Cross-dimensional change point detection methods and systems for artificial coke coal preparation process analysis

Abstract

The application discloses a cross-dimension variable point detection method for artificial coke preparation process analysis, and comprises the following steps: obtaining historical operation data of an artificial coke preparation production test system and obtaining an artificial coke production data set based on the historical operation data; performing a first-stage multi-mode time sequence anomaly score learning on the artificial coke production data set to obtain multi-mode time sequence anomaly score information, and performing anomaly score smoothing processing on the multi-mode time sequence anomaly score information; performing a second-stage variable point score learning on the multi-mode time sequence anomaly score data after the anomaly score smoothing processing to obtain multi-dimension time sequence variable point score data, performing anomaly score smoothing processing, obtaining a candidate variable point set based on dynamic threshold analysis, and outputting corresponding process adjustment nodes; performing stage segmentation output, and accurately segmenting the artificial coke production data set into multiple independent process stages with the process adjustment nodes as boundaries and by combining and comparing process logs and business event labels. A corresponding detection system is also disclosed.

Description

Technical Field

[0001] This invention relates to the field of coal chemical process analysis and state monitoring technology, specifically to a cross-dimensional change point detection method and system for artificial coking coal preparation process analysis, particularly a cross-dimensional change point detection method and system based on a two-layer learning structure for artificial coking coal preparation process analysis. Background Technology

[0002] Coking coal, as a core raw material in iron and steel smelting, directly determines coke strength and blast furnace efficiency. However, high-quality coking coal resources are increasingly scarce globally (primary coking coal and fat coal account for less than 30%), while low-caking coals (such as gas coal and lean coal) are difficult to utilize directly due to their weak caking properties (G value < 50) and poor coking ability. Artificial coking coal technology transforms low-caking coal into high-caking coal substitutes through catalytic hydrogenation, possessing the dual strategic value of resource substitution and cost reduction and efficiency improvement.

[0003] The key to advancing artificial coking coal technology from research feasibility to large-scale engineering application lies in conducting extensive process condition experiments and parameter optimization for the catalytic hydrogenation upgrading process. To reduce the costs of frequent start-ups and shutdowns and accelerate verification, coal chemical pilot plants need to continuously adjust process parameters under unsteady and variable operating conditions. Verification of a single process scheme requires dynamic segmentation of continuous production data, but existing technologies face two major bottlenecks:

[0004] 1. High dependence on manual labor: The process team needs to analyze dozens to hundreds of sensor parameters one by one, which is a heavy workload and easy to miss key equipment data, resulting in detection delays and inaccurate positioning;

[0005] 2. Fragmentation of multi-source information: Traditional methods ignore the spatiotemporal correlation of multi-source sensor signals (such as temperature, pressure, and flow rate). Isolated analysis makes it difficult to capture the dynamic inertia and nonlinear reaction processes of production systems, which limits the depth of multimodal collaborative analysis.

[0006] To address the above challenges, cross-dimensional change point detection technology integrating multi-source sensing is key to overcoming the technical problems.

[0007] 1. Precise segmentation of process stages: By automatically identifying abrupt changes in the statistical characteristics of distributed sensor data (such as abrupt changes in mean and variance), and combining process logs and business event tags, continuous production data is automatically segmented to locate process adjustment nodes;

[0008] 2. Multimodal collaborative analysis: Utilizing spatiotemporal dependency modeling, it captures the linkage effect of system inertia and nonlinearity, and improves the ability to identify structural changes;

[0009] 3. Supporting closed-loop process optimization: By correlating variable points with the root causes of production fluctuations (such as catalyst activity decay and localized reactor overheating), it provides data-driven basis for process parameter tuning, enhancing the robustness and economy of the process package.

[0010] Therefore, in order to reduce the workload of researchers developing artificial coking coal preparation technology packages and improve the economy and robustness of artificial coking coal process packages, it is necessary to conduct research on cross-dimensional variable point detection methods by integrating multi-source sensor information for the task of artificial coking coal preparation process analysis and parameter optimization. Summary of the Invention

[0011] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a cross-dimensional variable point detection method and system for process analysis of artificial coking coal preparation. By comprehensively considering the spatiotemporal dependence between monitoring sensor data of the artificial coking coal preparation pilot system, and relying on the multi-mode time-series anomaly calculation and two-layer adaptive learning structure proposed in the embodiments of this invention, it better adapts to the characteristics of frequent process adjustments, unknown types of fluctuations and high-noise environments in the artificial coking coal preparation pilot scenario, so as to solve the problems of high manual dependence, process stage segmentation detection delay and positioning inaccuracy in existing process analysis.

[0012] A first aspect of the present invention is to provide a cross-dimensional change point detection method for the analysis of artificial coking coal preparation processes, comprising:

[0013] S1, acquire historical operating data of the artificial coking coal preparation pilot system and obtain an artificial coking coal production dataset based on the historical operating data;

[0014] S2, perform a first-stage multi-modal temporal anomaly score learning on the artificial coking coal production dataset to obtain multi-modal temporal anomaly score information, and perform anomaly score smoothing processing on the multi-modal temporal anomaly score information.

[0015] S3, perform a second stage of variable point score learning on the multi-mode time series abnormal score data after the abnormal score smoothing process to obtain multi-dimensional time series variable point score data, perform abnormal score smoothing process on the multi-dimensional time series variable point score data, and obtain a candidate variable point set based on dynamic threshold analysis and output the corresponding process adjustment node.

[0016] S4, perform stage segmentation output, including: using the process adjustment node as the boundary, and combining the comparison of process logs and business event tags, accurately segment the artificial coking coal production dataset into multiple independent process stages, and use the data of each process stage independently to support process package data analysis and process optimization.

[0017] Preferably, S1 includes:

[0018] S11, Obtain historical operating data of the artificial coking coal preparation pilot system, including: obtaining historical operating data of the artificial coking coal preparation pilot system within a preset time period; the historical operating data is one or more sets; the historical operating data includes coal slurry flow rate, hydrogen pressure, hydrogen flow rate, catalyst feeding, coke pusher displacement sensor signal, reactor temperature, reactor vibration signal, coal preheating temperature, methane / CO concentration, gaseous product temperature, reactor pressure difference, and valve opening; the historical operating data of the artificial coking coal preparation system is used to provide underlying data support for the artificial coking coal preparation process analysis;

[0019] S12, perform a first processing on the historical operation data, and obtain a structured sensor data information matrix based on the historical operation data after the first processing; the first processing includes aligning the acquisition intervals of different sensor signals in the historical operation data through signal sampling;

[0020] S13, the structured sensor data information matrix is ​​subjected to a second processing, and structured historical operation data is obtained based on the structured sensor data information matrix after the second processing; wherein, the second processing includes normalizing the sensor data information matrix; the structured historical operation data forms the artificial coking coal production dataset.

[0021] Preferably, the historical operating data of the artificial coking coal preparation pilot system obtained over a preset time period are represented by the following formula (1):

[0022] (1-1);

[0023] , (1-2);

[0024] , (1-3);

[0025] In equations (1-1) to (1-3), Represents the time measurement point, Represents the sensor serial number. This represents a matrix of historical operating data for an artificial coking coal preparation system. This represents the number of historical operational data collection groups. The number of signal sensors deployed in the artificial coking coal preparation system; Representing a point in time The sensor signal vector collected by the monitoring system of the artificial coking coal preparation system; Representing the Historical operating data of individual sensors, such as historical operating data of coal slurry flow sensors; This represents a point in time. No. The measured value of sensor number 1.

[0026] Preferably, after acquiring multiple sets of historical operating data of the artificial coking coal preparation system within a preset time period, the Z-score normalization method is used to normalize each sensor signal in the historical operating data. The Z-score normalization operation is defined by the following formula (2):

[0027] (2);

[0028] in, Represents the time measurement point, Represents the sensor serial number. Represents a point in time No. The original measurement value of sensor number 1 This represents the normalized result that eliminates the dimensional characteristics of the sensor. and They represent the first The average value and standard deviation of the historical operating data of sensor number 1.

[0029] Preferably, S2 includes:

[0030] S21, apply various types of time-series anomaly detection algorithms to the artificial coking coal production dataset to obtain multi-mode time-series anomaly score information of historical operating data. The multi-mode time-series anomaly score information is used to train a time-series anomaly model based on joint detection of multi-source information.

[0031] S22, introduce data smoothing processing into the multi-mode time series anomaly score information to filter transient interference and highlight continuous change characteristics, and obtain smoothed multi-mode time series anomaly score data.

[0032] Preferably, S21 includes:

[0033] S211, the linear deviation anomaly is calculated based on the linear deviation anomaly measurement formula (3) of the historical operation data of the artificial coking coal preparation system. The formula (3) is as follows:

[0034]

[0035]

[0036]

[0037] (3);

[0038] In the formula, Representing a point in time The sensor signal vector collected by the monitoring system of the artificial coking coal preparation system. Represents the VAR model for time points The predicted value of the sensor signal vector. The lag order of the VAR model. This represents the number of historical operational data collection groups. The number of signal sensors deployed in the artificial coking coal preparation system. For the initial sparse matrix, for A matrix of coefficients to be estimated in dimension . for 3D white noise vector;

[0039] S212, the Euclidean distance between the VAR model prediction result and the actual signal vector is used as the linear deviation anomaly score at the current time point, and its definition is shown in the following equation (4):

[0040] (4);

[0041] in, Represents the current time point The VAR linear deviation anomaly score is at the current time point. Actual data Compared with VAR model-based predictions Vector distance between them;

[0042] S213, the nonlinear deviation anomaly score is calculated based on the measurement formula (5) of the degree of nonlinear deviation anomaly of the historical operation data of the artificial coking coal preparation system. The formula (5) is as follows:

[0043]

[0044]

[0045] (5);

[0046] in, Represents the number of isolated trees in an isolated forest. Representative data points In the Path length in an isolated tree Representative data points Average path length in an isolated forest; Representing the The total number of leaf nodes in the isolation tree. It is the negative logarithm of the total number of leaf nodes in all isolated trees in the forest. Representative data points Nonlinear deviation anomaly score;

[0047] S214, the complexity deviation anomaly score is calculated using formula (6) to measure the degree of complex anomaly deviation based on the long-term dependence fluctuations and multivariate cointegration relationship changes of historical operating data of artificial coking coal preparation system. Formula (6) is as follows: (6);

[0048] in, For encoder model, For decoder models, Representing a point in time The sensor signal vector collected by the monitoring system of the artificial coking coal preparation system. This represents the LSTM-AE model for time points. The reconstructed value of the sensor signal vector, Represents the current time point The complexity deviates from the abnormal score;

[0049] S215, Calculate the multi-mode time-series anomaly score matrix of historical operating data of artificial coking coal preparation system based on formula (6). The formula (7) is as follows:

[0050]

[0051]

[0052] (7);

[0053] in, This indicates the number of timing anomaly detection algorithms used in the application. This represents the number of historical operational data collection groups. Represents the specific A series of time anomaly detection algorithms, Representing a point in time The corresponding data is in the first Anomaly scores under various time-series anomaly detection algorithms.

[0054] Preferably, S22 includes:

[0055] The smoothed multi-mode time-series anomaly score of the historical operating data of the artificial coking coal preparation system is calculated based on formula (8), which is as follows:

[0056] (8);

[0057] in, Indicates at a point in time Application No. Anomaly scores of various time-series anomaly detection algorithms This represents the multi-modal temporal anomaly score after smoothing by the corresponding temporal anomaly algorithm. Represents the moving average window size. These are the weighting coefficients.

[0058] Preferably, S3 includes:

[0059] S31, apply a sequential iterative autoregressive model to each dimension of the smoothed multi-mode time-series anomaly score data, and calculate the change point score of each dimension through residual analysis, including: calculating the change point score of the historical operation data of the artificial coking coal preparation system based on formula (9), wherein formula (9) is defined as follows:

[0060]

[0061]

[0062]

[0063]

[0064]

[0065] (9);

[0066] in, Represents a point in time Application No. The smoothed temporal anomaly scores of various temporal anomaly detection algorithms are used hereafter for formula simplification; the symbols will be used in the following text. Substitute , The coefficients to be estimated in the AR model. The lag order of the AR model. This is the discount rate coefficient, used to control the weighting of historical data. It is white noise. Represents the SIAR model for time points Predicted values ​​of time-series anomaly scores. The order is represented as Historical vectors, Update the weights for the AR model coefficients. Let covariance matrix be the variance matrix. Representing time In the Variable point scores under abnormal modes;

[0067] S32, performing anomaly score smoothing processing on the multi-dimensional time-series variable point score data, including: performing cascade optimization on the multi-dimensional time-series variable point score data through secondary moving average processing to generate the final smoothed multi-dimensional time-series variable point score data; obtaining a candidate variable point set based on dynamic threshold analysis and outputting the corresponding process adjustment nodes, including:

[0068] S321, Calculate the smoothed variable point score of the historical operating data of the artificial coking coal preparation system based on formula (10), where formula (10) is defined as follows:

[0069] (10);

[0070] in, Representative Time In the Smoothed variable point score under dimensional anomalous mode. Represents the moving average window size. The weighting coefficients are distributed according to a Gaussian distribution, with a weight of 0.5 at the center and decreasing to 0.1 at the edges;

[0071] S322, based on dynamic threshold analysis, potential process adjustment nodes are extracted as variable points, and the dynamic threshold calculation is defined as shown in the following equation (11):

[0072] (11);

[0073] in, Representing the The mean of the smoothed change-point score full-cycle sequence under the dimensional anomalous mode. Representing the Standard deviation of the smoothed change-point score full-cycle sequence under the dimensional anomalous mode. For the first The threshold for triggering change point scores under any anomalous mode; under any anomalous mode, if Then mark that time point. By identifying potential process adjustment nodes, a set of candidate change points is obtained.

[0074] Preferably, S4 includes:

[0075] S41, based on the candidate variable point set, perform an event association verification operation, including: comparing the timestamps of the candidate nodes with those of the process log events, allowing a maximum time deviation Δt = 10 seconds; if the time deviation between the candidate variable point and the log event is within Δt, it is marked as a valid process adjustment point; otherwise, it is considered a false variable point and is removed.

[0076] S42, Perform process segmentation output operation, including: using the effective process adjustment point as the boundary, segment the continuous production data into independent process stages to form a structured process dataset, with each stage data being used independently to support process package data analysis and process optimization.

[0077] A second aspect of the present invention is to provide a cross-dimensional variable point detection system for process analysis of artificial coking coal preparation, for implementing the method of the first aspect, comprising:

[0078] The production dataset acquisition module (101) is used to acquire historical operating data of the artificial coking coal preparation pilot system and obtain an artificial coking coal production dataset based on the historical operating data.

[0079] The multi-modal temporal anomaly score information module (102) is used to perform a first-stage multi-modal temporal anomaly score learning on the artificial coking coal production dataset to obtain multi-modal temporal anomaly score information, and to perform anomaly score smoothing processing on the multi-modal temporal anomaly score information.

[0080] The variable point score information module (103) is used to perform a second stage of variable point score learning on the multi-mode time series abnormal score data after the abnormal score smoothing process to obtain multi-dimensional time series variable point score data, perform abnormal score smoothing process on the multi-dimensional time series variable point score data, and obtain a candidate variable point set based on dynamic threshold analysis and output the corresponding process adjustment node.

[0081] The stage segmentation output module (104) is used to perform stage segmentation output, including: taking the process adjustment node as the boundary, combining the process log and business event tags, accurately segmenting the artificial coking coal production dataset into multiple independent process stages, and the data of each process stage is used independently to support process package data analysis and process optimization.

[0082] A third aspect of the present invention provides an electronic device including a processor and a memory, the memory storing a plurality of instructions, the processor being configured to read the instructions and execute the method as described in the first aspect.

[0083] A fourth aspect of the present invention provides a computer-readable storage medium storing a plurality of instructions which can be read by a processor and executed as described in the first aspect.

[0084] The beneficial effects of the method and system of the present invention are as follows:

[0085] This provides a comprehensive and effective solution for precise segmentation and detection in the artificial coking coal preparation process. It not only effectively considers the spatiotemporal dependencies between monitoring sensor data in the artificial coking coal preparation pilot system, but also better adapts to the frequent process adjustments, unknown fluctuations, and high-noise environment characteristics in the artificial coking coal preparation pilot scenario by introducing multi-mode time-series anomaly calculation and a two-layer adaptive learning structure. It can effectively solve the problems of high reliance on manual process analysis, delays in process segmentation and detection, and inaccurate positioning in existing processes. Attached Figure Description

[0086] To more clearly illustrate the technical solutions in the specific embodiments or related technologies of the present invention, the drawings used in the description of the specific embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0087] Figure 1 This is a flowchart of a cross-dimensional variable point detection method for the analysis of artificial coking coal preparation process based on existing technology;

[0088] Figure 2 This is a diagram of a cross-dimensional variable point detection system for analyzing the artificial coking coal preparation process according to an embodiment of the present invention.

[0089] Figure 3 This is a flowchart illustrating the steps of a cross-dimensional variable point detection method for artificial coking coal preparation process analysis based on a two-layer learning structure, according to an embodiment of the present invention.

[0090] Figure 4 The image shows the effect of the cross-dimensional change point detection method based on a two-layer learning structure provided in this embodiment of the invention on the detection of process change points of mean translation type.

[0091] Figure 5 The diagram shows the effect of the cross-dimensional variable point detection method based on a two-layer learning structure provided in this embodiment of the invention on the detection of process variable points with variance fluctuation.

[0092] Figure 6 This is a structural diagram of an electronic device provided according to an embodiment of the present invention. Detailed Implementation

[0093] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0094] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0095] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0096] like Figure 1 As shown, a first aspect of the present invention is to provide a cross-dimensional change point detection method for the analysis of artificial coking coal preparation processes, comprising:

[0097] S1, acquire historical operating data of the artificial coking coal preparation pilot system and obtain an artificial coking coal production dataset based on the historical operating data.

[0098] In a preferred embodiment, S1 includes:

[0099] S11, Obtain historical operating data of the artificial coking coal production pilot system, including: obtaining historical operating data of the artificial coking coal production pilot system within a preset time period; the historical operating data is one or more sets;

[0100] In some embodiments of this example, acquiring historical operating data of the artificial coking coal preparation system over a preset time period includes: continuously collecting and storing the operating data of the artificial coking coal preparation system. Each set of operating data includes multiple signal data collected by the artificial coking coal preparation system at that time point, and a preset time interval exists between any two adjacent time points. The preset time interval is the time interval for the artificial coking coal preparation system to collect data, for example, 5 seconds.

[0101] In some embodiments of this example, the historical operating data of the artificial coking coal preparation system includes various types such as coal slurry flow rate, hydrogen pressure, hydrogen flow rate, catalyst feeding, coke pusher displacement sensor signal, reactor temperature, reactor vibration signal, coal preheating temperature, methane / CO concentration, gaseous product temperature, reactor pressure difference, and valve opening. The historical operating data of the artificial coking coal preparation system is used to provide underlying data support for the analysis of the artificial coking coal preparation process.

[0102] In some embodiments of this example, the historical operating data of the artificial coking coal production pilot system obtained over a preset time period are represented by the following formula (1):

[0103] (1-1);

[0104] , (1-2);

[0105] , (1-3);

[0106] In equations (1-1) to (1-3), Represents the time measurement point, Represents the sensor serial number. This represents a matrix of historical operating data for an artificial coking coal preparation system. This represents the number of historical operational data collection groups. The number of signal sensors deployed in the artificial coking coal preparation system; Representing a point in time The sensor signal vector collected by the monitoring system of the artificial coking coal preparation system; Representing the Historical operating data of individual sensors, such as historical operating data of coal slurry flow sensors; This represents a point in time. No. The measured value of sensor number 1.

[0107] S12, perform a first processing on the historical operation data, and obtain a structured sensor data information matrix based on the historical operation data after the first processing; the first processing includes aligning the acquisition intervals of different sensor signals in the historical operation data through signal sampling;

[0108] S13, the structured sensor data information matrix is ​​subjected to a second processing, and structured historical operation data is obtained based on the structured sensor data information matrix after the second processing; wherein, the second processing includes normalizing the sensor data information matrix to eliminate the dimensional differences between different sensor signal features, thereby obtaining structured historical operation data for subsequent change point detection algorithms; the structured historical operation data forms the artificial coking coal production dataset.

[0109] In this embodiment, after acquiring multiple sets of historical operating data of the artificial coking coal preparation system within a preset time period, the Z-score normalization method is used to normalize each sensor signal in the historical operating data to eliminate the influence of differences in the numerical dimensions of different sensor signals on the subsequent model, thereby improving the stability of the change point detection model.

[0110] The Z-score normalization operation is defined by the following formula (2):

[0111] (2);

[0112] in, Represents the time measurement point, Represents the sensor serial number. Represents a point in time No. The original measurement value of sensor number 1 This represents the normalized result that eliminates the dimensional characteristics of the sensor. and They represent the first The average value and standard deviation of the historical operating data of sensor number 1.

[0113] S2, perform a first-stage multi-modal temporal anomaly score learning on the artificial coking coal production dataset to obtain multi-modal temporal anomaly score information, and perform anomaly score smoothing processing on the multi-modal temporal anomaly score information.

[0114] In a preferred embodiment, S2 includes:

[0115] S21, apply various types of time-series anomaly detection algorithms to the artificial coking coal production dataset to obtain multi-mode time-series anomaly score information of historical operating data. The multi-mode time-series anomaly score information is used to train a time-series anomaly model based on joint detection of multi-source information.

[0116] In a preferred embodiment, S21 includes:

[0117] S211, the linear deviation anomaly is calculated based on the linear deviation anomaly measurement formula (3) of the historical operation data of the artificial coking coal preparation system. The formula (3) is as follows:

[0118]

[0119]

[0120] (3);

[0121] In the formula, Representing a point in time The sensor signal vector collected by the monitoring system of the artificial coking coal preparation system. Represents the VAR model for time points The predicted value of the sensor signal vector. The lag order of the VAR model. This represents the number of historical operational data collection groups. The number of signal sensors deployed in the artificial coking coal preparation system. For the initial sparse matrix, for A matrix of coefficients to be estimated in dimension . for 3D white noise vector;

[0122] S212, the Euclidean distance between the VAR model prediction result and the actual signal vector is used as the linear deviation anomaly score at the current time point, and its definition is shown in the following equation (4):

[0123] (4);

[0124] in, Represents the current time point The VAR linear deviation anomaly score is at the current time point. Actual data Compared with VAR model-based predictions Vector distance between them;

[0125] S213, the nonlinear deviation anomaly score is calculated based on the measurement formula (5) of the degree of nonlinear deviation anomaly of the historical operation data of the artificial coking coal preparation system. The formula (5) is as follows:

[0126]

[0127] (5);

[0128] in, Represents the number of isolated trees in an isolated forest. Representative data points In the Path length in an isolated tree Representative data points Average path length in an isolated forest; Representing the The total number of leaf nodes in the isolation tree. It is the negative logarithm of the total number of leaf nodes in all isolated trees in the forest. Representative data points Nonlinear deviation anomaly score;

[0129] S214, the complexity deviation score is calculated using formula (6) to measure the degree of complex anomaly deviation based on the long-term dependence fluctuation and multivariate cointegration relationship changes of historical operating data of artificial coking coal preparation system. Formula (6) is as follows: (6);

[0130] in, For encoder model, For decoder models, Representing a point in time The sensor signal vector collected by the monitoring system of the artificial coking coal preparation system. This represents the LSTM-AE model for time points. The reconstructed value of the sensor signal vector, Represents the current time point The complexity deviates from the abnormal score.

[0131] S215, Calculate the multi-mode time-series anomaly score matrix of historical operating data of artificial coking coal preparation system based on formula (7). The formula (7) is as follows:

[0132]

[0133]

[0134] (7);

[0135] in, This indicates the number of timing anomaly detection algorithms used in the application. This represents the number of historical operational data collection groups. Represents the specific A series of time anomaly detection algorithms, Representing a point in time The corresponding data is in the first Anomaly scores under various time-series anomaly detection algorithms.

[0136] S22, introduce data smoothing processing into the multi-mode time series anomaly score information to filter transient interference and highlight continuous change characteristics, and obtain smoothed multi-mode time series anomaly score data.

[0137] In a preferred embodiment, S22 includes:

[0138] The smoothed multi-mode time-series anomaly score of the historical operating data of the artificial coking coal preparation system is calculated based on formula (8), which is as follows:

[0139] (8);

[0140] in, Indicates a point in time Application No. Anomaly scores of various time-series anomaly detection algorithms This represents the multi-modal temporal anomaly score after smoothing by the corresponding temporal anomaly algorithm. Represents the moving average window size. These are the weighting coefficients.

[0141] S3, perform a second stage of variable point score learning on the multi-mode time series abnormal score data after the abnormal score smoothing process to obtain multi-dimensional time series variable point score data, perform abnormal score smoothing process on the multi-dimensional time series variable point score data, and obtain a candidate variable point set based on dynamic threshold analysis and output the corresponding process adjustment node.

[0142] In a preferred embodiment, S3 includes:

[0143] S31, apply a sequential iterative autoregressive model to each dimension of the smoothed multi-mode time-series anomaly score data, and calculate the change point score of each dimension through residual analysis;

[0144] In a preferred embodiment, S31 includes: calculating the change point score of the historical operating data of the artificial coking coal preparation system based on formula (9), wherein formula (9) is defined as follows:

[0145]

[0146]

[0147]

[0148]

[0149]

[0150]

[0151] (9);

[0152] in, Represents a point in time Application No. The smoothed temporal anomaly scores of various temporal anomaly detection algorithms are used hereafter for formula simplification; the symbols will be used in the following text. Substitute , The coefficients to be estimated in the AR model. The lag order of the AR model. This is the discount rate coefficient, used to control the weighting of historical data. It is white noise. Represents the SIAR model for time points Predicted values ​​of time-series anomaly scores. The order is represented as Historical vectors, Update the weights for the AR model coefficients. Let covariance matrix be the variance matrix. Representing time In the Variable point scores under abnormal modes.

[0153] S32, perform abnormal score smoothing processing on the multi-dimensional time-series variable point score data, including: performing cascade optimization on the multi-dimensional time-series variable point score data through secondary moving average processing to generate the final smoothed multi-dimensional time-series variable point score data, obtaining a candidate variable point set based on dynamic threshold analysis and outputting the corresponding process adjustment node.

[0154] In a preferred embodiment, S32 includes:

[0155] S321, Calculate the smoothed variable point score of the historical operating data of the artificial coking coal preparation system based on formula (10), where formula (10) is defined as follows:

[0156] (10);

[0157] in, Representative Time In the Smoothed variable point score under dimensional anomalous mode. Represents the moving average window size. The weighting coefficients are distributed according to a Gaussian distribution, with a weight of 0.5 at the center and decreasing to 0.1 at the edges;

[0158] S322, based on dynamic threshold analysis, potential process adjustment nodes are extracted as variable points, and the dynamic threshold calculation is defined as shown in the following equation (11):

[0159] (11);

[0160] in, Representing the The mean of the smoothed change-point score full-cycle sequence under the dimensional anomalous mode. Representing the Standard deviation of the smoothed change-point score full-cycle sequence under the dimensional anomalous mode. For the first The threshold for triggering change point scores under any anomalous mode; under any anomalous mode, if Then mark that time point. By identifying potential process adjustment nodes, a set of candidate change points is obtained.

[0161] S4, perform stage segmentation output, including: using the process adjustment node as the boundary, and combining the comparison of process logs and business event tags, accurately segment the artificial coking coal production dataset into multiple independent process stages (such as "preheating stage" and "hydrogenation reaction stage"). The data of each process stage is used independently to support process package data analysis and process optimization.

[0162] In a preferred embodiment, S4 includes:

[0163] S41, based on the candidate variable point set, perform an event association verification operation, including: comparing the timestamps of candidate nodes with process log events (such as "catalyst injection" and "reactor pressurization"), allowing a maximum time deviation Δt = 10 seconds (due to sensor response delay and system pipeline lag effect); if the time deviation between the candidate variable point and the log event is within Δt, it is marked as a valid process adjustment point; otherwise, it is considered a false variable point and discarded.

[0164] S42, Perform process segmentation and output operations, including: dividing the continuous production data into independent process stages (such as "constant temperature reaction stage" and "pressure reduction separation stage") using the effective process adjustment points as boundaries, forming a structured process dataset. The data of each stage is independently used to support process package data analysis and process optimization.

[0165] like Figure 2 As shown, a second aspect of the present invention provides a cross-dimensional variable point detection system for analyzing the artificial coking coal preparation process, for implementing the method of the first aspect, comprising:

[0166] The production dataset acquisition module (101) is used to acquire historical operating data of the artificial coking coal preparation pilot system and obtain an artificial coking coal production dataset based on the historical operating data.

[0167] The multi-modal temporal anomaly score information module (102) is used to perform a first-stage multi-modal temporal anomaly score learning on the artificial coking coal production dataset to obtain multi-modal temporal anomaly score information, and to perform anomaly score smoothing processing on the multi-modal temporal anomaly score information.

[0168] The variable point score information module (103) is used to perform a second stage of variable point score learning on the multi-mode time series abnormal score data after the abnormal score smoothing process to obtain multi-dimensional time series variable point score data, perform abnormal score smoothing process on the multi-dimensional time series variable point score data, and obtain a candidate variable point set based on dynamic threshold analysis and output the corresponding process adjustment node.

[0169] The stage segmentation output module (104) is used to perform stage segmentation output, including: taking the process adjustment node as the boundary, combining the process log and business event tags, accurately segmenting the artificial coking coal production dataset into multiple independent process stages (such as "preheating stage" and "hydrogenation reaction stage"), and the data of each process stage is used independently to support process package data analysis and process optimization.

[0170] Example 1

[0171] Figure 3 This is a flowchart of a cross-dimensional change point detection method for artificial coking coal preparation process analysis based on a two-layer learning structure, provided by one embodiment of the present invention. Figure 3 As shown, this invention provides a method for detecting change points in the artificial coking coal preparation process based on a two-layer learning structure. The method includes:

[0172] Step S101: Obtain historical operating data of the artificial coking coal production pilot system. There may be one or more sets of historical operating data.

[0173] In some embodiments of this example, the historical operating data of the artificial coking coal preparation system includes more than ten types of data, such as coal slurry flow rate, hydrogen pressure, hydrogen flow rate, catalyst feeding, coke pusher displacement sensor signal, reactor temperature, reactor vibration signal, coal preheating temperature, methane / CO concentration, gaseous product temperature, reactor pressure difference, and valve opening, providing underlying data support for the analysis of the artificial coking coal preparation process.

[0174] In some embodiments of this example, acquiring historical operating data of the artificial coking coal preparation system includes: acquiring historical operating data of the artificial coking coal preparation system within a preset time period; aligning the acquisition intervals of different sensor signals in the historical operating data through signal sampling to obtain a structured sensor data information matrix; and normalizing the sensor data information matrix to obtain structured historical operating data.

[0175] In the above implementation, the sensor data information matrix is ​​normalized to eliminate the dimensional differences between different sensor signal features, thereby obtaining historical operating data for subsequent change point detection algorithms.

[0176] In some embodiments of this example, acquiring historical operating data of the artificial coking coal preparation system over a preset time period includes: continuously collecting and storing the operating data of the artificial coking coal preparation system. Each set of operating data includes multiple signal data collected by the artificial coking coal preparation system at that time point, and there is a preset time interval between any two adjacent time points. The preset time interval is the time interval for the artificial coking coal preparation system to collect data, for example, 5 seconds.

[0177] In some embodiments of this example, obtaining multiple sets of historical operating data of the artificial coking coal preparation system over a preset time period can be represented by the following formula:

[0178]

[0179]

[0180]

[0181] In the formula, This represents a matrix of historical operating data for an artificial coking coal preparation system. This represents the number of historical operational data collection groups. The number of signal sensors deployed in the artificial coking coal preparation system; Representing a point in time The sensor signal vector collected by the monitoring system of the artificial coking coal preparation system; Representing the Historical operating data of individual sensors, such as historical operating data of coal slurry flow sensors; This represents a point in time. No. The measured value of sensor number 1.

[0182] After acquiring multiple sets of historical operating data of the artificial coking coal preparation system over a preset time period, the Z-score standardization method is used to normalize each sensor signal in the historical operating data to eliminate the influence of differences in the numerical dimensions of different sensor signals on the subsequent model, thereby improving the stability of the change point detection model.

[0183] The Z-score normalization operation is defined by the following formula:

[0184]

[0185] in, Represents the time measurement point, Represents the sensor serial number. Represents a point in time No. The original measurement value of sensor number 1 This represents the normalized result that eliminates the dimensional characteristics of the sensor. and Indicates the first The average value and standard deviation of the historical operating data of sensor number 1.

[0186] Step S102: Based on multiple time-series anomaly detection algorithms, obtain multi-mode time-series anomaly score information of historical operation data of artificial coking coal preparation system.

[0187] In the pilot production of artificial coking coal, frequent process adjustments, nonlinear coupling of the production system, and a high-noise environment lead to frequent and difficult-to-distinguish abnormal fluctuations in historical operating data. This results in problems such as missed changes, detection delays, and inaccurate positioning in traditional process stage segmentation methods. To address these issues, this invention proposes a two-layer adaptive learning structure: the first stage obtains multi-modal anomaly scores by detecting temporal anomalies in historical data; the second stage further extracts change point features based on these scores to achieve accurate capture of process mutations.

[0188] In some implementations of this embodiment, firstly, various time-series anomaly detection algorithms are applied to comprehensively identify potential fluctuation patterns in system operation, including single-point anomalies, mean drift, trend reversals, and variance mutations. It is important to emphasize that time-series anomaly detection focuses on instantaneous behavioral deviations at a single measurement point, while change-point detection focuses on the migration of statistical characteristics across segments; time-series anomalies are more sensitive in terms of detection granularity. Through collaborative analysis of multiple algorithms, the aforementioned fluctuation patterns are quantitatively represented as multi-dimensional anomaly scores, completing the first stage of "multi-mode time-series anomaly score" learning and laying the data foundation for the second stage of "change-point score" learning.

[0189] In this implementation, a Vector Autoregression (VAR) model is used to assess the degree of linear anomaly deviation at historical time points, an Isolation Forest (iForest) detection algorithm is used to assess the degree of nonlinear anomaly deviation, and an LSTM-AE (Long Short-Term Memory neural network based AutoEncoders) algorithm is used to identify complex anomalies such as long-term dependency fluctuations and changes in multivariate cointegration relationships hidden in historical data. This results in the definition of multiple combinations of time-series anomaly detectors. :

[0190]

[0191] in, Indicates the number of primitive anomaly detection algorithms applied; These represent a time-series anomaly detection algorithm, such as the iForest anomaly detection algorithm for isolated forests.

[0192] Specifically, in this step, a vector autoregression (VAR) model is used to fit and learn the historical operating data. The prediction error of the VAR model results is used to quantitatively characterize the degree of linear deviation anomaly in the historical operating data of the artificial coking coal preparation system.

[0193] The VAR algorithm takes normalized historical running data as input, and the formula is defined as follows:

[0194] in, Representing a point in time The sensor signal vector collected by the monitoring system of the artificial coking coal preparation system. Represents the VAR model for time points The predicted value of the sensor signal vector. The lag order of the VAR model. This represents the number of historical operational data collection groups. The number of signal sensors deployed in the artificial coking coal preparation system. for A matrix of coefficients to be estimated in dimension . for 3D white noise vector.

[0195] Preferably, the Euclidean distance between the VAR model prediction result and the actual signal vector is used as the linear deviation anomaly score at the current time point, and its definition is as follows:

[0196]

[0197] in, Represents the current time point The VAR linear deviation anomaly score is at the current time point. Actual data Compared with VAR model-based predictions Calculate the vector distance.

[0198] Specifically, in this step, the Isolation Forest (iForest) algorithm is used to identify the degree of nonlinear deviation anomalies in historical running data.

[0199] The core idea of ​​iForest is to isolate data points by randomly partitioning the feature space. For the isolation trees in an isolated forest, the path length is defined as the number of layers a data point traverses from the root to the isolated leaf node. In normal data distributions, the path length of data points is usually long; however, outliers, due to sparse features and ease of isolation, have significantly shorter path lengths. Based on this characteristic, the iForest algorithm generates anomaly scores by calculating path lengths, effectively characterizing nonlinear deviation patterns in historical data, defined as follows:

[0200]

[0201]

[0202]

[0203] in, Represents the number of isolated trees in an isolated forest. Representative data points In the Path length in an isolated tree Representative data points Average path length in an isolated forest; Representing the The total number of leaf nodes in the isolation tree. It is the negative logarithm of the total number of leaf nodes in all isolated trees in the forest. Representative data points The nonlinear deviation anomaly score.

[0204] Specifically, in this step, an autoencoder algorithm based on a long short-term memory neural network (LSTM-AE) is used to identify complex anomalies and deviations in historical data, such as long-term dependency fluctuations and changes in multivariate cointegration relationships.

[0205] The core idea of ​​the LSTM-AE algorithm is to reconstruct time-series data using an LSTM autoencoder and use the reconstruction error as an anomaly score. Since outliers are difficult to represent effectively in a low-dimensional feature space, their reconstruction errors are significantly higher than those of normal data. Therefore, it is reasonable to use the reconstruction error of the LSTM-AE algorithm to quantify the complexity deviation of historical data points and the anomaly score.

[0206] Specifically, the LSTM-AE time series anomaly detection model comprises two modules: an encoder and a decoder. The encoder extracts and compresses the original multidimensional time series into a latent feature vector space; the decoder reconstructs the latent features into the output space, obtaining the reconstructed output data. The encoding and decoding process can be represented as follows:

[0207] in, For encoder model, For decoder models, Representing a point in time The sensor signal vector collected by the monitoring system of the artificial coking coal preparation system. This represents the LSTM-AE model for time points. The reconstructed value of the sensor signal vector, Represents the current time point The complexity deviates from the abnormal score.

[0208] The encoder module extracts the long-term dependencies and temporal evolution patterns among various sensing features in historical operation data through an LSTM network. The encoder acts as a sequence folding layer to obtain the latent features of historical operation data information.

[0209] The decoder module reconstructs the latent features output by the encoder using an LSTM network. Starting from the last time step, the decoder reconstructs the original input step by step in reverse. The latent state at each time point is affected by the latent state at the current and next time steps, and finally, the decoder outputs the reconstructed sequence of running data observations.

[0210] Furthermore, in the training process of the LSTM-AE model, mean squared error is used as the loss function, the Adam optimizer is used for parameter optimization, the learning rate is set to 0.001, and the number of training epochs is 100.

[0211] A combination of various types of time-series anomaly detectors is applied to the historical runtime data processed in step S101. The "multi-mode time series anomaly score" matrix of historical operating data The definition is as follows:

[0212]

[0213]

[0214]

[0215] in, This indicates the number of timing anomaly detection algorithms used in the application. This represents the number of historical operational data collection groups. This represents the r-th type of time series anomaly detection algorithm. Representing a point in time The corresponding data is in the first Anomaly scores under various time-series anomaly detection algorithms.

[0216] Step S103: Perform a moving average on the "multi-mode time series anomaly score" data output in step S102 to obtain smoothed multi-mode time series anomaly score data.

[0217] In the pilot production scenario of artificial coking coal preparation, the superposition of multiple factors such as proactive process adjustments, transient production disturbances, and electromechanical anomalies leads to both single-point anomalies and continuous changes (such as process adjustment changes) exhibiting high scores in the first stage of "multi-mode time-series anomaly score," causing drastic fluctuations in the anomaly score. To filter out transient interference (such as single-point anomalies) and highlight the core objective of process analysis—the characteristics of continuous changes—anomaly score smoothing processing needs to be introduced.

[0218] In some implementations of this embodiment, a moving average algorithm is used to smooth the multi-mode time-series anomaly scores obtained in step S102 to eliminate the influence of transient interference and random noise. The calculation method is defined as follows:

[0219]

[0220] in, Indicates at a point in time Application No. Anomaly scores of various time-series anomaly detection algorithms This represents the anomaly score data after smoothing by the corresponding time-series anomaly algorithm. Represents the moving average window size. These are the weighting coefficients.

[0221] In this embodiment, a targeted smoothing process for "multi-mode timing anomaly scores" is added to prevent instantaneous anomalies at a single measurement point from being misjudged as change points. This allows for more accurate analysis of persistent changes (process adjustment change points) in historical operational data. By suppressing random noise, stable multi-mode timing anomaly score data is output, providing a high signal-to-noise ratio input for subsequent change point detection.

[0222] Step S104: On the smoothed multi-mode time-series anomaly score data, apply the Sequentially Iterative Auto-Regression (SIAR) model to each dimension, and calculate the "change point score" under the corresponding anomaly mode based on residual analysis.

[0223] Based on the smoothed multi-modal time-series anomaly score data, a sequential iterative autoregressive (SIAR) model is applied to each dimension. The "change point score" of each dimension is calculated through residual analysis to complete the second stage of "change point score" learning. The iterative learning mechanism of the SIAR model can effectively distinguish between instantaneous anomalies (such as single-point noise) and continuous changes (such as process adjustment change points)—the former has random fluctuations in residuals, while the latter produces significantly high residuals due to the transfer of statistical characteristics.

[0224] Specifically, in this embodiment, autoregressive (AR) modeling is first performed on the single-dimensional (modal) anomaly score sequence, and the calculation process is defined as follows:

[0225] in, Represents a point in time Application No. The smoothed temporal anomaly scores of various temporal anomaly detection algorithms are used hereafter for formula simplification; the symbols will be used in the following text. Substitute , The coefficients to be estimated in the AR model. The lag order of the AR model. It is white noise. Represents the AR model at a given time point Predicted values ​​for time-series anomaly scores.

[0226] Furthermore, the Sequential Iterative Autoregressive (SIAR) model algorithm essentially iterates the process parameters sequentially based on the recursive least squares principle during the autoregressive modeling process. The parameter update formula is as follows:

[0227]

[0228]

[0229]

[0230]

[0231] in, This represents the smoothed time series anomaly score after applying the r-th time series anomaly detection algorithm at time point t. The order is represented as Historical vectors, Update the weights for the AR model coefficients. It is the covariance matrix; This is the discount rate coefficient, used to control the weight of historical data.

[0232] Preferably, the SIAR model is used to predict the results. and raw data The Hellinger distance is used to calculate the second-stage "variable point score," which is defined as follows:

[0233]

[0234]

[0235] in, This represents the anomaly score at a certain anomalous mode at time point t. This represents the predicted score of the time series anomaly at time point t by the SIAR model. The change point score represents the time t in the d-th dimension of the anomalous mode.

[0236] Step S105: Perform moving average processing again on the multi-dimensional variable point score data output in step S104 to form cascade optimization, and collect a set of candidate process variable points based on threshold analysis.

[0237] Based on the multi-dimensional variable point score data generated in step S104, this step achieves cascade optimization through secondary moving average processing. The purpose is to suppress pseudo fluctuations in the variable point score (such as brief equipment jitter and process fine-tuning disturbances), form a two-level noise filtering mechanism, and highlight continuous process adjustment signals. Figure 4 The data processing method described in this example can effectively identify continuous process adjustment signals of the mean shift type. Figure 5 The method shown in this embodiment can also effectively highlight persistent process adjustment signals of variance fluctuation type.

[0238] In some implementations of this example, the variable point score sequence for each dimension The weighted moving average is calculated as follows:

[0239]

[0240] in, Representative Time In the Smoothed variable point score under dimensional anomalous mode. Represents the moving average window size. The weighting coefficients are distributed according to a Gaussian distribution (0.5 for the center and decreasing to 0.1 at the edges).

[0241] Furthermore, based on dynamic threshold analysis, potential process adjustment nodes (change points) are identified and extracted. The threshold calculation is defined as follows:

[0242]

[0243] in, This represents the mean of the smoothed change-point score across the entire time series under the d-th dimension anomalous mode. Representing the Standard deviation of the smoothed change-point score full-cycle sequence under the dimensional anomalous mode. For the first The threshold for triggering change point scores under abnormal modalities.

[0244] In any anomalous mode, if If so, then the time point t is marked as a potential process adjustment node, and a set of candidate change points is obtained.

[0245] Step S106: Stage segmentation output. Using the output process adjustment node as the boundary, and combining the comparison process log, the continuous production data is divided into independent process stages.

[0246] Based on the candidate variable point set output in step S105, perform the event association verification operation: compare the timestamps of the candidate nodes with those of process log events (such as "catalyst injection" and "reactor pressurization"), allowing a maximum time deviation of Δt = 10 seconds (due to sensor response delay and system pipeline lag effect).

[0247] Specifically, the verification mechanism used in some implementations of this embodiment is described in detail as follows: if the time deviation between the candidate change point and the log event is within Δt, it is marked as a valid process adjustment point; otherwise, it is considered a false change point and is discarded.

[0248] Furthermore, process segmentation output operations are carried out, that is, using effective adjustment points as boundaries, continuous production data is divided into independent process stages (such as "constant temperature reaction stage" and "pressure reduction separation stage") to form a structured process dataset. The data of each stage is used independently to support process package data analysis and process optimization.

[0249] The solutions and methods provided in this embodiment offer a comprehensive and effective approach to accurately segment and detect changes in the process stages of artificial coking coal preparation. They not only effectively consider the spatiotemporal dependencies between monitoring sensor data in the artificial coking coal preparation pilot system, but also better adapt to the frequent process adjustments, unknown fluctuations, and high-noise environment characteristics present in the artificial coking coal preparation pilot scenario by introducing multi-mode time-series anomaly calculation and a two-layer adaptive learning structure. This effectively solves the problems of high reliance on manual process analysis, delays in process stage segmentation and detection, and inaccurate positioning in existing processes.

[0250] The present invention also provides a memory that stores multiple instructions for implementing the method as described in Embodiment 1.

[0251] like Figure 6 As shown, the present invention also provides an electronic device, including a processor 301 and a memory 302 connected to the processor 301. The memory 302 stores a plurality of instructions, which can be loaded and executed by the processor to enable the processor to perform methods as described in Embodiments 2 and 3.

[0252] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cross-dimensional variable point detection method for the analysis of artificial coking coal preparation processes, characterized in that, include: S1, acquire historical operating data of the artificial coking coal preparation pilot system and obtain an artificial coking coal production dataset based on the historical operating data; S2, perform a first-stage multi-modal temporal anomaly score learning on the artificial coking coal production dataset to obtain multi-modal temporal anomaly score information, and perform anomaly score smoothing processing on the multi-modal temporal anomaly score information. S3, perform a second stage of variable point score learning on the multi-mode time series abnormal score data after the abnormal score smoothing process to obtain multi-dimensional time series variable point score data, perform abnormal score smoothing process on the multi-dimensional time series variable point score data, and obtain a candidate variable point set based on dynamic threshold analysis and output the corresponding process adjustment node. S4, perform stage segmentation output, including: using the process adjustment node as the boundary, and combining the comparison of process logs and business event tags, accurately segment the artificial coking coal production dataset into multiple independent process stages, and use the data of each process stage independently to support process package data analysis and process optimization.

2. The cross-dimensional variable point detection method for the analysis of artificial coking coal preparation process according to claim 1, characterized in that, S1 includes: S11, Obtain historical operating data of the artificial coking coal preparation pilot system, including: obtaining historical operating data of the artificial coking coal preparation pilot system within a preset time period; the historical operating data is one or more sets; the historical operating data includes coal slurry flow rate, hydrogen pressure, hydrogen flow rate, catalyst feeding, coke pusher displacement sensor signal, reactor temperature, reactor vibration signal, coal preheating temperature, methane / CO concentration, gaseous product temperature, reactor pressure difference, and valve opening; the historical operating data of the artificial coking coal preparation system is used to provide underlying data support for the artificial coking coal preparation process analysis; S12, perform a first processing on the historical operation data, and obtain a structured sensor data information matrix based on the historical operation data after the first processing; the first processing includes aligning the acquisition intervals of different sensor signals in the historical operation data through signal sampling; S13, the structured sensor data information matrix is ​​subjected to a second processing, and structured historical operation data is obtained based on the structured sensor data information matrix after the second processing; wherein, the second processing includes normalizing the sensor data information matrix; the structured historical operation data forms the artificial coking coal production dataset.

3. The cross-dimensional variable point detection method for the analysis of artificial coking coal preparation process according to claim 2, characterized in that, The historical operating data of the artificial coking coal production pilot system obtained over the past preset time period are represented by the following formula (1): （1-1）； ， （1-2）； ， （1-3）； In equations (1-1) to (1-3), Represents the time measurement point, Represents the sensor serial number. This represents a matrix of historical operating data for an artificial coking coal preparation system. This represents the number of historical operational data collection groups. The number of signal sensors deployed in the artificial coking coal preparation system; Representing a point in time The sensor signal vector collected by the monitoring system of the artificial coking coal preparation system; Representing the Historical operating data of each sensor; This represents a point in time. No. The measured value of sensor number 1.

4. The cross-dimensional variable point detection method for the analysis of artificial coking coal preparation process according to claim 3, characterized in that, After obtaining multiple sets of historical operating data of the artificial coking coal preparation system within a preset time period, the Z-score normalization method is used to normalize each sensor signal in the historical operating data. The Z-score normalization operation is defined by the following formula (2): （2）； in, Represents the time measurement point, Represents the sensor serial number. Represents a point in time No. The original measurement value of sensor number 1 This represents the normalized result that eliminates the dimensional characteristics of the sensor. and They represent the first The average value and standard deviation of the historical operating data of sensor number 1.

5. The cross-dimensional variable point detection method for analyzing the artificial coking coal preparation process according to claim 4, characterized in that, S2 includes: S21, apply various types of time-series anomaly detection algorithms to the artificial coking coal production dataset to obtain multi-mode time-series anomaly score information of historical operating data. The multi-mode time-series anomaly score information is used to train a time-series anomaly model based on joint detection of multi-source information. S22, introduce data smoothing processing into the multi-mode time series anomaly score information to filter transient interference and highlight continuous change characteristics, and obtain smoothed multi-mode time series anomaly score data.

6. The cross-dimensional variable point detection method for the analysis of artificial coking coal preparation process according to claim 5, characterized in that, S21 includes: S211, the linear deviation anomaly is calculated based on the linear deviation anomaly measurement formula (3) of the historical operation data of the artificial coking coal preparation system. The formula (3) is as follows: ； ； ； （3）； In the formula, Representing a point in time The sensor signal vector collected by the monitoring system of the artificial coking coal preparation system. Represents the VAR model for time points The predicted value of the sensor signal vector. The lag order of the VAR model. This represents the number of historical operational data collection groups. The number of signal sensors deployed in the artificial coking coal preparation system. For the initial sparse matrix, for A matrix of coefficients to be estimated in dimension . for 3D white noise vector; S212, the Euclidean distance between the VAR model prediction result and the actual signal vector is used as the linear deviation anomaly score at the current time point, and its definition is shown in the following equation (4): （4）； in, Represents the current time point The VAR linear deviation anomaly score is at the current time point. Actual data Compared with VAR model-based predictions Vector distance between them; S213, the nonlinear deviation anomaly score is calculated based on the measurement formula (4) of the degree of nonlinear deviation anomaly of the historical operation data of the artificial coking coal preparation system. The formula (4) is as follows: ； ； （4）； in, Represents the number of isolated trees in an isolated forest. Representative data points In the Path length in an isolated tree Representative data points Average path length in an isolated forest; Representing the The total number of leaf nodes in the isolation tree. It is the negative logarithm of the total number of leaf nodes in all isolated trees in the forest. Representative data points Nonlinear deviation anomaly score; S214, the complexity deviation score is calculated using formula (5) to measure the degree of complex anomaly deviation based on the long-term dependence fluctuation and multivariate cointegration relationship changes of historical operating data of artificial coking coal preparation system. Formula (5) is as follows: （5）； in, For encoder model, For decoder models, Representing a point in time The sensor signal vector collected by the monitoring system of the artificial coking coal preparation system. This represents the LSTM-AE model for time points. The reconstructed value of the sensor signal vector, Represents the current time point The complexity deviates from the abnormal score; S215, Calculate the multi-mode time-series anomaly score matrix of historical operating data of artificial coking coal preparation system based on formula (6). The formula (6) is as follows: ； ； （6）； in, This indicates the number of timing anomaly detection algorithms used in the application. This represents the number of historical operational data collection groups. Represents the specific A series of time anomaly detection algorithms, Representing a point in time The corresponding data is in the first Anomaly scores under various time-series anomaly detection algorithms.

7. The cross-dimensional variable point detection method for the analysis of artificial coking coal preparation process according to claim 6, characterized in that, S22 includes: The smoothed multi-mode time-series anomaly score of the historical operating data of the artificial coking coal preparation system is calculated based on formula (7), which is as follows: （7）； in, Indicates at a point in time Application No. Anomaly scores of various time-series anomaly detection algorithms This represents the multi-modal temporal anomaly score after smoothing by the corresponding temporal anomaly algorithm. Represents the moving average window size. These are the weighting coefficients.

8. The cross-dimensional variable point detection method for the analysis of artificial coking coal preparation process according to claim 7, characterized in that, S3 includes: S31, apply a sequential iterative autoregressive model to each dimension of the smoothed multi-mode time-series anomaly score data, and calculate the change point score of each dimension through residual analysis, including: calculating the change point score of the historical operation data of the artificial coking coal preparation system based on formula (8), wherein formula (8) is defined as follows: ； ； ； ； ； ； ； （8）； in, Represents a point in time Application No. The smoothed temporal anomaly scores of various temporal anomaly detection algorithms are used hereafter for formula simplification; the symbols will be used in the following text. Substitute , The coefficients to be estimated in the AR model. The lag order of the AR model. This is the discount rate coefficient, used to control the weighting of historical data. It is white noise. Represents the SIAR model for time points Predicted values ​​of time-series anomaly scores. The order is represented as Historical vectors, Update the weights for the AR model coefficients. Let covariance matrix be the variance matrix. Representing time In the Variable point scores under abnormal modes; S32, performing anomaly score smoothing processing on the multi-dimensional time-series variable point score data, including: performing cascade optimization on the multi-dimensional time-series variable point score data through secondary moving average processing to generate the final smoothed multi-dimensional time-series variable point score data; obtaining a candidate variable point set based on dynamic threshold analysis and outputting the corresponding process adjustment nodes, including: S321, Calculate the smoothed variable point score of the historical operating data of the artificial coking coal preparation system based on formula (9), where formula (9) is defined as follows: （9）； in, Representative Time In the Smoothed variable point score under dimensional anomalous mode. Represents the moving average window size. The weighting coefficients are distributed according to a Gaussian distribution, with a weight of 0.5 at the center and decreasing to 0.1 at the edges; S322, based on dynamic threshold analysis, potential process adjustment nodes are extracted as variable points, and the dynamic threshold calculation is defined as shown in the following equation (10): （10）； in, Representing the The mean of the smoothed change-point score full-cycle sequence under the dimensional anomalous mode. Representing the Standard deviation of the smoothed change-point score full-cycle sequence under the dimensional anomalous mode. For the first The threshold for triggering change point scores under any anomalous mode; under any anomalous mode, if Then mark that time point. By identifying potential process adjustment nodes, a set of candidate change points is obtained.

9. The cross-dimensional variable point detection method for the analysis of artificial coking coal preparation process according to claim 8, characterized in that, S4 includes: S41, based on the candidate variable point set, perform an event association verification operation, including: comparing the timestamps of the candidate nodes with those of the process log events, allowing a maximum time deviation Δt = 10 seconds; if the time deviation between the candidate variable point and the log event is within Δt, it is marked as a valid process adjustment point; otherwise, it is considered a false variable point and is removed. S42, Perform process segmentation output operation, including: using the effective process adjustment point as the boundary, segment the continuous production data into independent process stages to form a structured process dataset, with each stage data being used independently to support process package data analysis and process optimization.

10. A cross-dimensional variable point detection system for analyzing the artificial coking coal preparation process, used to implement the method according to any one of claims 1-9, characterized in that, include: The production dataset acquisition module (101) is used to acquire historical operating data of the artificial coking coal preparation pilot system and obtain an artificial coking coal production dataset based on the historical operating data. The multi-modal temporal anomaly score information module (102) is used to perform a first-stage multi-modal temporal anomaly score learning on the artificial coking coal production dataset to obtain multi-modal temporal anomaly score information, and to perform anomaly score smoothing processing on the multi-modal temporal anomaly score information. The variable point score information module (103) is used to perform a second stage of variable point score learning on the multi-mode time series abnormal score data after the abnormal score smoothing process to obtain multi-dimensional time series variable point score data, perform abnormal score smoothing process on the multi-dimensional time series variable point score data, and obtain a candidate variable point set based on dynamic threshold analysis and output the corresponding process adjustment node. The stage segmentation output module (104) is used to perform stage segmentation output, including: taking the process adjustment node as the boundary, combining the process log and business event tags, accurately segmenting the artificial coking coal production dataset into multiple independent process stages, and the data of each process stage is used independently to support process package data analysis and process optimization.

Images