Method and device for detecting static pressure abnormity of blast furnace shaft

Abstract

The invention provides a blast furnace body static pressure anomaly detection method and device, and the method comprises the steps: obtaining the time sequence data of the static pressure of a blast furnace body, decomposing the time sequence data, obtaining a plurality of time sequence components, calculating the sample entropy of each time sequence component, carrying out the clustering of the sample entropy of each time sequence component, and carrying out the detection of the static pressure anomaly of the blast furnace body according to the clustering result. Superposing the time sequence components belonging to the same type to obtain a superposed component, inputting the superposed component into an adversarial self-encoding network model to obtain an estimated component of the superposed component, calculating a score of the estimated component of the superposed component, and calculating the score of the estimated component of the superposed component according to a comparison result of the score of the estimated component of the superposed component and a preset score threshold value. Determining abnormal fragments of the time series data, and combining the abnormal fragments of the time series data to obtain an anomaly detection result; the technical problems that monitoring of the static pressure abnormal state of the blast furnace body does not have a unified standard and cannot be quantified, and the working intensity is high due to a manual monitoring mode are solved, and the production and control stability of the blast furnace can be ensured.

Description

Technical Field

[0001] This application relates to the field of industrial time-series signal analysis technology, and in particular to a method and apparatus for detecting abnormal static pressure in blast furnace body. Background Art

[0002] The blast furnace is a key piece of equipment in the metallurgical industry used for ironmaking, smelting iron ore (such as hematite and magnetite) into liquid pig iron through a reduction reaction. The gas flow distribution within the blast furnace is a crucial factor affecting its stable operation, and it is characterized by the static pressure measurement signal. The static pressure is measured by pressure sensors evenly installed around the cooling walls of certain layers of the blast furnace. Abnormalities in the static pressure signal often indicate abnormal gas flow distribution within the blast furnace.

[0003] In related technologies, the main method for monitoring abnormal static pressure in the furnace body is for blast furnace operators to observe the static pressure signal curve in real time and make judgments on whether the static pressure is abnormal or normal based on their own experience. This requires a lot of effort, and because different operators have different analytical logic or experience, it is difficult to achieve standardized and quantifiable monitoring of the static pressure in the furnace body.

[0004] Therefore, it is necessary to improve the methods for detecting abnormal static pressure in the blast furnace body in related technologies. Summary of the Invention

[0005] In view of the shortcomings of the prior art described above, this application provides a method and device for detecting abnormal static pressure in blast furnace body, so as to solve the technical problems of lack of unified standards, inability to quantify, and time-consuming and labor-intensive monitoring of abnormal static pressure in blast furnace body.

[0006] According to one aspect of the embodiments of this application, a method for detecting abnormal static pressure in a blast furnace body is provided. The method includes: acquiring time-series data of static pressure in the blast furnace body; the static pressure in the blast furnace body is used to characterize the pressure of gas in the blast furnace body under static conditions; decomposing the time-series data to obtain multiple time-series components; calculating the sample entropy of each time-series component, clustering the sample entropy of each time-series component, and superimposing time-series components belonging to the same type according to the clustering results to obtain superimposed components, the superimposed components including: trend term time-series components, period term time-series components, and residual term time-series components; inputting the superimposed components into an adversarial autoencoder network model to obtain estimated components of the superimposed components; the adversarial autoencoder network model is trained based on the sample components; calculating the score of the estimated components of the superimposed components; determining abnormal segments of the time-series data based on the comparison result of the score of the estimated components of the superimposed components with a preset score threshold; merging the abnormal segments of the time-series data to obtain the abnormal detection result of the static pressure in the blast furnace body.

[0007] In one embodiment of this application, the process of calculating the sample entropy of each time-series component includes: segmenting each time-series component according to a first preset length to obtain a first segmentation vector sequence set for each time-series component; calculating the distance value between different first segmentation vector sequences in each first segmentation vector sequence set, denoted as a first distance value; filtering the first distance value according to the dimension of each first segmentation vector sequence to obtain the distance value associated with each first segmentation vector sequence; determining the distance quantity ratio of each first segmentation vector sequence set based on the distance value associated with each first segmentation vector sequence and the number of first segmentation vector sequences in each first segmentation vector sequence set; and segmenting each time-series component according to a second preset length. Obtain the second segmentation vector sequence set for each time series component; calculate the distance value between different second segmentation vector sequences in each second segmentation vector sequence set, denoted as the second distance value; and filter the second distance value according to the dimension of each second segmentation vector sequence to obtain the distance value associated with each second segmentation vector sequence; determine the distance quantity ratio of each second segmentation vector sequence set based on the distance value associated with each second segmentation vector sequence and the number of second segmentation vector sequences in each second segmentation vector sequence set; the second preset length is greater than the first preset length; determine the sample entropy of each time series component based on the distance quantity ratio of each first segmentation vector sequence set and the distance quantity ratio of each second segmentation vector sequence set.

[0008] In one embodiment of this application, the process of determining the distance quantity ratio of each first segmentation vector sequence set based on the distance value associated with each first segmentation vector sequence and the number of first segmentation vector sequences in each first segmentation vector sequence set includes: selecting one first segmentation vector sequence set as a first target segmentation vector sequence set; determining the distance quantity ratio of the first target segmentation vector sequence set based on the distance value associated with each first segmentation vector sequence in the first target segmentation vector sequence set and the number of first segmentation vector sequences in the first target segmentation vector sequence set; selecting another first segmentation vector sequence set as a second target segmentation vector sequence set; determining the distance quantity ratio of the second target segmentation vector sequence set based on the distance value associated with each first segmentation vector sequence in the second target segmentation vector sequence set and the number of first segmentation vector sequences in the second target segmentation vector sequence set; and so on, until all first segmentation vector sequence sets have been selected, thereby obtaining the distance quantity ratio of each first segmentation vector sequence set.

[0009] In one embodiment of this application, the process of determining the distance quantity ratio of the first target segmentation vector sequence set based on the distance value associated with each first segmentation vector sequence in the first target segmentation vector sequence set and the number of first segmentation vector sequences in the first target segmentation vector sequence set includes: selecting one first segmentation vector sequence in the first target segmentation vector sequence set as the first target segmentation vector sequence; counting the number of distance values ​​associated with the first target segmentation vector sequence that are greater than a first filtering threshold, and recording it as the first filtering distance quantity; deleting the first target segmentation vector sequence from the first target segmentation vector sequence set to obtain the first remaining segmentation vector sequence quantity in the first target segmentation vector sequence set, and multiplying the first remaining segmentation vector sequence quantity by the first preset length as the first remaining data quantity in the first target segmentation vector sequence set; the first filtering threshold is determined by the standard deviation of the first target segmentation vector sequence and the preset threshold coefficient; The ratio of the first filtered distance count to the first remaining data volume is taken as the distance count percentage of the first target segmentation vector sequence; another first segmentation vector sequence in the first target segmentation vector sequence set is selected as the second target segmentation vector sequence; the distance count percentage of the second target segmentation vector sequence is determined based on the distance value associated with the second target segmentation vector sequence, the second filtering threshold, the number of first segmentation vector sequences in the first target segmentation vector sequence set, and the first preset length, until all first segmentation vector sequences in the first target segmentation vector sequence set are selected, thus obtaining the distance count percentage of all first segmentation vector sequences in the first target segmentation vector sequence set; the second filtering threshold is determined by the standard deviation of the second target segmentation vector sequence and the preset threshold coefficient; the average of the distance count percentages of all first segmentation vector sequences in the first target segmentation vector sequence set is taken as the distance count percentage of the first target segmentation vector sequence set.

[0010] In one embodiment of this application, the process of determining the sample entropy of each temporal component based on the distance quantity ratio of each first segmentation vector sequence set and the distance quantity ratio of each second segmentation vector sequence set includes: selecting one temporal component as a first target temporal component, calculating the sample entropy of the first target temporal component based on the distance quantity ratio of the first segmentation vector sequence set corresponding to the first target temporal component and the distance quantity ratio of the second segmentation vector sequence set corresponding to the first target temporal component; selecting another temporal component as a second target temporal component, calculating the sample entropy of the second target temporal component based on the distance quantity ratio of the first segmentation vector sequence set corresponding to the second target temporal component and the distance quantity ratio of the second segmentation vector sequence set corresponding to the second target temporal component, until all temporal components have been selected, and obtaining the sample entropy of each temporal component.

[0011] In one embodiment of this application, if the sample components include: trend term time-series sample components, periodic term time-series sample components, and residual term time-series sample components, then the process of training a preset adversarial autoencoder network model based on the sample components to obtain the adversarial autoencoder network model includes: training the preset adversarial autoencoder network model using the trend term time-series sample components to obtain a trend adversarial autoencoder network model; the preset adversarial autoencoder network model includes: an encoder, a first decoder, and a second decoder; the trend adversarial autoencoder network model is used to output an estimated component of the trend term time-series sample components when the trend term time-series sample components are taken as input; training the preset adversarial autoencoder network model using the periodic term time-series sample components to obtain the trend adversarial autoencoder network model includes: training the preset adversarial autoencoder network model using the trend term time-series sample components; training the preset adversarial autoencoder network model using the periodic term time-series sample components to obtain the trend adversarial autoencoder network model. An adversarial autoencoder network model is trained to obtain a periodic adversarial autoencoder network model. This model outputs an estimated component of the periodic term's time-series sample component when the time-series sample component of the periodic term is taken as input. A preset adversarial autoencoder network model is trained using the time-series sample component of the residual term to obtain a residual adversarial autoencoder network model. This model outputs an estimated component of the residual term's time-series sample component when the time-series sample component of the residual term is taken as input. The trend adversarial autoencoder network model, the periodic adversarial autoencoder network model, and the residual adversarial autoencoder network model are combined to obtain the final adversarial autoencoder network model.

[0012] In one embodiment of this application, if the preset adversarial autoencoder network model includes: a first preset adversarial autoencoder network model and a second preset adversarial autoencoder network model, then the process of training the preset adversarial autoencoder network model using the trend term time-series sample components to obtain a trend adversarial autoencoder network model includes: inputting the trend term time-series sample components into the first preset adversarial autoencoder network model to obtain a first-order estimated component of the trend term time-series sample components; and inputting the first-order estimated component of the trend term time-series sample components into the second preset adversarial autoencoder network model to obtain a second-order estimated component of the trend term time-series sample components; the parameter values ​​in the first preset adversarial autoencoder network model are different from the parameter values ​​in the second preset adversarial autoencoder network model; based on the trend term time-series sample components and the trend term time-series sample components... The difference between the second-order estimated components of the trend term is used to determine the loss function; a first preset hyperparameter is used as a hyperparameter in the loss function to obtain a first preset loss function, and a second preset hyperparameter is used as a hyperparameter in the loss function to obtain a second preset loss function; the trend term time series sample component and its second-order estimated component are used as inputs to the first preset loss function, and the first preset adversarial autoencoder network model is trained with the goal of maximizing the output of the first preset loss function; the trend term time series sample component and its second-order estimated component are used as inputs to the second preset loss function, and the second preset adversarial autoencoder network model is trained with the goal of minimizing the output of the second preset loss function, thus obtaining the trend adversarial autoencoder network model.

[0013] In one embodiment of this application, the process of calculating the score of the estimated component of the superimposed component includes: inputting the estimated component of the superimposed component into a first preset adversarial autoencoder network model to obtain a first-order estimated component of the superimposed component; inputting the first-order estimated component of the superimposed component into a second preset adversarial autoencoder network model to obtain a second-order estimated component of the superimposed component; using a third preset hyperparameter as a hyperparameter in the loss function to obtain a third preset loss function, and using a fourth preset hyperparameter as a hyperparameter in the loss function to obtain a fourth preset loss function; constructing a scoring function based on the third preset loss function and the fourth preset loss function; and inputting the estimated component of the superimposed component, the first-order estimated component of the superimposed component, and the second-order estimated component of the superimposed component into the scoring function to obtain a score of the estimated component of the superimposed component.

[0014] In one embodiment of this application, if the preset scoring threshold includes a trend item scoring threshold, a period item scoring threshold, and a residual item scoring threshold, then the process of determining the abnormal segment of the time series data based on the comparison result of the estimated component score of the superimposed component with the preset scoring threshold includes: if the estimated component score of the trend item time series component is greater than the trend item scoring threshold, then the time series segment corresponding to the trend item time series component is determined to be an abnormal segment of the time series data; if the estimated component score of the period item time series component is greater than the period item scoring threshold, then the time series segment corresponding to the period item time series component is determined to be an abnormal segment of the time series data; if the estimated component score of the residual item time series component is greater than the residual item scoring threshold, then the time series segment corresponding to the residual item time series component is determined to be an abnormal segment of the time series data.

[0015] According to one aspect of the embodiments of this application, a blast furnace body static pressure anomaly detection device is provided, comprising: a data acquisition module for acquiring time-series data of blast furnace body static pressure; the blast furnace body static pressure is used to characterize the pressure of gas in the blast furnace body under static conditions; a data decomposition module for decomposing the time-series data to obtain multiple time-series components; calculating the sample entropy of each time-series component, clustering the sample entropy of each time-series component, and superimposing time-series components belonging to the same type according to the clustering results to obtain a superimposed component, the superimposed component including: a trend term time-series component, a period... The system comprises: a time-series component of the superimposed component and a time-series component of the residual component; a component estimation module, used to input the superimposed component into an adversarial autoencoder network model to obtain an estimated component of the superimposed component; the adversarial autoencoder network model is trained on a preset adversarial autoencoder network model based on the sample component; an anomaly detection module, used to calculate the score of the estimated component of the superimposed component, and determine the abnormal segments of the time-series data based on the comparison result of the score of the estimated component of the superimposed component with a preset score threshold; and to merge the abnormal segments of the time-series data to obtain the anomaly detection result of the static pressure of the blast furnace body.

[0016] The beneficial effects of this application are as follows: This application acquires time-series data of blast furnace body static pressure, decomposes the time-series data to obtain multiple time-series components, calculates the sample entropy of each time-series component, clusters the sample entropy of each time-series component, and superimposes time-series components belonging to the same type according to the clustering results to obtain superimposed components. The superimposed components are input into an adversarial autoencoder network model to obtain estimated components of the superimposed components. The scores of the estimated components of the superimposed components are calculated. Based on the comparison results of the estimated scores of the superimposed components with a preset score threshold, abnormal segments of the time-series data are identified. These abnormal segments are then merged to obtain the anomaly detection results of blast furnace body static pressure. The process involves decomposing, clustering, and superimposing time-series data to obtain superimposed components. An adversarial autoencoder network model is then used to output estimated components of the superimposed components. After obtaining scores for these estimated components, the system compares these scores with preset scoring thresholds to identify abnormal segments in the time-series data. This results in the detection of abnormal blast furnace static pressure. The system achieves standardized and unified anomaly detection methods for blast furnace static pressure, solving the technical problems of lacking unified standards and quantifiability in monitoring abnormal blast furnace static pressure, as well as the high workload associated with manual monitoring. This approach helps ensure the stability of blast furnace production and control.

[0017] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings:

[0019] Figure 1 This is a schematic diagram illustrating an exemplary system architecture as shown in an exemplary embodiment of this application;

[0020] Figure 2 This is a flowchart illustrating an exemplary embodiment of the method for detecting abnormal static pressure in a blast furnace body, as shown in this application.

[0021] Figure 3 This is a flowchart illustrating a method for detecting abnormal static pressure in the blast furnace body, as shown in another exemplary embodiment of this application.

[0022] Figure 4 This is a block diagram illustrating an exemplary embodiment of the blast furnace body static pressure anomaly detection device.

[0023] Figure 5 This is a schematic diagram of the structure of a computer system for an electronic device, as illustrated in an exemplary embodiment of this application. DETAILED DESCRIPTION

[0024] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0025] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0026] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the present application. However, it will be apparent to those skilled in the art that embodiments of the present application may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the present application.

[0027] Figure 1 This is a schematic diagram illustrating an exemplary system architecture as shown in an exemplary embodiment of this application.

[0028] Reference Figure 1As shown, the system architecture may include a data acquisition device 101 and a computer device 102. The computer device 102 may be at least one of a desktop graphics processing unit (GPU) computer, a GPU computing cluster, or a neural network computer. Technical personnel can use this computer device 102 to acquire time-series data of blast furnace static pressure, decompose the time-series data to obtain multiple time-series components, calculate the sample entropy of each time-series component, cluster the sample entropy of each time-series component, and, according to the clustering results, superimpose time-series components belonging to the same type to obtain superimposed components. The superimposed components are then input into an adversarial autoencoder network model to obtain estimated components of the superimposed components. A score for the estimated components of the superimposed components is calculated. Based on the comparison between the score of the estimated components of the superimposed components and a preset score threshold, abnormal segments of the time-series data are identified. These abnormal segments are then merged to obtain the anomaly detection result of the blast furnace static pressure. The data acquisition device 101 is used to collect static pressure data of the blast furnace body. After arranging the static pressure data of the blast furnace body in chronological order, the time-series data of the static pressure of the blast furnace body is obtained and provided to the computer device 102 for processing.

[0029] In a schematic manner, computer device 102 acquires time-series data from acquisition device 101, decomposes the time-series data to obtain multiple time-series components, calculates the sample entropy of each time-series component, clusters the sample entropy of each time-series component, and superimposes time-series components belonging to the same type according to the clustering results to obtain superimposed components. The superimposed components are input into an adversarial autoencoder network model to obtain estimated components of the superimposed components. The score of the estimated components of the superimposed components is calculated. Based on the comparison result of the estimated component scores of the superimposed components with a preset score threshold, abnormal segments of the time-series data are identified. These abnormal segments of the time-series data are merged to obtain the abnormal detection result of the blast furnace body static pressure. The above process involves decomposing, clustering, and superimposing time-series data to obtain superimposed components. An estimated component of the superimposed components is then output through an adversarial autoencoder network model. After obtaining the score of the estimated component of the superimposed components, the abnormal segments of the time-series data are identified based on the comparison between the score of the estimated component of the superimposed components and the preset score threshold. This yields the abnormal detection result of the blast furnace body static pressure. This method enables the detection of abnormal blast furnace body static pressure through a standardized and unified method, solving the technical problems of lacking unified standards and quantifiability in monitoring abnormal states of blast furnace body static pressure, as well as the high workload caused by manual monitoring. This is beneficial for ensuring the stability of blast furnace production and control.

[0030] It should be noted that the blast furnace body static pressure abnormality detection method provided in this application embodiment is generally executed by computer equipment 102, and correspondingly, the blast furnace body static pressure abnormality detection device is generally installed in computer equipment 102.

[0031] The implementation details of the technical solutions in the embodiments of this application are described in detail below:

[0032] Figure 2 This is a flowchart illustrating an exemplary embodiment of a blast furnace body static pressure anomaly detection method. This method can be executed by a computational processing device, which may be... Figure 2 The computer device 102 shown is illustrated. (Refer to...) Figure 2 As shown, the method for detecting abnormal static pressure in the blast furnace body includes at least steps S210 to S230, which are described in detail below:

[0033] In step S210, time-series data of blast furnace body static pressure is acquired. In one embodiment of this application, blast furnace body static pressure is used to characterize the pressure of gas in the blast furnace body under static conditions. The blast furnace body static pressure data is measured by pressure sensors uniformly installed around the cooling walls of the blast furnace layers. After obtaining the blast furnace body static pressure data, the data is sorted, deduplicated, and missing values ​​are filled according to time sequence to obtain the time-series data of blast furnace body static pressure.

[0034] In step S220, the time-series data is decomposed to obtain multiple time-series components; the sample entropy of each time-series component is calculated, and the sample entropy of each time-series component is clustered. According to the clustering results, time-series components belonging to the same type are superimposed to obtain superimposed components. In one embodiment of this application, the superimposed components include: trend term time-series components, period term time-series components, and residual term time-series components. The time-series data is decomposed using the VMD (Variational Mode Decomposition) method. The VMD method is a non-recursive signal processing method that decomposes time-series data into a series of intrinsic mode functions (IMFs) with finite bandwidth. Compared with the empirical mode decomposition (EMD) method, the VMD method can effectively avoid mode aliasing by iteratively optimizing the IMF components, making the decomposition results more accurate and stable. The objective function of VMD is shown below:

[0035]

[0036] in, This represents the partial derivative function, δ(·) represents the Dirac distribution function, and u k f(t) represents the k-th time component, K represents the number of time components, and f(t) represents the time data.

[0037] In another embodiment of this application, the method for clustering the sample entropy of each time series component is k-means clustering. The clustering results are used to characterize which time series components belong to the same type and which belong to different types. Time series components belonging to the same type are then superimposed to obtain superimposed components. The types of time series components include: trend terms, periodic terms, and residual terms.

[0038] In step S230, the superimposed components are input into the adversarial autoencoder network model to obtain the estimated components of the superimposed components. In one embodiment of this application, the adversarial autoencoder network model is trained on a preset adversarial autoencoder network model based on sample components. The adversarial autoencoder network model includes: a trend adversarial autoencoder network model, a periodic adversarial autoencoder network model, and a residual adversarial autoencoder network model. The process of inputting the superimposed components into the adversarial autoencoder network model to obtain the estimated components of the superimposed components includes: inputting the trend term time series component into the trend adversarial autoencoder network model to obtain the estimated components of the trend term time series component; inputting the periodic term time series component into the periodic adversarial autoencoder network model to obtain the estimated components of the periodic term time series component; and inputting the residual term time series component into the residual adversarial autoencoder network model to obtain the estimated components of the residual term time series component.

[0039] In step S240, the score of the estimated component of the superimposed component is calculated. Based on the comparison result of the score of the estimated component of the superimposed component with the preset score threshold, abnormal segments of the time series data are determined. The abnormal segments of the time series data are merged to obtain the abnormal detection result of the static pressure of the blast furnace body. In one embodiment of this application, the process of calculating the score of the estimated component of the superimposed component includes: inputting the estimated component of the superimposed component into a first preset adversarial autoencoder network model to obtain the first-order estimated component of the superimposed component; inputting the first-order estimated component of the superimposed component into a second preset adversarial autoencoder network model to obtain the second-order estimated component of the superimposed component; using a third preset hyperparameter as a hyperparameter in the loss function to obtain a third preset loss function, and using a fourth preset hyperparameter as a hyperparameter in the loss function to obtain a fourth preset loss function; constructing a scoring function based on the third preset loss function and the fourth preset loss function; and inputting the estimated component of the superimposed component, the first-order estimated component of the superimposed component, and the second-order estimated component of the superimposed component into the scoring function to obtain the score of the estimated component of the superimposed component.

[0040] In one embodiment of this application, time-series data is decomposed, clustered, and superimposed to obtain superimposed components. An estimated component of the superimposed components is output through an adversarial autoencoder network model. After obtaining the score of the estimated component of the superimposed components, the abnormal segments of the time-series data are determined based on the comparison result of the score of the estimated component of the superimposed components with the preset score threshold. Thus, the abnormal detection result of the blast furnace body static pressure is obtained. This realizes the abnormal detection of blast furnace body static pressure through a standardized and unified abnormal detection method, which solves the technical problems of no unified standard, non-quantitative nature of monitoring abnormal state of blast furnace body static pressure, and high workload caused by manual monitoring. This is conducive to ensuring the stability of blast furnace production and control.

[0041] In one embodiment of this application, the process of calculating the sample entropy of each time-series component includes:

[0042] According to a first preset length, each time component is segmented to obtain a first segmentation vector sequence set for each time component; the distance value between different first segmentation vector sequences in each first segmentation vector sequence set is calculated and recorded as the first distance value; and the first distance value is filtered according to the dimension of each first segmentation vector sequence to obtain the distance value associated with each first segmentation vector sequence; the proportion of distances in each first segmentation vector sequence set is determined based on the distance value associated with each first segmentation vector sequence and the number of first segmentation vector sequences in each first segmentation vector sequence set. In one embodiment of this application, the first preset length is determined according to the actual situation. The process of calculating the distance value between different first segmentation vector sequences in each first segmentation vector sequence set includes: selecting two different first segmentation vector sequences in the same first segmentation vector sequence set; calculating the absolute value of the difference between corresponding elements in the two different first segmentation vector sequences; taking the maximum value of the absolute value of the difference between all corresponding elements as the distance value between the two different first segmentation vector sequences; and calculating the distance value between all different first segmentation vector sequences in the same first segmentation vector sequence set according to the above method until the distance value between all different first segmentation vector sequences in all first segmentation vector sequence sets is calculated. The formula for calculating the distance between two different first segmentation vector sequences is as follows:

[0043] d ij =max{|x i,a (t)-x j,a (t)|} Equation (2)

[0044] Where, d ij Let x represent the distance between the i-th and j-th first segmentation vector sequences. i,a (t) represents the element at position a in the i-th first segmentation vector sequence, x j,a(t) represents the element at position a in the j-th first segmentation vector sequence, where a = {0, 1, ..., m-1} and m is the length of the first segmentation vector sequence, i.e., the first preset length.

[0045] In another embodiment of this application, the process of determining the distance quantity ratio of each first segmentation vector sequence set based on the distance value associated with each first segmentation vector sequence and the number of first segmentation vector sequences in each first segmentation vector sequence set includes: selecting one first segmentation vector sequence set as a first target segmentation vector sequence set; determining the distance quantity ratio of the first target segmentation vector sequence set based on the distance value associated with each first segmentation vector sequence in the first target segmentation vector sequence set and the number of first segmentation vector sequences in the first target segmentation vector sequence set; selecting another first segmentation vector sequence set as a second target segmentation vector sequence set; determining the distance quantity ratio of the second target segmentation vector sequence set based on the distance value associated with each first segmentation vector sequence in the second target segmentation vector sequence set and the number of first segmentation vector sequences in the second target segmentation vector sequence set; and so on, until all first segmentation vector sequence sets have been selected, thereby obtaining the distance quantity ratio of each first segmentation vector sequence set.

[0046] According to a second preset length, each time component is segmented to obtain a second segmentation vector sequence set for each time component. The distance value between different second segmentation vector sequences in each second segmentation vector sequence set is calculated and denoted as the second distance value. The second distance values ​​are then filtered according to the dimension of each second segmentation vector sequence to obtain the distance value associated with each second segmentation vector sequence. Based on the distance value associated with each second segmentation vector sequence and the number of second segmentation vector sequences in each second segmentation vector sequence set, the proportion of distances in each second segmentation vector sequence set is determined. In one embodiment of this application, the second preset length is greater than the first preset length; the process of calculating the distance value between different second segmentation vector sequences in each second segmentation vector sequence set is the same as the process of calculating the distance value between different first segmentation vector sequences in each first segmentation vector sequence set.

[0047] In another embodiment of this application, the process of determining the distance quantity ratio of each second segmentation vector sequence set based on the distance value associated with each second segmentation vector sequence and the number of second segmentation vector sequences in each second segmentation vector sequence set includes: selecting one second segmentation vector sequence set as a third target segmentation vector sequence set; determining the distance quantity ratio of the third target segmentation vector sequence set based on the distance value associated with each second segmentation vector sequence in the third target segmentation vector sequence set and the number of second segmentation vector sequences in the third target segmentation vector sequence set; selecting another second segmentation vector sequence set as a fourth target segmentation vector sequence set; determining the distance quantity ratio of the fourth target segmentation vector sequence set based on the distance value associated with each second segmentation vector sequence in the fourth target segmentation vector sequence set and the number of second segmentation vector sequences in the fourth target segmentation vector sequence set; until all second segmentation vector sequence sets have been selected, the distance quantity ratio of each second segmentation vector sequence set is obtained.

[0048] In one embodiment of this application, the process of determining the distance quantity ratio of the third target segmentation vector sequence set based on the distance value associated with each second segmentation vector sequence in the third target segmentation vector sequence set and the number of second segmentation vector sequences in the third target segmentation vector sequence set is the same as the process of determining the distance quantity ratio of the first target segmentation vector sequence set based on the distance value associated with each first segmentation vector sequence in the first target segmentation vector sequence set and the number of first segmentation vector sequences in the first target segmentation vector sequence set. Similarly, the process of determining the distance quantity ratio of the fourth target segmentation vector sequence set based on the distance value associated with each second segmentation vector sequence in the fourth target segmentation vector sequence set and the number of second segmentation vector sequences in the fourth target segmentation vector sequence set is the same as the process of determining the distance quantity ratio of the second target segmentation vector sequence set based on the distance value associated with each first segmentation vector sequence in the second target segmentation vector sequence set and the number of first segmentation vector sequences in the second target segmentation vector sequence set.

[0049] The sample entropy of each time-series component is determined based on the distance quantity ratio of each first segmentation vector sequence set and the distance quantity ratio of each second segmentation vector sequence set. In one embodiment of this application, the process of determining the sample entropy of each time-series component based on the distance quantity ratio of each first segmentation vector sequence set and the distance quantity ratio of each second segmentation vector sequence set includes: selecting one time-series component as a first target time-series component, calculating the sample entropy of the first target time-series component based on the distance quantity ratio of the first segmentation vector sequence set corresponding to the first target time-series component and the distance quantity ratio of the second segmentation vector sequence set corresponding to the first target time-series component; selecting another time-series component as a second target time-series component, calculating the sample entropy of the second target time-series component based on the distance quantity ratio of the first segmentation vector sequence set corresponding to the second target time-series component and the distance quantity ratio of the second segmentation vector sequence set corresponding to the second target time-series component, until all time-series components have been selected, and obtaining the sample entropy of each time-series component.

[0050] In one embodiment of this application, the process of determining the distance ratio of each first segmentation vector sequence set based on the distance value associated with each first segmentation vector sequence and the number of first segmentation vector sequences in each first segmentation vector sequence set includes:

[0051] A first segmentation vector sequence set is selected as the first target segmentation vector sequence set; the distance quantity ratio of the first target segmentation vector sequence set is determined based on the distance value associated with each first segmentation vector sequence in the first target segmentation vector sequence set and the number of first segmentation vector sequences in the first target segmentation vector sequence set. In one embodiment of this application, the process of determining the distance quantity ratio of the first target segmentation vector sequence set based on the distance value associated with each first segmentation vector sequence in the first target segmentation vector sequence set and the number of first segmentation vector sequences in the first target segmentation vector sequence set includes: selecting one first segmentation vector sequence in the first target segmentation vector sequence set as the first target segmentation vector sequence; counting the number of distance values ​​greater than a first screening threshold among the distance values ​​associated with the first target segmentation vector sequence, and recording it as the first screening distance quantity; deleting the first target segmentation vector sequence from the first target segmentation vector sequence set to obtain the first remaining segmentation vector sequence quantity in the first target segmentation vector sequence set, and multiplying the first remaining segmentation vector sequence quantity by a first preset length as the first remaining data quantity in the first target segmentation vector sequence set; the first screening threshold is determined by the standard deviation of the first target segmentation vector sequence and a preset threshold coefficient. The ratio of the first filtered distance quantity to the total amount of the first remaining data is taken as the distance quantity proportion of the first target segmentation vector sequence. Another first segmentation vector sequence in the first target segmentation vector sequence set is selected as the second target segmentation vector sequence. The distance quantity proportion of the second target segmentation vector sequence is determined based on the distance value associated with the second target segmentation vector sequence, the second filtering threshold, the number of first segmentation vector sequences in the first target segmentation vector sequence set, and the first preset length. This process continues until all first segmentation vector sequences in the first target segmentation vector sequence set have been selected, resulting in the distance quantity proportion of all first segmentation vector sequences in the first target segmentation vector sequence set. The second filtering threshold is determined by the standard deviation of the second target segmentation vector sequence and the preset threshold coefficient. The average of the distance quantity proportions of all first segmentation vector sequences in the first target segmentation vector sequence set is taken as the distance quantity proportion of the first target segmentation vector sequence set.

[0052] Select another first segmentation vector sequence set as the second target segmentation vector sequence set. Based on the distance value associated with each first segmentation vector sequence in the second target segmentation vector sequence set and the number of first segmentation vector sequences in the second target segmentation vector sequence set, determine the distance quantity ratio of the second target segmentation vector sequence set. Continue until all first segmentation vector sequence sets have been selected to obtain the distance quantity ratio of each first segmentation vector sequence set. In one embodiment of this application, the process of determining the proportion of distance values ​​in the second target segmentation vector sequence set based on the distance values ​​associated with each first segmentation vector sequence in the second target segmentation vector sequence set and the number of first segmentation vector sequences in the second target segmentation vector sequence set includes: selecting one first segmentation vector sequence in the second target segmentation vector sequence set as the third target segmentation vector sequence; counting the number of distance values ​​associated with the third target segmentation vector sequence that are greater than a third screening threshold, and recording this as the third screening distance quantity; deleting the third target segmentation vector sequence from the second target segmentation vector sequence set to obtain the number of first remaining segmentation vector sequences in the second target segmentation vector sequence set, and multiplying the number of first remaining segmentation vector sequences by a first preset length as the total amount of third remaining data in the second target segmentation vector sequence set; the third screening threshold is determined by the standard deviation of the third target segmentation vector sequence and a preset threshold coefficient. The ratio of the third filtered distance quantity to the total amount of the third remaining data is taken as the distance quantity proportion of the third target segmentation vector sequence. Another first segmentation vector sequence in the second target segmentation vector sequence set is selected as the fourth target segmentation vector sequence. The distance quantity proportion of the fourth target segmentation vector sequence is determined based on the distance value associated with the fourth target segmentation vector sequence, the fourth filtering threshold, the number of first segmentation vector sequences in the second target segmentation vector sequence set, and the first preset length. This process continues until all first segmentation vector sequences in the second target segmentation vector sequence set have been selected, resulting in the distance quantity proportion of all first segmentation vector sequences in the second target segmentation vector sequence set. The fourth filtering threshold is determined by the standard deviation of the fourth target segmentation vector sequence and the preset threshold coefficient. The average of the distance quantity proportions of all first segmentation vector sequences in the second target segmentation vector sequence set is taken as the distance quantity proportion of the second target segmentation vector sequence set.

[0053] In one embodiment of this application, the process of determining the distance quantity ratio of the first target segmentation vector sequence set based on the distance value associated with each first segmentation vector sequence in the first target segmentation vector sequence set and the number of first segmentation vector sequences in the first target segmentation vector sequence set includes:

[0054] One first segmentation vector sequence is selected from the set of first target segmentation vector sequences as the first target segmentation vector sequence. In one embodiment of this application, the length of the first segmentation vector sequence is a first preset length. The selection of the first target segmentation vector sequence can be random, or it can be selected according to the chronological order of all the first segmentation vector sequences in the set of first target segmentation vector sequences.

[0055] The number of distance values ​​associated with the first target segmentation vector sequence that are greater than a first filtering threshold is counted and denoted as the first filtering distance count. The first target segmentation vector sequence is then removed from the first target segmentation vector sequence set to obtain the number of first remaining segmentation vector sequences in the first target segmentation vector sequence set. The product of the number of first remaining segmentation vector sequences and a first preset length is taken as the total amount of first remaining data in the first target segmentation vector sequence set. In one embodiment of this application, the first filtering threshold is determined by the standard deviation of the first target segmentation vector sequence and a preset threshold coefficient. The calculation formula for the first filtering threshold is as follows:

[0056] F1 = r * SD1 (Equation 3)

[0057] Where F1 represents the first screening threshold, r represents the preset threshold coefficient, and SD1 represents the standard deviation of the first target segmentation vector sequence.

[0058] The ratio of the first filtered distance count to the total amount of the first remaining data is used as the proportion of the distance count in the first target segmentation vector sequence. In one embodiment of this application, the formula for calculating the proportion of the distance count in the first target segmentation vector sequence is as follows:

[0059] C m (1)=sumd1 / sumt1 Formula (4)

[0060] Among them, C m (1) represents the proportion of distances in the first target segmentation vector sequence, sumd1 represents the first filtering distance, and sumt1 represents the first remaining data.

[0061] Another first segmentation vector sequence from the first target segmentation vector sequence set is selected as the second target segmentation vector sequence. Based on the distance value associated with the second target segmentation vector sequence, a second filtering threshold, the number of first segmentation vector sequences in the first target segmentation vector sequence set, and a first preset length, the distance percentage of the second target segmentation vector sequence is determined. This process continues until all first segmentation vector sequences in the first target segmentation vector sequence set have been selected, resulting in the distance percentage of all first segmentation vector sequences in the first target segmentation vector sequence set. In one embodiment of this application, the second filtering threshold is determined by the standard deviation of the second target segmentation vector sequence and a preset threshold coefficient. The formula for calculating the second filtering threshold is as follows:

[0062] F2 = r * SD2 (5)

[0063] Where F2 represents the second screening threshold, r represents the preset threshold coefficient, and SD2 represents the standard deviation of the second target segmentation vector sequence.

[0064] In another embodiment of this application, the process of determining the distance quantity ratio of the second target segmentation vector sequence based on the distance value associated with the second target segmentation vector sequence, the second filtering threshold, the number of first segmentation vector sequences in the first target segmentation vector sequence set, and the first preset length includes: counting the number of distance values ​​greater than the second filtering threshold among the distance values ​​associated with the second target segmentation vector sequence, and recording them as the second filtering distance quantity; deleting the second target segmentation vector sequence from the first target segmentation vector sequence set to obtain the number of second remaining segmentation vector sequences in the first target segmentation vector sequence set, and taking the product of the number of second remaining segmentation vector sequences and the first preset length as the total amount of second remaining data in the first target segmentation vector sequence set; and taking the ratio of the second filtering distance quantity to the total amount of second remaining data as the distance quantity ratio of the second target segmentation vector sequence.

[0065] The average of the distance percentages of all first segmentation vector sequences in the first target segmentation vector sequence set is taken as the distance percentage of the first target segmentation vector sequence set. In one embodiment of this application, if there are Y first segmentation vector sequences in the first target segmentation vector sequence set, then the distance percentage of the first target segmentation vector sequence set is the average of the distance percentages of the Y first segmentation vector sequences.

[0066] In one embodiment of this application, the process of determining the sample entropy of each time-series component based on the distance quantity proportion of each first segmentation vector sequence set and the distance quantity proportion of each second segmentation vector sequence set includes:

[0067] A time-series component is selected as the first target time-series component. Based on the proportion of distances in the first segmentation vector sequence set corresponding to the first target time-series component and the proportion of distances in the second segmentation vector sequence set corresponding to the first target time-series component, the sample entropy of the first target time-series component is calculated. In one embodiment of this application, the formula for calculating the sample entropy of the first target time-series component is as follows:

[0068]

[0069] Where SampEn(t) represents the sample entropy of the first target time-series component. This represents the proportion of distances in the first segmentation vector sequence set corresponding to the first target temporal component. This represents the proportion of distances in the second segmentation vector sequence set corresponding to the first target temporal component.

[0070] Another time-series component is selected as the second target time-series component. The sample entropy of the second target time-series component is calculated based on the proportion of distances in the first segmentation vector sequence set corresponding to the second target time-series component and the proportion of distances in the second segmentation vector sequence set corresponding to the second target time-series component. This process is repeated until all time-series components have been selected, resulting in the sample entropy of each time-series component. In one embodiment of this application, the process of calculating the sample entropy of the second target time-series component based on the proportion of distances in the first segmentation vector sequence set corresponding to the second target time-series component and the proportion of distances in the second segmentation vector sequence set corresponding to the second target time-series component is the same as the process of calculating the sample entropy of the first target time-series component based on the proportion of distances in the first segmentation vector sequence set corresponding to the first target time-series component and the proportion of distances in the second segmentation vector sequence set corresponding to the first target time-series component.

[0071] In one embodiment of this application, if the sample components include: trend term time-series sample components, period term time-series sample components, and residual term time-series sample components, then the process of training a preset adversarial autoencoder network model based on the sample components to obtain an adversarial autoencoder network model includes:

[0072] A trend adversarial autoencoder (TAG) model is obtained by training a pre-defined adversarial autoencoder (ADAG) model with trend term time-series sample components. In one embodiment of this application, the ATG model is used to output an estimated component of the trend term time-series sample components as input. The pre-defined ATG model includes an encoder, a first decoder, and a second decoder. The trend term time-series sample components are input to the encoder. After the trend term time-series sample components are input to the encoder, a trend term encoded component is obtained. The trend term encoded component is input to the first decoder. After the trend term encoded component is input to the first decoder, a trend term decoded component is obtained. The trend term decoded component is input to the second decoder. After the trend term decoded component is input to the second decoder, an estimated component of the trend term time-series sample components is obtained.

[0073] A pre-defined adversarial autoencoder (AAE) model is obtained by training a pre-defined adversarial autoencoder model using periodic term temporal sample components. In one embodiment of this application, the pre-defined AAE model is used to output an estimated component of the periodic term temporal sample components when the periodic term temporal sample components are used as input. The pre-defined AAE model includes an encoder, a first decoder, and a second decoder. The periodic term temporal sample components are input to the encoder. After the periodic term temporal sample components are input to the encoder, a periodic term encoded component is obtained. The periodic term encoded component is input to the first decoder. After the periodic term encoded component is input to the first decoder, a periodic term decoded component is obtained. The periodic term decoded component is input to the second decoder. After the periodic term decoded component is input to the second decoder, an estimated component of the periodic term temporal sample components is obtained.

[0074] A residual adversarial autoencoder (SAE) network model is obtained by training a pre-defined adversarial autoencoder model using residual term temporal sample components. In one embodiment of this application, the SAE network model is used to output an estimated component of the residual term temporal sample components when the residual term temporal sample components are used as input. The pre-defined SAE network model includes an encoder, a first decoder, and a second decoder. The residual term temporal sample components are input to the encoder. After the residual term temporal sample components are input to the encoder, the residual term encoded components are obtained. The residual term encoded components are input to the first decoder. After the residual term encoded components are input to the first decoder, the residual term decoded components are obtained. The residual term decoded components are input to the second decoder. After the residual term decoded components are input to the second decoder, the estimated component of the residual term temporal sample components is obtained.

[0075] A trend adversarial autoencoder (AACE) model, a periodic adversarial autoencoder (PAA) model, and a residual adversarial autoencoder (RASA) model are combined to obtain an adversarial autoencoder (RASA) model. In one embodiment of this application, after obtaining the superimposed components, the trend term time series component is input into the trend AACE model to obtain the estimated component of the trend term time series component; the periodic term time series component is input into the periodic adversarial autoencoder (PAA) model to obtain the estimated component of the periodic term time series component; and the residual term time series component is input into the residual adversarial autoencoder (RASA) model to obtain the estimated component of the residual term time series component.

[0076] In one embodiment of this application, if the preset adversarial autoencoder network model includes: a first preset adversarial autoencoder network model and a second preset adversarial autoencoder network model, then the process of training the preset adversarial autoencoder network model using the trend term time-series sample components to obtain the trend adversarial autoencoder network model includes:

[0077] The trend term time series sample components are input into a first preset adversarial autoencoder network model to obtain the first-order estimated components of the trend term time series sample components; and the first-order estimated components of the trend term time series sample components are input into a second preset adversarial autoencoder network model to obtain the second-order estimated components of the trend term time series sample components. In one embodiment of this application, the parameter values ​​in the first preset adversarial autoencoder network model are different from the parameter values ​​in the second preset adversarial autoencoder network model. The parameter values ​​in the first preset adversarial autoencoder network model include adjustable hyperparameter values, which are set according to actual conditions. The hyperparameter values ​​in the second preset adversarial autoencoder network model also include adjustable hyperparameter values, which are set according to actual conditions.

[0078] A loss function is determined based on the difference between the time-series sample components of the trend term and their second-order estimated components. A first preset hyperparameter is used as a hyperparameter in the loss function to obtain a first preset loss function, and a second preset hyperparameter is used as a hyperparameter in the loss function to obtain a second preset loss function. In one embodiment of this application, the first preset hyperparameter and the second preset hyperparameter are set according to actual conditions. The expression for the loss function is as follows:

[0079]

[0080]

[0081] in, Let W represent the loss function. i Represents the time series sample components of the trend term, AE2(AE1(W i )) represents the second-order estimated component of the time series sample component of the trend term, AE1(Wi ) represents the first-order estimated component of the trend term time series sample component, AE1(·) represents the first preset adversarial autoencoder network model, AE2(·) represents the second preset adversarial autoencoder network model, and δ represents the absolute error threshold.

[0082] In one embodiment of this application, the process of determining the loss function based on the difference between the trend term time series sample component and the second-order estimated component of the trend term time series sample component includes: if the absolute value of the difference between the trend term time series sample component and the second-order estimated component of the trend term time series sample component is less than or equal to the absolute error threshold, then the first function in formula (7) is used as the loss function; if the absolute value of the difference between the trend term time series sample component and the second-order estimated component of the trend term time series sample component is greater than the absolute error threshold, then the second function in formula (7) is used as the loss function. Specifically, when the absolute value of the difference between the trend term time series sample component and its second-order estimated component is less than or equal to the absolute error threshold, the product of the square of the difference between the trend term time series sample component and its second-order estimated component and half of the difference is used as a function to calculate the reconstruction loss value during the training of the preset adversarial autoencoder network model. When the absolute value of the difference between the trend term time series sample component and its second-order estimated component is greater than the absolute error threshold, the product of the absolute value of the difference between the trend term time series sample component and its second-order estimated component and the absolute error threshold is calculated as the first product; the product of the square of the absolute error threshold and its half is calculated as the second product. The difference between the first product and the second product is used as a function to calculate the reconstruction loss value during the training of the preset adversarial autoencoder network model.

[0083] The trend term time-series sample components and their second-order estimated components are used as inputs to a first preset loss function, and the first preset adversarial autoencoder network model is trained with the goal of maximizing the output of the first preset loss function. Conversely, the trend term time-series sample components and their second-order estimated components are used as inputs to a second preset loss function, and the second preset adversarial autoencoder network model is trained with the goal of minimizing the output of the second preset loss function, thus obtaining a trend adversarial autoencoder network model. In one embodiment of this application, the functional expression for training the preset adversarial autoencoder network model is as follows:

[0084]

[0085] Among them, W i Represents the time series sample components of the trend term, AE2(AE1(W i )) represents the second-order estimated component of the time series sample component of the trend term, AE1(W i) represents the first-order estimated component of the trend term time series sample component, AE1(·) represents the first pre-defined adversarial autoencoder network model, and AE2(·) represents the second pre-defined adversarial autoencoder network model. This represents the first preset loss function. This represents the second preset loss function.

[0086] In one embodiment of this application, the process of training a preset adversarial autoencoder network model using periodic term time-series sample components to obtain a periodic adversarial autoencoder network model is the same as the process of training a preset adversarial autoencoder network model using trend term time-series sample components to obtain a trend adversarial autoencoder network model. Similarly, the process of training a preset adversarial autoencoder network model using residual term time-series sample components to obtain a residual adversarial autoencoder network model is the same as the process of training a preset adversarial autoencoder network model using trend term time-series sample components to obtain a trend adversarial autoencoder network model.

[0087] In one embodiment of this application, the process of calculating the score of the estimated component of the superimposed component includes:

[0088] The estimated components of the superimposed components are input into a first preset adversarial autoencoder network model to obtain a first-order estimated component of the superimposed components; the first-order estimated component of the superimposed components is then input into a second preset adversarial autoencoder network model to obtain a second-order estimated component of the superimposed components. In one embodiment of this application, the first preset adversarial autoencoder network model includes an encoder, a first decoder, and a second decoder; the second preset adversarial autoencoder network model includes an encoder, a first decoder, and a second decoder; and the parameter values ​​in the first preset adversarial autoencoder network model are different from the parameter values ​​in the second preset adversarial autoencoder network model.

[0089] The third preset hyperparameter is used as a hyperparameter in the loss function to obtain the third preset loss function, and the fourth preset hyperparameter is used as a hyperparameter in the loss function to obtain the fourth preset loss function. In one embodiment of this application, the third preset hyperparameter is different from the first preset hyperparameter, the third preset hyperparameter is different from the second preset hyperparameter, the fourth preset hyperparameter is different from the first preset hyperparameter, and the fourth preset hyperparameter is different from the second preset hyperparameter.

[0090] A scoring function is constructed based on the third and fourth preset loss functions; the estimated components of the superimposed components, the first-order estimated components of the superimposed components, and the second-order estimated components of the superimposed components are input into the scoring function to obtain the score of the estimated components of the superimposed components. In one embodiment of this application, the expression of the scoring function is as follows:

[0091]

[0092] in, Represents the scoring function. The score represents the estimated component of the superimposed components. This represents the estimated component of the superposition components. This represents the first-order estimated component of the superposition components. Let AE1(·) represent the second-order estimated component of the superimposed component, AE2(·) represent the first pre-defined adversarial autoencoder network model, and AE2(·) represent the second pre-defined adversarial autoencoder network model. This represents the third preset loss function. Let α represent the fourth preset loss function, β represent the first preset weight coefficient, and β represent the second preset weight coefficient.

[0093] In one embodiment of this application, if the preset scoring thresholds include: a trend item scoring threshold, a period item scoring threshold, and a residual item scoring threshold, then the process of determining abnormal segments of time series data based on the comparison result of the scores of the estimated components of the superimposed components with the preset scoring thresholds includes:

[0094] If the score of the estimated component of the trend term time series component is greater than the trend term score threshold, the time series segment corresponding to the trend term time series component is determined to be an abnormal segment of the time series data. In one embodiment of this application, after obtaining the score of the estimated component of the trend term time series component, the score of the estimated component of the trend term time series component is input into a quantile function to obtain the trend term score threshold. The calculation formula for the trend term score threshold is as follows:

[0095]

[0096] Where T represents the trend item rating threshold, Q 0.99 (·) denotes the quantile function, Q 0.99 The quantiles (·) are set according to the actual situation, and A(·) represents the scoring function. This represents the score of the estimated component of the superimposed component.

[0097] If the score of the estimated component of the periodic term time series component is greater than the periodic term score threshold, the time series segment corresponding to the periodic term time series component is determined to be an abnormal segment of the time series data. In one embodiment of this application, after obtaining the score of the estimated component of the periodic term time series component, the score of the estimated component of the periodic term time series component is input into a quantile function to obtain the periodic term score threshold. The calculation formula for the periodic term score threshold is the same as the calculation formula for the trend term score threshold.

[0098] If the score of the estimated component of the residual term's time series component is greater than the residual term score threshold, then the time series segment corresponding to the residual term's time series component is determined to be an outlier segment of the time series data. In one embodiment of this application, after obtaining the score of the estimated component of the residual term's time series component, the score of the estimated component of the residual term's time series component is input into a quantile function to obtain the residual term score threshold. The calculation formula for the residual term score threshold is the same as the calculation formula for the trend term score threshold.

[0099] Figure 3 This is a flowchart illustrating a method for detecting abnormal static pressure in a blast furnace body, as shown in another exemplary embodiment of this application. Figure 3 As shown, the method for detecting anomalies in blast furnace static pressure includes: performing VMD decomposition on the time-series data of blast furnace static pressure to obtain multiple time-series components; calculating the sampling entropy of each time-series component; performing k-means clustering on the sampling entropy of each time-series component; combining time-series components of the same type according to the clustering results to obtain trend, periodic, and residual time-series components; inputting the trend, periodic, and residual time-series components into an adversarial autoencoder network model to obtain estimated components of the combined components; calculating the score of the estimated components of the combined components; determining the abnormal segments of the time-series data based on the comparison between the score of the estimated components of the combined components and a preset score threshold; and merging the abnormal segments of the time-series data. Results) are obtained to detect abnormal static pressure in the blast furnace body. In this data, Trend represents the trend term time-series component, Seasonal represents the period term time-series component, Residual represents the residual term time-series component, Encoder represents the encoder, Decoder1 represents the first decoder, and Decoder2 represents the second decoder.

[0100] In one embodiment of this application, the trend term time series component is beneficial for reflecting the overall trend of time series data, the period term time series component is beneficial for reflecting the periodic changes of time series data, and the residual term time series component is beneficial for reflecting the detailed fluctuations of time series data.

[0101] In one embodiment of this application, the loss function is determined based on the difference between the trend term time-series sample component and its second-order estimated component. This helps to control the sensitivity to outlier data through an absolute error threshold and enhances robustness to outliers. A trend adversarial autoencoder (AACE) model is obtained by training a pre-defined adversarial autoencoder model using the trend term time-series sample component, and the trend term time-series component is estimated using this model. Similarly, a periodic adversarial autoencoder model is obtained by training a pre-defined adversarial autoencoder model using the periodic term time-series sample component, and the periodic term time-series component is estimated using this model. Furthermore, a residual adversarial autoencoder model is obtained by training a pre-defined adversarial autoencoder model using the residual term time-series sample component, and the residual term time-series component is estimated using this model. This approach can adapt to various anomalies in time-series data, such as trend shifts, fluctuations, or occasional extreme values.

[0102] The following describes an embodiment of the apparatus described in this application, which can be used to execute the blast furnace body static pressure anomaly detection method described in the above embodiments of this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the blast furnace body static pressure anomaly detection method described in the above applications.

[0103] Figure 4 This is a block diagram illustrating an exemplary embodiment of a blast furnace body static pressure anomaly detection device. This device can be applied to... Figure 1 The implementation environment shown is specifically configured in computer device 102. This device can also be applied to other exemplary implementation environments and specifically configured in other devices. This embodiment does not limit the implementation environment to which the device is applicable.

[0104] like Figure 4 As shown, the exemplary blast furnace body static pressure abnormality detection device 400 includes:

[0105] The data acquisition module 401 is used to acquire the time-series data of the static pressure of the blast furnace body.

[0106] The data decomposition module 402 is used to decompose the time series data to obtain multiple time series components; calculate the sample entropy of each time series component, cluster the sample entropy of each time series component, and according to the clustering results, superimpose the time series components belonging to the same type to obtain superimposed components.

[0107] The component estimation module 403 is used to input the superimposed components into the adversarial autoencoder network model to obtain the estimated components of the superimposed components.

[0108] The anomaly detection module 404 is used to calculate the score of the estimated component of the superimposed component. Based on the comparison result of the score of the estimated component of the superimposed component with the preset score threshold, the abnormal segments of the time series data are determined. The abnormal segments of the time series data are merged to obtain the anomaly detection result of the static pressure of the blast furnace body.

[0109] In one embodiment of this application, the blast furnace static pressure is used to characterize the pressure of the gas inside the blast furnace in a static state. The blast furnace static pressure data is measured by pressure sensors uniformly installed around the cooling walls of the blast furnace layers. After obtaining the blast furnace static pressure data, the data is sorted, deduplicated, and missing values ​​are filled in chronological order to obtain the time-series data of the blast furnace static pressure.

[0110] In one embodiment of this application, the superimposed components include: a trend term time series component, a period term time series component, and a residual term time series component. The time series data is decomposed using the VMD (Variational Mode Decomposition) method. VMD is a non-recursive signal processing method that decomposes time series data into a series of intrinsic mode functions (IMFs) with finite bandwidth. Compared with the empirical mode decomposition (EMD) method, the VMD method can effectively avoid mode aliasing by iteratively optimizing the IMF components, making the decomposition results more accurate and stable. The objective function of VMD is shown in the following formula (1).

[0111] In another embodiment of this application, the method for clustering the sample entropy of each time series component is k-means clustering. The clustering results are used to characterize which time series components belong to the same type and which belong to different types. Time series components belonging to the same type are then superimposed to obtain superimposed components. The types of time series components include: trend terms, periodic terms, and residual terms.

[0112] In one embodiment of this application, the adversarial autoencoder (AACE) model is trained on a preset AACE model based on sample components. The AACE model includes a trend AACE model, a periodic AACE model, and a residual AACE model. The process of inputting the superimposed components into the AACE model to obtain estimated components of the superimposed components includes: inputting the trend term time series components into the trend AACE model to obtain estimated components of the trend term time series components; inputting the periodic term time series components into the periodic AACE model to obtain estimated components of the periodic term time series components; and inputting the residual term time series components into the residual AACE model to obtain estimated components of the residual term time series components.

[0113] In one embodiment of this application, the process of calculating the score of the estimated component of the superimposed component includes: inputting the estimated component of the superimposed component into a first preset adversarial autoencoder network model to obtain a first-order estimated component of the superimposed component; inputting the first-order estimated component of the superimposed component into a second preset adversarial autoencoder network model to obtain a second-order estimated component of the superimposed component; using a third preset hyperparameter as a hyperparameter in the loss function to obtain a third preset loss function, and using a fourth preset hyperparameter as a hyperparameter in the loss function to obtain a fourth preset loss function; constructing a scoring function based on the third preset loss function and the fourth preset loss function; and inputting the estimated component of the superimposed component, the first-order estimated component of the superimposed component, and the second-order estimated component of the superimposed component into the scoring function to obtain the score of the estimated component of the superimposed component.

[0114] In one embodiment of this application, time-series data is decomposed, clustered, and superimposed to obtain superimposed components. An estimated component of the superimposed components is output through an adversarial autoencoder network model. After obtaining the score of the estimated component of the superimposed components, the abnormal segments of the time-series data are determined based on the comparison result of the score of the estimated component of the superimposed components with the preset score threshold. Thus, the abnormal detection result of the blast furnace body static pressure is obtained. This realizes the abnormal detection of blast furnace body static pressure through a standardized and unified abnormal detection method, which solves the technical problems of no unified standard, non-quantitative nature of monitoring abnormal state of blast furnace body static pressure, and high workload caused by manual monitoring. This is conducive to ensuring the stability of blast furnace production and control.

[0115] It should be noted that the blast furnace body static pressure anomaly detection device and the blast furnace body static pressure anomaly detection method provided in the above embodiments belong to the same concept. The specific operation methods of each module and unit have been described in detail in the method embodiments and will not be repeated here. In practical applications, the blast furnace body static pressure anomaly detection device provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0116] Embodiments of this application also provide an electronic device, including: one or more processors; and a storage device for storing one or more programs, which, when executed by one or more processors, enable the electronic device to implement the blast furnace body static pressure anomaly detection method provided in the above embodiments.

[0117] Figure 5 This is a schematic diagram illustrating the structure of a computer system for an electronic device, as shown in an exemplary embodiment of this application. It should be noted that... Figure 5The computer system 500 of the electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0118] like Figure 5 As shown, the computer system 500 includes a Central Processing Unit (CPU) 501, which can perform various appropriate actions and processes, such as executing the methods described in the above embodiments, based on programs stored in Read-Only Memory (ROM) 502 or programs loaded from storage portion 508 into Random Access Memory (RAM) 503. The RAM 503 also stores various programs and data required for system operation. The CPU 501, ROM 502, and RAM 503 are interconnected via a bus 504. An Input / Output (I / O) interface 505 is also connected to the bus 504.

[0119] The following components are connected to I / O interface 505: an input section 506 including a keyboard, mouse, etc.; an output section 507 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 508 including a hard disk, etc.; and a communication section 509 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 509 performs communication processing via a network such as the Internet. A drive 510 is also connected to I / O interface 505 as needed. Removable media 511, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 510 as needed so that computer programs read from them can be installed into storage section 508 as needed.

[0120] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including a computer program for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 509, and / or installed from removable medium 511. When the computer program is executed by central processing unit (CPU) 501, it performs various functions defined in the system of this application.

[0121] Another aspect of this application provides a computer-readable storage medium storing computer-readable instructions thereon. When these computer-readable instructions are executed by a computer's processor, the computer performs the blast furnace body static pressure anomaly detection method provided in the above embodiments. This computer-readable storage medium may be included in the electronic device described in the above embodiments, or it may exist independently and not be assembled into the electronic device.

[0122] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying a computer-readable computer program. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The computer program contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0123] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method for detecting abnormal static pressure in a blast furnace body, characterized in that, The method includes: Acquire time-series data of blast furnace body static pressure; the blast furnace body static pressure is used to characterize the pressure of gas in the blast furnace body under static conditions; The time series data is decomposed to obtain multiple time series components; the sample entropy of each time series component is calculated, and the sample entropy of each time series component is clustered. According to the clustering results, the time series components belonging to the same type are superimposed to obtain superimposed components. The superimposed components include: trend term time series components, period term time series components, and residual term time series components. The superimposed components are input into an adversarial autoencoder network model to obtain estimated components of the superimposed components; the adversarial autoencoder network model is obtained by training a preset adversarial autoencoder network model based on the sample components; Calculate the score of the estimated component of the superimposed component, and determine the abnormal segments of the time series data based on the comparison result of the score of the estimated component of the superimposed component with the preset score threshold; merge the abnormal segments of the time series data to obtain the abnormal detection result of the static pressure of the blast furnace body.

2. The method for detecting abnormal static pressure in the blast furnace body according to claim 1, characterized in that, The process of calculating the sample entropy of each time series component includes: Each time-series component is segmented according to a first preset length to obtain a first segmentation vector sequence set for each time-series component; the distance value between different first segmentation vector sequences in each first segmentation vector sequence set is calculated and denoted as the first distance value; the first distance value is then filtered according to the dimension of each first segmentation vector sequence to obtain the distance value associated with each first segmentation vector sequence; the distance quantity ratio of each first segmentation vector sequence set is determined based on the distance value associated with each first segmentation vector sequence and the number of first segmentation vector sequences in each first segmentation vector sequence set. According to the second preset length, each time component is segmented to obtain a second segmentation vector sequence set for each time component; the distance value between different second segmentation vector sequences in each second segmentation vector sequence set is calculated and denoted as the second distance value; and the second distance value is filtered according to the dimension of each second segmentation vector sequence to obtain the distance value associated with each second segmentation vector sequence; based on the distance value associated with each second segmentation vector sequence and the number of second segmentation vector sequences in each second segmentation vector sequence set, the distance quantity ratio of each second segmentation vector sequence set is determined; the second preset length is greater than the first preset length; The sample entropy of each time component is determined based on the proportion of distances in each first segmentation vector sequence set and the proportion of distances in each second segmentation vector sequence set.

3. The method for detecting abnormal static pressure in the blast furnace body according to claim 2, characterized in that, The process of determining the proportion of distances in each set of first segmentation vector sequences based on the distance value associated with each first segmentation vector sequence and the number of first segmentation vector sequences in each set of first segmentation vector sequences includes: Select a first segmentation vector sequence set as the first target segmentation vector sequence set; determine the distance quantity ratio of the first target segmentation vector sequence set based on the distance value associated with each first segmentation vector sequence in the first target segmentation vector sequence set and the number of first segmentation vector sequences in the first target segmentation vector sequence set; Select another first segmentation vector sequence set as the second target segmentation vector sequence set. Based on the distance value associated with each first segmentation vector sequence in the second target segmentation vector sequence set and the number of first segmentation vector sequences in the second target segmentation vector sequence set, determine the distance quantity ratio of the second target segmentation vector sequence set. Continue until all first segmentation vector sequence sets have been selected to obtain the distance quantity ratio of each first segmentation vector sequence set.

4. The method for detecting abnormal static pressure in the blast furnace body according to claim 3, characterized in that, The process of determining the proportion of distances in the first target segmentation vector sequence set based on the distance value associated with each first segmentation vector sequence in the first target segmentation vector sequence set and the number of first segmentation vector sequences in the first target segmentation vector sequence set includes: Select one first segmentation vector sequence from the first target segmentation vector sequence set as the first target segmentation vector sequence; The number of distance values ​​associated with the first target segmentation vector sequence that are greater than a first filtering threshold is counted and denoted as the first filtering distance number; the first target segmentation vector sequence is deleted from the first target segmentation vector sequence set to obtain the first remaining segmentation vector sequence number in the first target segmentation vector sequence set, and the product of the first remaining segmentation vector sequence number and the first preset length is taken as the first remaining data volume in the first target segmentation vector sequence set; the first filtering threshold is determined by the standard deviation of the first target segmentation vector sequence and the preset threshold coefficient; The ratio of the first number of filtered distances to the total amount of the first remaining data is taken as the proportion of the distances in the first target segmentation vector sequence; Another first segmentation vector sequence in the first target segmentation vector sequence set is selected as the second target segmentation vector sequence. The distance ratio of the second target segmentation vector sequence is determined based on the distance value associated with the second target segmentation vector sequence, a second filtering threshold, the number of first segmentation vector sequences in the first target segmentation vector sequence set, and the first preset length. This process continues until all first segmentation vector sequences in the first target segmentation vector sequence set have been selected, resulting in the distance ratio of all first segmentation vector sequences in the first target segmentation vector sequence set. The second filtering threshold is determined by the standard deviation of the second target segmentation vector sequence and the preset threshold coefficient. The average of the distance count percentages of all first segmentation vector sequences in the first target segmentation vector sequence set is taken as the distance count percentage of the first target segmentation vector sequence set.

5. The method for detecting abnormal static pressure in the blast furnace body according to claim 2, characterized in that, The process of determining the sample entropy of each time-series component based on the proportion of distances in each first segmentation vector sequence set and the proportion of distances in each second segmentation vector sequence set includes: Select a time series component as the first target time series component, and calculate the sample entropy of the first target time series component based on the distance quantity ratio of the first segmentation vector sequence set corresponding to the first target time series component and the distance quantity ratio of the second segmentation vector sequence set corresponding to the first target time series component. Select another time series component as the second target time series component. Based on the distance quantity ratio of the first segmentation vector sequence set corresponding to the second target time series component and the distance quantity ratio of the second segmentation vector sequence set corresponding to the second target time series component, calculate the sample entropy of the second target time series component. Continue until all time series components have been selected to obtain the sample entropy of each time series component.

6. The method for detecting abnormal static pressure in the blast furnace body according to any one of claims 1-5, characterized in that, If the sample components include: trend term time series sample components, period term time series sample components, and residual term time series sample components, then the process of training the preset adversarial autoencoder network model based on the sample components to obtain the adversarial autoencoder network model includes: The trend adversarial autoencoder network model is trained using the trend term time series sample components to obtain the trend adversarial autoencoder network model. The pre-defined adversarial autoencoder network model includes an encoder, a first decoder, and a second decoder. The trend adversarial autoencoder network model is used to output the estimated components of the trend term time series sample components when the trend term time series sample components are taken as input. The preset adversarial autoencoder network model is trained using the periodic term time series sample components to obtain a periodic adversarial autoencoder network model; the periodic adversarial autoencoder network model is used to output the estimated components of the periodic term time series sample components when the periodic term time series sample components are used as input. The preset adversarial autoencoder network model is trained using the residual term time series sample components to obtain a residual adversarial autoencoder network model; the residual adversarial autoencoder network model is used to output the estimated components of the residual term time series sample components when the residual term time series sample components are used as input. The trend adversarial autoencoder network model, the periodic adversarial autoencoder network model, and the residual adversarial autoencoder network model are combined to obtain the adversarial autoencoder network model.

7. The method for detecting abnormal static pressure in the blast furnace body according to claim 6, characterized in that, If the preset adversarial autoencoder network model includes: a first preset adversarial autoencoder network model and a second preset adversarial autoencoder network model, then the process of training the preset adversarial autoencoder network model using the trend term time series sample components to obtain the trend adversarial autoencoder network model includes: The trend term time series sample component is input into the first preset adversarial autoencoder network model to obtain the first-order estimated component of the trend term time series sample component; and the first-order estimated component of the trend term time series sample component is input into the second preset adversarial autoencoder network model to obtain the second-order estimated component of the trend term time series sample component; the parameter values ​​in the first preset adversarial autoencoder network model are different from the parameter values ​​in the second preset adversarial autoencoder network model; Based on the difference between the time series sample component of the trend term and the second-order estimated component of the time series sample component of the trend term, a loss function is determined; a first preset hyperparameter is used as a hyperparameter in the loss function to obtain a first preset loss function, and a second preset hyperparameter is used as a hyperparameter in the loss function to obtain a second preset loss function; The trend term time series sample component and its second-order estimated component are used as inputs to the first preset loss function, and the first preset adversarial autoencoder network model is trained with the goal of maximizing the output of the first preset loss function. The trend adversarial autoencoder network model is trained with the trend term time series sample component and its second-order estimated component as inputs to the second preset loss function, and the second preset adversarial autoencoder network model is trained with the goal of minimizing the output of the second preset loss function.

8. The method for detecting abnormal static pressure in the blast furnace body according to claim 7, characterized in that, The process of calculating the score of the estimated component of the superimposed component includes: The estimated component of the superimposed component is input into the first preset adversarial autoencoder network model to obtain the first-order estimated component of the superimposed component; the first-order estimated component of the superimposed component is input into the second preset adversarial autoencoder network model to obtain the second-order estimated component of the superimposed component. The third preset hyperparameter is used as a hyperparameter in the loss function to obtain the third preset loss function, and the fourth preset hyperparameter is used as a hyperparameter in the loss function to obtain the fourth preset loss function; Based on the third and fourth preset loss functions, a scoring function is constructed; and the estimated component of the superimposed component, the first-order estimated component of the superimposed component, and the second-order estimated component of the superimposed component are input into the scoring function to obtain the score of the estimated component of the superimposed component.

9. The method for detecting abnormal static pressure in the blast furnace body according to any one of claims 1-5, characterized in that, If the preset scoring thresholds include: a trend item scoring threshold, a period item scoring threshold, and a residual item scoring threshold, then the process of determining the abnormal segments of the time series data based on the comparison result of the estimated component scores of the superimposed components with the preset scoring thresholds includes: If the score of the estimated component of the trend item time series component is greater than the trend item score threshold, then the time series segment corresponding to the trend item time series component is determined to be an abnormal segment of the time series data. If the score of the estimated component of the periodic item time series component is greater than the periodic item score threshold, then the time series segment corresponding to the periodic item time series component is determined to be an abnormal segment of the time series data. If the score of the estimated component of the residual term time series component is greater than the residual term score threshold, then the time series segment corresponding to the residual term time series component is determined to be an abnormal segment of the time series data.

10. A device for detecting abnormal static pressure in a blast furnace body, characterized in that, include: The data acquisition module is used to acquire time-series data of the static pressure in the blast furnace body; The static pressure of the blast furnace body is used to characterize the pressure of the gas inside the blast furnace body in a static state. The data decomposition module is used to decompose the time-series data to obtain multiple time-series components; Calculate the sample entropy of each time series component, cluster the sample entropy of each time series component, and according to the clustering results, superimpose time series components belonging to the same type to obtain superimposed components. The superimposed components include: trend term time series components, period term time series components, and residual term time series components. The component estimation module is used to input the superimposed components into the adversarial autoencoder network model to obtain the estimated components of the superimposed components; the adversarial autoencoder network model is obtained by training a preset adversarial autoencoder network model based on the sample components; An anomaly detection module is used to calculate the score of the estimated component of the superimposed component, determine the abnormal segments of the time series data based on the comparison result of the score of the estimated component of the superimposed component with the preset score threshold, and merge the abnormal segments of the time series data to obtain the anomaly detection result of the static pressure of the blast furnace body.

Images