as an example, the following steps are performed: (1) data preprocessing, using "time series interpolation method" (filling with the mean value of adjacent time points) to process the missing values; (2) feature extraction, performing standardization: Min-Max standardization is performed on the data of parameters with explicit physical range; Z-score standardization is used for parameter data without fixed range but with stable distribution; (3) sample feature vector construction, obtaining the data at each time point in a window, and obtaining the feature matrix corresponding to the window by summarizing, the matrix dimension is , each row corresponds to the data at a time point, and each column represents a dimension parameter. The matrix is a single input sample of the weighted LSTM model; (4) sample label annotation: obtain the known label samples in the historical samples by combining the historical operation records, artificial inspection results and other annotation information of the heat storage tank, and define the label as , indicating normal state, , indicating different fault types, , the total number of fault categories), and the label value corresponds to the actual health status of the heat storage tank at the last time point of the time sequence window.

[0025] S2: Construct a fault feature initial graph according to the feature vectors of each sample in the historical sample set, establish a dynamic association rule of local density and Gaussian kernel width of nodes, dynamically update the similarity weight of nodes, and obtain a fault feature graph; The nodes in the fault feature graph represent samples, and the edges represent the neighbor relationship between samples.

[0026] Specifically, based on the feature vector of the node, a fault feature initial graph is constructed by a sparse strategy (KNN neighbor method), and the Euclidean distance measurement method is selected to calculate the distance between nodes. For the dimension feature matrix of the node (each row corresponds to a time point, and each column corresponds to a parameter dimension), the Euclidean distance between nodes is calculated row by row according to time ​Total distance between feature matrixes is obtained by reaccumulation and averaging ; Selecting hyperparameters , find 8 nearest neighbor nodes for each node , each node is connected to only nearest neighbors to reduce the amount of calculation, and a graph structure for label propagation is constructed:

[0027] ; wherein, represents the total distance from the th initial graph node of the fault feature to the th initial graph node of the fault feature; represents the Euclidean distance at the th time from the th initial graph node of the fault feature to the th initial graph node of the fault feature, .

[0028] Specifically, the corresponding node similarity weight is calculated using a Gaussian kernel function, and the distance between the K nearest neighbors of each fault feature graph node and the fault feature graph node is mapped to a weight value between 0 and 1 using the Gaussian kernel function, and the closer the distance, the higher the weight.

[0029] In this embodiment, after the KNN neighbor method is used to construct the initial fault feature graph, the distance between nodes is converted into node similarity weight using a Gaussian kernel function: for the K nearest neighbors of each initial fault feature graph node, the distance between the center node and the center node is mapped to a weight value between 0 and 1 using the Gaussian kernel function, and the closer the distance, the higher the weight, thereby providing reasonable propagation strength for label propagation, and high-weight edges will allow "correct information" (such as similar labels) to be efficiently propagated, and low-weight edges will suppress the interference of "incorrect information" (such as dissimilar labels). If node is one of the neighbors of node , the node similarity weight is calculated using a Gaussian kernel function, otherwise the node similarity weight is set to 0. The expression for calculating the node similarity weight using the Gaussian kernel function is as follows:

[0030] ; wherein, represents the total distance from the th initial graph node of the fault feature to the th initial graph node of the fault feature; represents the Gaussian kernel width of the th initial graph node of the fault feature, which controls the speed of weight decay with distance; represents the th initial graph node of the fault feature and the The node similarity weights of the initial graph nodes for each fault feature; Indicates to Perform negative correlation normalization.

[0031] Through the above steps, a dynamic association rule between the local density of nodes and the width of the Gaussian kernel is established to dynamically update the node similarity weights, determine the distance between each fault feature initial graph node and its nearest neighbor nodes and its corresponding node similarity weights, and finally form a fault feature graph.

[0032] The dynamic association rule is as follows: calculate the local density of the node; take the average local density of all nodes as the overall distribution density; obtain the optimal Gaussian kernel width corresponding to the node by comparing the local density and the overall distribution density; and calculate the node similarity weight using the Gaussian kernel function based on the optimal Gaussian kernel width and the distance between nodes.

[0033] This application also provides a method for dynamically updating node similarity weight calculation. Specifically, it designs an adaptive kernel width strategy based on data distribution to find the optimal kernel width. This is used to apply the Gaussian kernel function to calculate node similarity weights, so as to dynamically update the node similarity weights. Local density of nodes is then established. With Gaussian kernel width The dynamic association rule is as follows: the local density of nodes reflects the degree of clustering of nodes in the fault feature map. The kernel width is adjusted based on the local density of nodes. The higher the local density of nodes, the larger the kernel width, which strengthens the connection between similar samples; the lower the local density of nodes, the smaller the kernel width, which reduces the interference of erroneous information.

[0034] The steps for dynamically updating node similarity weights based on dynamic association rules are as follows: Calculate the average distance between the current node and its nearest neighbors, and calculate the local density corresponding to the current node. Calculate the average local density of all nodes, defining it as the overall distribution density. Compare the local density value of the current node with the overall distribution density to obtain the optimal Gaussian kernel width for that node. Calculate the node similarity weight based on the optimal Gaussian kernel width and the distance.

[0035] Specifically, for each fault feature in the initial graph, the average distance between its K nearest neighbors is calculated. In this embodiment, K is preset to 8, and local density is defined. To achieve higher density with smaller distances, the average local density of all nodes in the initial fault feature graph is calculated and defined as the overall distribution density. .

[0036] ;in, It refers to the first From the initial graph node of the fault feature to the first the total distance of the K nearest neighbors of the initial graph node of the fault feature; is the Kth nearest neighbor of the initial graph node of the fault feature; is the Kth nearest neighbor of the initial graph node of the fault feature;

[0037] ; wherein, is the Kth nearest neighbor of the initial graph node of the fault feature; is the Kth nearest neighbor of the initial graph node of the fault feature; is the local density of the initial graph node of the fault feature; is the local density of the initial graph node of the fault feature;

[0038] ; wherein, is the local density of the initial graph node of the fault feature; is the local density of the initial graph node of the fault feature; is the overall distribution density of all initial graph nodes of the fault feature; is the total number of initial graph nodes of the fault feature.

[0039] Specifically, the node initial Gaussian kernel width of each initial graph node of the fault feature is set as The local density value of each initial graph node of the fault feature is compared with the overall distribution density: if the local density of the current node is smaller, the corresponding optimal Gaussian kernel width is calculated by the product of the ratio of the overall distribution density and the local density of the current node and the node initial Gaussian kernel width; if the local density of the current node is larger, the corresponding optimal Gaussian kernel width is calculated by the product of the inverse of the ratio of the overall distribution density and the local density of the current node and the node initial Gaussian kernel width. If the local density value is high, the kernel width is enlarged, so that the higher the density, the larger the kernel width, the greater the node similarity weight, and the stronger the contact of the same kind of samples; if the local density value is low, the kernel width is reduced, so that the lower the density, the smaller the kernel width, the smaller the node similarity weight, and the weaker the interference of the error information.

[0040] ; wherein, is the local density of the initial graph node of the fault feature; is the local density of the initial graph node of the fault feature; is the Gaussian kernel width of the initial graph node of the fault feature; is the Gaussian kernel width of the initial graph node of the fault feature; ​​​​​​​​​​​​​​It represents the overall distribution density of all nodes.

[0041] Specifically, calculate the local density of the current node. Overall distribution density of nodes in the initial fault feature map Compare the two sizes; if the local density of the current node is... A smaller kernel width indicates a smaller average distance between the current node and its nearest neighbors. This necessitates increasing the kernel width and strengthening the connections between similar samples, based on the overall distribution density. and the local density of the current node The ratio of the initial Gaussian kernel width of the node The optimal Gaussian kernel width is calculated for the product. If the local density of the current node A large value indicates that the average distance between the current node and its immediate neighbors is large, requiring a reduction in kernel width to decrease interference from erroneous information, which can be addressed by adjusting the overall distribution density. and the local density of the current node The negative of the ratio and the initial Gaussian kernel width of the node The optimal Gaussian kernel width is calculated for the product. .

[0042] Specifically, the fault feature map can be represented as G=(V, E, W), where V is the set of nodes; E is the set of edges (only the edges of the K nearest neighbors are retained); and W is the node similarity weight matrix composed of node similarity weights. diagonal elements The fault feature map can reflect the feature correlation of the health status samples (normal / faulty) of the heat storage tank in real time.

[0043] S3: Based on the fault feature map, the label propagation algorithm is used to update the labels of unlabeled samples in the historical sample set. Based on the updated historical sample set, a sample type label and propagation probability are introduced to distinguish whether a sample is a real label sample or a propagation label sample. A weighted cross-entropy loss function is constructed, a weighted LSTM model is trained, and a fault diagnosis model is obtained.

[0044] The loss function assigns a weight to each sample in the historical sample set. The weight is determined based on the sample's type label and propagation probability: the weight of a sample with a true label is 1; the weight of a sample with a propagation label is its corresponding propagation probability. The propagation probability is the probability that a sample with a propagation label will obtain its final label during the label propagation process.

[0045] In one embodiment, the method for updating the historical sample set is: taking the fault label value of the known fault category node as a seed to give the initial fault label value to the node of the fault feature graph, explicitly propagating the starting point, and propagating the fault label value based on the node similarity weight: the fault label value of the unknown fault category node is updated to the weighted sum of the fault label values of its adjacent known fault category nodes, and the iteration is continuously iterated until the fault label value of the unknown fault category node no longer changes. The label value of the converged sample node is rounded to obtain the final label of the node, that is, the fault category value; and the propagation probability of the sample node belonging to the final label is obtained.

[0046] In one embodiment, the fault label value iterative calculation expression can be: ; wherein, is the fault label value of the i th fault feature graph node in the j th iteration process; is the fault label value of the i th fault feature graph node in the j th iteration process; is the fault label value of the i th fault feature graph node in the j th iteration process; is the fault label value of the i th fault feature graph node in the j th iteration process; is the fault label value of the i th fault feature graph node in the j th iteration process; is the fault label value of the i th fault feature graph node in the j th iteration process; is the fault label value of the i th fault feature graph node in the j th iteration process; is the fault label value of the i th fault feature graph node in the j th iteration process; is the fault label value of the i th fault feature graph node in the j th iteration process; is the fault label value of the i th fault feature graph node in the j th iteration process; is the fault label value of the i th fault feature graph node in the j th iteration process; is the fault label value of the i th fault feature graph node in the j th iteration process; is the fault label value of the i th fault feature graph node in the j th iteration process; is the fault label value of the i th fault feature graph node in the j th iteration process; is the fault label value of the i th fault feature graph node in the j th iteration process; is the fault label value of the i th fault feature graph node in the j th iteration process; is the fault label value of the i th fault feature graph node in the j th iteration process;

[0047] The weighted LSTM model is trained based on the updated historical sample set to obtain a fault diagnosis model capable of mapping the "time sequence feature-health state" relationship.

[0048] The model sample set is determined based on the updated historical sample set. Through two steps of sample type labeling and label encoding conversion, a model sample set adapted for LSTM model training is constructed, as follows: Define sample type labels. In order to distinguish the source attributes of sample labels, define sample type labels G, where G=0 indicates that the sample is an original labeled real label sample (the label comes from real labeling information such as historical operation records of thermal storage tanks and manual inspection results); G=1 indicates that the sample is a mapped label sample obtained by the fault feature map label propagation algorithm in step S2 (the label is updated through neighbor node label propagation), and the propagation probability P of the sample belonging to this label is retrieved simultaneously.

[0049] The label encoding needs to be converted. Since the output of the LSTM model is a multi-class probability distribution, the labels of all samples in the updated historical sample set need to be converted into an encoding format (e.g., label Y=0 corresponds to [1,0,0,...,0], Y=1 corresponds to [0,1,0,...,0], the encoding dimension is C+1, which matches the total number of fault categories and normal state).

[0050] Through the above sample type labeling and encoding conversion, a model sample set for training the LSTM model is obtained.

[0051] To balance the training weights of real-label samples and mapped-label samples and improve the model's diagnostic accuracy, a weighted cross-entropy loss function is designed: Based on the sample type label G and the propagation probability P of the mapped-label samples, the weight R of the m-th sample in the model sample set is determined. m ; with the weight R of the m-th sample m We construct a weighted cross-entropy loss function using the weighting coefficients to calculate the error between the model's predicted probability distribution and the sample's true encoded label.

[0052] ;in, This refers to the weight of the m-th sample; This refers to the type label of the m-th sample; This refers to the propagation probability of the label of the m-th sample.

[0053] In this embodiment, according to The proportion is determined by stratified sampling to divide the model sample set, and the feature matrix of the input layer of the weighted LSTM model is calculated based on the received samples. The system uses a dimensional matrix (of dimensionality) to learn the mapping relationship between temporal features and labels of samples through hidden layers, Dropout layers, and fully connected layers. It then calculates the predicted probability distribution through forward propagation, calculates the prediction error using the cross-entropy loss function, and updates the weight parameters of each layer through backpropagation. The final output layer has a dimension of . The probability distribution is used to achieve the mapping learning of "temporal features - health status labels".

[0054] S4: Collecting real-time operation data of the heat storage tank, inputting the fault diagnosis model after the same pre-processing as the historical sample, and outputting the diagnosis result of the health status of the heat storage tank.

[0055] Specifically, in the embodiment, the continuous time series data of the multi-dimensional parameters of the heat storage tank is collected in real time according to the sampling frequency and the sampling window of step S1, the pre-processing operation consistent with the historical sample is performed, the feature matrix corresponding to the real-time window is obtained, the trained LSTM fault diagnosis model is input, and the probability distribution with a dimension of corresponding to the "normal state and various fault states" is output, the category corresponding to the maximum value in the probability distribution is extracted as the health status diagnosis result of the current heat storage tank, and if it is not the normal state, the corresponding early warning information (specifically, the type of fault) is output.

[0056] In the specific application of the embodiment of the application, the fault label value of a small amount of known fault category nodes can be used to realize the diagnosis of real-time data, the dependence on a large amount of labeled data is reduced, and the data acquisition cost is reduced.

[0057] The embodiment of the application also discloses an LSTM-based online early warning system for the health status of a heat storage tank, comprising a processor and a memory, and the memory stores computer program instructions, which realize the LSTM-based online early warning method for the health status of a heat storage tank according to the application when executed by the processor.

[0058] The above system also comprises other components such as a communication bus and a communication interface which are well known to those skilled in the art, and the settings and functions thereof are known in the art, so they will not be described here.

[0059] In the application, the aforementioned memory can be any tangible medium containing or storing a program, which can be used or combined with an instruction execution system, device or instrument. For example, the computer readable storage medium can be any appropriate magnetic storage medium or magneto-optical storage medium, such as resistive memory, dynamic random access memory, static random access memory, etc., or any other medium that can be used to store the required information and can be accessed by an application program, module or both. Any such computer storage medium can be part of the device or accessible or connectable to the device.

[0060] The above are the preferred embodiments of the application, which do not limit the protection scope of the application, therefore: any equivalent changes made on the basis of the structure, shape, principle of the application should be covered within the protection scope of the application.

Claims

1. An online health state early warning method for a thermal storage tank based on LSTM, characterized in that, The method comprises the steps of: collecting multi-dimensional time sequence data of the operation of the heat storage tank and pre-processing the data to obtain a historical sample set containing known label samples and unlabelled samples; constructing a fault feature initial graph according to feature vectors of the samples in the historical sample set, establishing a dynamic association rule of local density of nodes and Gaussian kernel width, dynamically updating the similarity weight of nodes, and obtaining a fault feature graph; in the fault feature graph, the nodes represent the samples, and the edges represent the neighbor relationship between the samples; According to the fault feature graph, the label propagation algorithm is used to update the labels of the unlabelled samples in the historical sample set, and an updated historical sample set is obtained. Based on the updated historical sample set, a sample type label and a propagation probability for distinguishing whether a sample is a real label sample or a propagated label sample are introduced, a weighted cross-entropy loss function is constructed, a weighted LSTM model is trained, and a fault diagnosis model is obtained; wherein the propagation probability is a probability value of a propagated label sample obtaining its final label in the label propagation process; the loss function gives each sample in the historical sample set a weight, and the weight is determined according to the type label and the propagation probability of the sample: the weight of the sample as a real label sample is 1; the weight of the sample as a propagated label sample is the corresponding propagation probability; Collecting real-time operation data of the heat storage tank, pre-processing the data in the same way as the historical samples, and inputting the data into the fault diagnosis model to output a diagnosis result of the health status of the heat storage tank. The dynamic association rule is: calculating the local density of the node; taking the average of the local densities of all nodes as the overall distribution density; obtaining the best Gaussian kernel width corresponding to the node by comparing the local density of the node with the overall distribution density; and calculating the similarity weight of the node using the Gaussian kernel function according to the best Gaussian kernel width and the distance between the nodes. Setting the node initial Gaussian kernel width of each fault feature initial graph node, comparing the local density value of each fault feature initial graph node with the overall distribution density: if the local density of the current node is small, the best Gaussian kernel width corresponding to the current node is calculated by multiplying the ratio of the overall distribution density to the local density of the current node with the node initial Gaussian kernel width; if the local density of the current node is large, the best Gaussian kernel width corresponding to the current node is calculated by multiplying the inverse of the ratio of the overall distribution density to the local density of the current node with the node initial Gaussian kernel width. According to the fault feature graph, the label propagation algorithm is used to update the labels of the unlabelled samples in the historical sample set, and an updated historical sample set is obtained. 2.The LSTM-based online health state early warning method for a thermal storage tank according to claim 1, characterized in that, The local density of the node is the inverse of the average distance between the node and its neighbor nodes. The local density of the node is the inverse of the average distance between the node and its neighbor nodes. 3.The LSTM-based online health state early warning method for a thermal storage tank according to claim 2, characterized in that, The average distance between the node and its neighbor nodes is calculated using a Euclidean distance measurement method; the Euclidean distance measurement method comprises: for each node, selecting K neighbors to connect, and calculating the average distance between the node and the neighbor nodes; K is a hyperparameter. 4.The LSTM-based online health state early warning method for a thermal storage tank according to claim 3, characterized in that, The calculation of the node similarity weight using the Gaussian kernel function comprises: mapping the distance between the node and its neighbor nodes into a weight value between 0 and 1 by using the Gaussian kernel function on the K neighbors of each fault feature initial graph node, and the closer the distance, the higher the weight. 5.The LSTM-based online health state early warning method for a thermal storage tank according to claim 1, characterized in that, The training of the weighted LSTM model according to the updated historical sample set to obtain the fault diagnosis model comprises: converting the labels of all samples in the updated historical sample set into an encoded format to obtain a model sample set, proportionally stratified sampling the model sample set, training the weighted LSTM model, and obtaining the fault diagnosis model.

6. The online health state early warning system for the thermal storage tank based on LSTM, characterized in that, Comprise: A processor and a memory, the memory stores computer program instructions, when the computer program instructions are executed by the processor, the method for online early warning of the health state of the heat accumulation tank based on the LSTM according to any one of claims 1-5 is realized.

Images