Method and system for monitoring energy consumption anomalies in steel production process based on data space
By constructing a data space-based method and utilizing mutual information algorithms and autoencoders, the problems of multi-source data integration and spatiotemporal feature extraction in energy consumption anomaly monitoring of the steel production process were solved, achieving high-precision and robust energy consumption anomaly monitoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- UNIV OF SCI & TECH BEIJING
- Filing Date
- 2026-02-25
- Publication Date
- 2026-06-02
Smart Images
Figure CN122132697A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of industrial process monitoring technology, and specifically refers to a method and system for monitoring abnormal energy consumption in steel production processes based on data space. Background Technology
[0002] The steel industry is a pillar industry of the national economy. Its production process encompasses multiple key stages, including blast furnace ironmaking, converter steelmaking, hot rolling, and cold rolling. Its total energy consumption accounts for a significant proportion of total industrial energy consumption, making it a typical high-energy-consuming industry. The energy system in the steel production process, as the core link connecting various production stages, is responsible for coordinating the generation, distribution, transmission, and utilization of multiple energy sources such as gas, steam, and electricity, directly impacting production efficiency, product costs, and environmental compliance. With the steel industry's transformation towards intelligent and green operations, higher demands are placed on the safe, stable, efficient, and economical operation of the energy system. Accurate monitoring of abnormal operating conditions and abnormal energy consumption fluctuations during the energy system's operation has become crucial for ensuring continuous steel production and reducing energy waste.
[0003] The energy system in the steel production process is significantly complex: on the one hand, it involves multi-level coordination at the equipment level, process level, and overall process level. The energy consumption of coke ovens, blast furnaces, and converters, the power distribution of rolling mills, and the recovery and utilization of waste heat are closely coupled. Energy flow, material flow, and information flow interact with each other, and any abnormality in any link or equipment may cause an imbalance in the energy supply and demand of the entire process through hierarchical transmission. On the other hand, the energy system data exhibits multi-source heterogeneous characteristics, including structured data such as historical operating data and real-time monitoring data, as well as unstructured knowledge such as process specifications, equipment correlation logic, and energy planning and scheduling. Moreover, the data exhibits complex spatiotemporal dynamic evolution characteristics with factors such as steel grade switching, operating condition adjustment, and equipment maintenance.
[0004] Existing methods for monitoring energy consumption anomalies in steel production processes have several limitations: First, traditional monitoring methods are mostly based on single-dimensional data or simple statistical models, failing to effectively integrate multi-source data and domain expert knowledge, making it difficult to characterize the spatiotemporal correlation characteristics of the energy system operation process and the intrinsic relationship between energy consumption indicators and various variables; Second, existing feature extraction methods often fail to effectively separate temporal and spatial features, leading to feature redundancy or missing key information, affecting the accuracy of energy consumption anomaly identification; Third, facing the complex dependencies of multi-source heterogeneous data, there is a lack of effective data organization and fusion frameworks, making it difficult to quantify the correlation strength between variables and their impact weight on energy consumption, resulting in high false alarm rates and delayed anomaly responses; Fourth, some methods do not consider the dynamic changes in steel production processes, resulting in insufficient model adaptability and difficulty in coping with changes in monitoring needs caused by operating condition switching, equipment aging, etc.
[0005] Furthermore, existing technologies lack depth in data and knowledge integration, failing to fully uncover the hidden energy consumption correlations within the data, and failing to effectively integrate domain expert knowledge into feature extraction and decision-making processes, resulting in insufficient reliability and interpretability of monitoring results. Therefore, how to construct a unified framework that balances data organization and knowledge integration to achieve accurate separation and extraction of spatiotemporal features, and improve the accuracy, robustness, and adaptability of energy consumption anomaly monitoring in steel production processes, has become an urgent technical problem to be solved. Summary of the Invention
[0006] To address the technical problems existing in the prior art, the present invention provides a method and system for monitoring energy consumption anomalies in steel production processes based on data space, the technical solution of which is as follows: On the one hand, a method for monitoring energy consumption anomalies in steel production processes based on data space is provided, the method comprising: S1. Acquire and preprocess real-time energy monitoring data and domain expert knowledge from the steel production process; S2. Calculate the spatial correlation strength between variables in the real-time energy monitoring data using the mutual information algorithm, clarify the variable dependencies, and combine domain expert knowledge to quantify the weight score of each variable's impact on energy consumption indicators, thereby achieving deep integration of multi-source heterogeneous data and domain knowledge. S3. Based on the spatial correlation strength and influence weight score, a three-dimensional multi-layer data space integrating time dimension, spatial dimension and energy consumption index dimension is constructed. The time dimension carries the temporal evolution characteristics of the data, the spatial dimension carries the correlation of different equipment and links, and the energy consumption index dimension carries the core evaluation indicators of energy consumption. Furthermore, an independent data space is constructed for each energy consumption index, and these independent data spaces for different indicators are then stacked and integrated to form a multi-layer data space including multiple index dimensions, thereby simultaneously carrying the spatiotemporal correlation information corresponding to multiple energy consumption indicators. S4. Extract time-series data from the time dimension of the data space and input it into the trained multi-channel stacked autoencoder SAISAE that combines the self-attention mechanism with the influence weight score. The encoder of SAISAE fuses the self-attention mechanism with the influence weight score and embeds the influence weight score into the attention weight calculation to guide the model to focus on the time-series features that are key to the performance indicators, thereby obtaining energy consumption weighted time features that carry energy consumption correlation information. The decoder of SAISAE reconstructs the energy consumption weighted time features to generate a first reconstructed variable that is sensitive to the energy consumption indicators of interest. S5. Each variable in the data space is individually input into a single-channel stacked autoencoder MISAE that combines a mutual information adjacency matrix. The encoder of the MISAE independently extracts the spatial features of each variable and combines them with the mutual information adjacency matrix constructed from the spatial correlation strength between the variables to obtain spatial correlation features that characterize the collaborative correlation characteristics of each variable. The decoder of the MISAE reconstructs the spatial correlation features to generate a second reconstructed variable that is sensitive to spatial correlation information. S6. Based on the first and second reconstructed variables, construct corresponding SPE monitoring statistics to characterize the operating status of the energy production, consumption, transfer and storage process under the two feature dimensions. Then, perform decision-level fusion of the two monitoring statistics through a Bayesian fusion algorithm, and comprehensively utilize the state information of the spatiotemporal dual dimensions to output the final monitoring result of energy consumption anomalies in the steel production process.
[0007] On the other hand, a data space-based energy consumption anomaly monitoring system for steel production processes is provided, the system comprising: The acquisition and preprocessing module is used to acquire and preprocess real-time energy monitoring data and domain expert knowledge in the steel production process. The computational quantization module is used to calculate the spatial correlation strength between variables in the real-time energy monitoring data through the mutual information algorithm, clarify the variable dependencies, and combine domain expert knowledge to quantify the weight score of the impact of each variable on the energy consumption index, so as to achieve deep integration of multi-source heterogeneous data and domain knowledge. The construction module is used to construct a three-dimensional multi-layer data space that integrates time dimension, spatial dimension and energy consumption index dimension based on the spatial correlation strength and influence weight score. The time dimension carries the temporal evolution characteristics of the data, the spatial dimension carries the correlation of different devices and links, and the energy consumption index dimension carries the core evaluation indicators of energy consumption. Furthermore, an independent data space is constructed for each energy consumption index, and these independent data spaces for different indicators are then stacked and integrated to form a multi-layer data space including multiple index dimensions, thereby simultaneously carrying the spatiotemporal correlation information corresponding to multiple energy consumption indicators. A multi-channel stacked autoencoder (SAISAE) is used to extract temporal data from the temporal dimension of the data space. The SAISAE is trained and combined with a self-attention mechanism and the influence weight score. The encoder of the SAISAE integrates the self-attention mechanism and the influence weight score, embeds the influence weight score into the attention weight calculation, guides the model to focus on the temporal features that are key to the performance indicators, and obtains energy consumption weighted time features carrying energy consumption correlation information. The decoder of the SAISAE reconstructs the energy consumption weighted time features and generates a first reconstructed variable that is sensitive to the energy consumption indicators of interest. A single-channel stacked autoencoder (MISAE) is used to input each variable in the data space individually into a single-channel stacked autoencoder MISAE that incorporates a mutual information adjacency matrix. The encoder of the MISAE independently extracts the spatial features of each variable and combines them with the mutual information adjacency matrix constructed from the spatial correlation strength between the variables to obtain spatial correlation features that characterize the collaborative correlation characteristics of each variable. The decoder of the MISAE reconstructs the spatial correlation features to generate a second reconstructed variable that is sensitive to spatial correlation information. A fusion output module is constructed to build corresponding SPE monitoring statistics based on the first and second reconstructed variables, respectively, to characterize the operating status of the energy production, consumption, transfer and storage process under the two feature dimensions. The two types of monitoring statistics are then fused at the decision level using a Bayesian fusion algorithm. By comprehensively utilizing the state information in both the spatiotemporal dimensions, the final monitoring results of energy consumption anomalies in the steel production process are output.
[0008] On the other hand, an electronic device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, which is loaded and executed by the processor to implement the above-mentioned data space-based method for monitoring energy consumption anomalies in steel production processes.
[0009] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, and the at least one instruction is loaded and executed by a processor to implement the above-described method for monitoring energy consumption anomalies in steel production processes based on data space.
[0010] The beneficial effects of the technical solution provided by this invention include at least the following: 1) Based on the three-dimensional multi-layer data space architecture and knowledge fusion mechanism, the integrated integration of multi-source heterogeneous data of steel process production and energy with domain expert knowledge is realized, a unified data organization and association framework is built, data silos are broken, energy consumption related information is transmitted more efficiently, and accurate data support is provided for monitoring. 2) By designing a dual model of SAISAE and MISAE and using a mutually exclusive optimization joint training strategy, we can achieve effective separation and accurate extraction of spatiotemporal features, avoid the problems of feature redundancy and missing key information in traditional methods, and significantly improve the representation ability of energy consumption-related features. 3) By adopting a Bayesian decision-level fusion algorithm, the monitoring statistics of both spatiotemporal dimensions are integrated to reduce the risk of false alarms and false alarms caused by a single feature dimension, and to significantly improve the accuracy and robustness of identifying abnormal energy consumption conditions in the steel process. 4) Deeply integrate data correlation quantification with domain expert knowledge, not only to uncover the implicit energy consumption correlation patterns in the data, but also to optimize the monitoring logic through knowledge guidance, taking into account the reliability and interpretability of the monitoring results, and helping the steel process energy system to operate efficiently, economically, safely and stably. Attached Figure Description
[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0012] Figure 1 This is a flowchart of a method for monitoring abnormal energy consumption in a steel production process based on data space, provided by an embodiment of the present invention. Figure 2 This is a schematic diagram of a three-dimensional multi-layer data space provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the SAISAE and MISAE processing procedures provided in the embodiments of the present invention; Figure 4 This is a schematic diagram illustrating the collaborative training of SAISAE and MISAE using a joint feature extraction training strategy with mutually exclusive optimization objectives, as provided in this embodiment of the invention. Figure 5 This is a block diagram of an energy consumption anomaly monitoring system for steel production process based on data space, provided by an embodiment of the present invention. Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation
[0013] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.
[0014] This invention provides a method for monitoring energy consumption anomalies in steel production processes based on data space. This method can be implemented by an electronic device, which can be a terminal or a server. Figure 1 The diagram shown is a flowchart of the method. The processing flow may include the following steps: S1. Acquire and preprocess real-time energy monitoring data and domain expert knowledge from the steel production process; The real-time energy monitoring data in this embodiment of the invention includes equipment power, energy transmission efficiency, and environmental parameter data collected in real time through field terminals. Each variable in these data refers to an observable physical quantity or operating parameter collected by different sensors, instruments, or systems in the steel production-energy system. The collection frequency is set to 1 time / minute to ensure the real-time and continuous nature of the data. Domain expert knowledge includes the logic of equipment association and the conversion and coupling relationship of multi-media energy. It also covers the tacit knowledge accumulated by experts, such as the experience in judging energy consumption anomalies and the transmission law of equipment association faults, which provides guidance for data space construction, association weight quantification and feature extraction.
[0015] The preprocessing includes normalizing the data.
[0016] As an optional embodiment, historical data and real-time monitoring data can be normalized based on the corresponding statistical characteristics of historical data. The normalization process in S1 can specifically include the following steps: As an optional embodiment, the mean of each variable in the historical data is first calculated, and then the mean is subtracted from the data of each variable to obtain the change of each sample relative to the overall average.
[0017] As an alternative implementation, the standard deviation of each variable in the historical data is first calculated, and then the centered data is divided by the standard deviation. This avoids certain special variables from being dominant, thus obtaining the normalized historical data.
[0018] As an optional implementation, the real-time monitoring data is normalized using the mean and standard deviation of historical data.
[0019] S2. Calculate the spatial correlation strength between variables in the real-time energy monitoring data using the mutual information algorithm, clarify the variable dependencies, and combine domain expert knowledge to quantify the weight score of each variable's impact on energy consumption indicators, thereby achieving deep integration of multi-source heterogeneous data and domain knowledge. Optionally, in step S2, the spatial correlation strength between variables in the real-time energy monitoring data is calculated using a mutual information algorithm, specifically including: S21. Represent the real-time energy monitoring data as a time series. , subscript Indicates the first The nth time series, that is, the nth time series One variable, This represents the time series, i.e., the number of variables. Given the length of the time series, calculate the mutual information values between all variables as follows:
[0020] in, and There are two variables. yes right mutual information value, For joint probability distribution, , It represents a marginal probability distribution; S22. Normalize each mutual information value as follows:
[0021] in, It is the minimum of all mutual information values. It is the maximum value of all mutual information values. It is the normalized mutual information value; S23. Sum all the mutual information values of each variable corresponding to the other variables to obtain the spatial correlation strength of that variable, calculated as follows:
[0022] in, It is a variable The sum of mutual information values represents the implicit spatial correlation of the variable. The larger the value, the more variable it represents. The stronger the spatial correlation with other variables.
[0023] Optionally, in step S2, domain expert knowledge is incorporated to quantify the weighted score of the impact of each variable on the energy consumption index, specifically including: S24. Combining the knowledge of domain experts, select the core variables that directly drive the change of a certain energy consumption indicator, and denote them as fundamental variables. This serves as the benchmark variable for weight calculation; S25. Use the mutual information algorithm to quantify the degree of nonlinear correlation between the remaining variables and the fundamental variable, assuming the set of fundamental variables is... The remaining set of variables to be calculated is Then the variable and The mutual information value is calculated as follows: Iterate through all variables to be calculated to obtain a set of mutual information values for each variable; S26. Normalize the sum of mutual information values of each variable to obtain the influence weight score of each variable on the energy consumption index (the score range of all variables is uniformly mapped to the [0,1] interval). The larger the score, the higher the influence of the variable on the energy consumption index.
[0024] S3. Based on the spatial correlation strength and influence weight score, construct a three-dimensional multi-layer data space that integrates time dimension, spatial dimension, and energy consumption index dimension, such as... Figure 2As shown, the time dimension carries the temporal evolution characteristics of the data, the spatial dimension carries the correlation between different devices and links, and the energy consumption index dimension carries the core evaluation indicators of energy consumption. Furthermore, an independent data space is constructed for each energy consumption index, and these independent data spaces for different indicators are then layered and integrated to form a multi-layer data space that includes multiple index dimensions, thereby simultaneously carrying the spatiotemporal correlation information corresponding to multiple energy consumption indicators. Among them, the core evaluation indicators of the energy consumption index dimension include one or more of the following: unit product energy consumption, energy utilization rate, energy consumption fluctuation coefficient, and key equipment energy consumption threshold. Unit product energy consumption quantifies the total amount of various types of energy required to produce a unit of steel product (ton of iron or ton of steel), and is a core indicator for measuring the economic efficiency of energy utilization. Energy utilization rate characterizes the efficiency of energy utilization from input to effective production capacity, reflecting the degree of loss during energy transmission and conversion. Energy consumption fluctuation coefficient calculates the dispersion of energy consumption data within a certain time window, accurately capturing abnormal fluctuations in energy consumption. Key equipment energy consumption threshold is set based on historical operating data and expert experience, and is an important reference for determining whether key equipment is in an abnormal energy consumption condition.
[0025] This space implicitly structures data and information, binding each variable to its corresponding spatial correlation strength and energy consumption impact weight score, forming a unified data organization of "variable-correlation strength-weight score", providing a structured data foundation for subsequent feature extraction.
[0026] S4. Extract time-series data from the time dimension of the data space and input it into the trained multi-channel stacked autoencoder SAISAE that combines the self-attention mechanism with the influence weight score. The encoder of SAISAE fuses the self-attention mechanism with the influence weight score and embeds the influence weight score into the attention weight calculation to guide the model to focus on the time-series features that are key to the performance indicators, thereby obtaining energy consumption weighted time features that carry energy consumption correlation information. The decoder of SAISAE reconstructs the energy consumption weighted time features to generate a first reconstructed variable that is sensitive to the energy consumption indicators of interest. Optionally, such as Figure 3 As shown, the SAISAE processing procedure specifically includes: S41. The SAISAE encoder extracts global temporal features of multivariate time series. :
[0027] in, For SAISAE encoders; S42. Input the weighted score of the impact of each variable on the energy consumption index into the trained SAISAE, and calculate as follows:
[0028] in, Characteristics representing the rating, The influence weight score of the variable on the energy consumption index is given. The influence weight score is for each variable and cannot be directly integrated with the extracted time features. Therefore, the influence weight score is used as the input of the trained SAISAE to obtain the score features. S43. The extracted temporal features are processed using a self-attention mechanism, and the influence weight scores are fused together to calculate the following: in, For attention weights, For querying, by The retrieval intent, carrying the current features, is obtained through linear transformation. For key, also by The linear transformation yields the identification information carrying each feature. The dimension representing the time feature; S44. Then, the energy consumption weighted time feature is obtained through dot product, calculated as follows:
[0029] in, For energy consumption weighted time characteristics, For value, by The content information carrying each feature is obtained through linear transformation; S45. The SAISAE decoder reconstructs the energy consumption weighted time features to generate reconstructed variables that are sensitive to the energy consumption indicators of interest, calculated as follows: in, For the first reconstructed variable, This is the decoder for SAISAE.
[0030] Optionally, the loss function for SAISAE training is: in Representing variables exist The value at time, and , and This refers to the encoder and decoder functions of a stacked autoencoder in SAISAE.
[0031] This invention optimizes temporal reconstruction through a dedicated loss function, overcoming the shortcomings of traditional SAE in capturing long temporal dependencies.
[0032] S5. Each variable in the data space is individually input into a single-channel stacked autoencoder MISAE that combines a mutual information adjacency matrix. The encoder of the MISAE independently extracts the spatial features of each variable and combines them with the mutual information adjacency matrix constructed from the spatial correlation strength between the variables to obtain spatial correlation features that characterize the collaborative correlation characteristics of each variable. The decoder of the MISAE reconstructs the spatial correlation features to generate a second reconstructed variable that is sensitive to spatial correlation information. Optionally, such as Figure 3 As shown, the MISAE processing procedure specifically includes: S51, The encoder of the MISAE independently extracts variables. Spatial characteristics:
[0033] in, Represents spatial characteristics, Represents the corresponding variable in MISAE The encoder function of a single-channel stacked autoencoder; S52. Construct a mutual information adjacency matrix based on the spatial correlation strength among the variables. The calculation is as follows: S53. Then use the mutual information matrix Weighted fusion of features for each variable: in, This represents the spatial association feature; S54. Then, the MISAE decoder reconstructs the spatial correlation features to generate the second reconstructed variable: in, For the MISAE decoder, This is the reconstructed output data for MISAE.
[0034] Optionally, the loss function for MISAE training is a weighted combination of reconstruction loss and entropy loss, defined as follows:
[0035] The reconstruction loss is:
[0036] The entropy loss is:
[0037] Because the mutual information adjacency matrix still contains many redundant spatial connections, an entropy loss constraint is added to make it sparse, retaining only the important and key spatial connections. The greater the entropy loss, the sparser the matrix, retaining only the key spatial associations and improving the targeting of spatial feature extraction. This represents a hyperparameter used to balance the reconstruction task and the sparsification task.
[0038] This invention, through adding entropy loss constraints to the sparsification matrix, eliminates redundant spatial connections and accurately captures nonlinear spatial dependencies between variables.
[0039] Optionally, SAISAE and MISAE are trained collaboratively using a joint feature extraction training strategy with mutually exclusive optimization objectives. The total loss function for collaborative training is:
[0040] Among them, such as Figure 4 As shown, input data After passing through SAISAE and MISAE, the reconstructed output data is obtained. and Assuming that SAISAE and MISAE have successfully extracted the temporal and spatial features of the input data, then it is considered that... It does not contain spatial features. It does not contain time features, and then... Input SAISAE to obtain reconstructed output data. ,Will Input MISAE to obtain reconstructed output data ,because Lacking time characteristics Lacking spatial features, SAISAE and MISAE cannot rely on these incomplete features to accurately reconstruct the original input data. This will significantly increase the reconstruction error. Therefore, by encouraging this reconstruction error in the loss function, we can inversely constrain the two models to focus on the extraction of temporal and spatial features respectively, thereby avoiding feature redundancy and mutual interference, and achieving effective separation and purity improvement of the two types of features.
[0041] S6. Based on the first and second reconstructed variables, construct corresponding SPE monitoring statistics to characterize the operating status of the energy production, consumption, transfer and storage process under the two feature dimensions. Then, perform decision-level fusion of the two monitoring statistics through a Bayesian fusion algorithm, and comprehensively utilize the state information of the spatiotemporal dual dimensions to output the final monitoring result of energy consumption anomalies in the steel production process.
[0042] Optionally, S6 specifically includes: S61. Based on the first reconstructed variable and the second reconstructed variable, construct the corresponding SPE monitoring statistics respectively, wherein... Squared prediction error (SPE) is represented as follows: Among them, input data After passing through SAISAE and MISAE, the reconstructed output data is obtained. and ; This indicates the degree of deviation of energy consumption fluctuations from historical normal patterns over a time dimension. The larger the value, the more significant the deviation between the time-series energy consumption characteristics of the current energy system and the historical normal pattern, and the higher the possibility of anomalies. This indicates the degree of deviation between the energy consumption correlation of each device / link in the spatial dimension and the historical normal pattern. The larger the value, the more significant the deviation between the current energy consumption coordination relationship between devices and the historical normal pattern, and the higher the possibility of local energy consumption anomalies or device-related failures. S62. A decision-level fusion strategy is adopted to integrate monitoring information from both spatiotemporal dimensions and reduce the risk of misjudgment from a single dimension. This is achieved through Bayesian fusion of these two dimensions. The monitoring statistics are as follows: in, and express Posterior probability of energy consumption anomalies in monitoring statistics; S63, Control Limits Set ,in It is a preset confidence level. During monitoring, if Exceeding the control limit indicates an abnormal energy consumption.
[0043] like Figure 5 As shown in the figure, this embodiment of the invention also provides a data space-based energy consumption anomaly monitoring system for steel production processes, the system comprising: The acquisition and preprocessing module 510 is used to acquire and preprocess real-time energy monitoring data and domain expert knowledge in the steel production process. The computational quantization module 520 is used to calculate the spatial correlation strength between variables in the real-time energy monitoring data through the mutual information algorithm, clarify the variable dependency relationship, and combine domain expert knowledge to quantify the weight score of the impact of each variable on the energy consumption index, so as to realize the deep integration of multi-source heterogeneous data and domain knowledge. The construction module 530 is used to construct a three-dimensional multi-layer data space that integrates time dimension, spatial dimension and energy consumption index dimension based on the spatial correlation strength and influence weight score. The time dimension carries the temporal evolution characteristics of the data, the spatial dimension carries the correlation relationship of different equipment and links, and the energy consumption index dimension carries the core evaluation indicators of energy consumption. Furthermore, an independent data space is constructed for each energy consumption index, and these independent data spaces for different indicators are then stacked and integrated to form a multi-layer data space including multiple index dimensions, thereby simultaneously carrying the spatiotemporal correlation information corresponding to multiple energy consumption indicators. The multi-channel stacked autoencoder SAISAE540 is used to extract temporal data from the temporal dimension of the data space. It is input into the trained multi-channel stacked autoencoder SAISAE that combines a self-attention mechanism with the influence weight score. The encoder of SAISAE fuses the self-attention mechanism with the influence weight score and embeds the influence weight score into the attention weight calculation to guide the model to focus on the temporal features that are key to the performance indicators, thereby obtaining energy consumption weighted time features that carry energy consumption correlation information. The decoder of SAISAE reconstructs the energy consumption weighted time features to generate a first reconstructed variable that is sensitive to the energy consumption indicators of interest. The single-channel stacked autoencoder MISAE550 is used to input each variable in the data space individually into the single-channel stacked autoencoder MISAE combined with a mutual information adjacency matrix. The encoder of the MISAE independently extracts the spatial features of each variable and combines them with the mutual information adjacency matrix constructed from the spatial correlation strength between the variables to obtain spatial correlation features characterizing the cooperative correlation characteristics of each variable. The decoder of the MISAE reconstructs the spatial correlation features to generate a second reconstructed variable that is sensitive to spatial correlation information. A fusion output module 560 is constructed to construct corresponding SPE monitoring statistics based on the first and second reconstructed variables, respectively, to characterize the operating status of the energy production, consumption, transfer and storage process under the two feature dimensions. The two types of monitoring statistics are then fused at the decision level using a Bayesian fusion algorithm. By comprehensively utilizing the state information in both the spatiotemporal dimensions, the final monitoring results of energy consumption anomalies in the steel production process are output.
[0044] The energy consumption anomaly monitoring system for steel production process based on data space provided in this embodiment of the invention has a functional structure that corresponds to the energy consumption anomaly monitoring method for steel production process based on data space provided in this embodiment of the invention, and will not be described again here.
[0045] Figure 6 This is a schematic diagram of the structure of an electronic device 600 provided in an embodiment of the present invention. The electronic device 600 may vary considerably due to different configurations or performance. It may include one or more central processing units (CPUs) 601 and one or more memories 602. The memory 602 stores at least one instruction, which is loaded and executed by the processor 601 to implement the steps of the above-mentioned method for monitoring abnormal energy consumption in the steel production process based on data space.
[0046] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions that can be executed by a processor in a terminal to complete the aforementioned data space-based method for monitoring energy consumption anomalies in steel production processes. For example, the computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, or optical data storage device.
[0047] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.
[0048] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for monitoring energy consumption anomalies in steel production processes based on data space, characterized in that, The method includes: S1. Acquire and preprocess real-time energy monitoring data and domain expert knowledge from the steel production process; S2. Calculate the spatial correlation strength between variables in the real-time energy monitoring data using the mutual information algorithm, clarify the variable dependencies, and combine domain expert knowledge to quantify the weight score of each variable's impact on energy consumption indicators, thereby achieving deep integration of multi-source heterogeneous data and domain knowledge. S3. Based on the spatial correlation strength and influence weight score, a three-dimensional multi-layer data space integrating time dimension, spatial dimension and energy consumption index dimension is constructed. The time dimension carries the temporal evolution characteristics of the data, the spatial dimension carries the correlation of different equipment and links, and the energy consumption index dimension carries the core evaluation indicators of energy consumption. Furthermore, an independent data space is constructed for each energy consumption index, and these independent data spaces for different indicators are then stacked and integrated to form a multi-layer data space including multiple index dimensions, thereby simultaneously carrying the spatiotemporal correlation information corresponding to multiple energy consumption indicators. S4. Extract time-series data from the time dimension of the data space and input it into the trained multi-channel stacked autoencoder SAISAE that combines the self-attention mechanism with the influence weight score. The encoder of SAISAE fuses the self-attention mechanism with the influence weight score and embeds the influence weight score into the attention weight calculation to guide the model to focus on the time-series features that are key to the performance indicators, thereby obtaining energy consumption weighted time features that carry energy consumption correlation information. The decoder of SAISAE reconstructs the energy consumption weighted time features to generate a first reconstructed variable that is sensitive to the energy consumption indicators of interest. S5. Each variable in the data space is individually input into a single-channel stacked autoencoder MISAE that combines a mutual information adjacency matrix. The encoder of the MISAE independently extracts the spatial features of each variable and combines them with the mutual information adjacency matrix constructed from the spatial correlation strength between the variables to obtain spatial correlation features that characterize the collaborative correlation characteristics of each variable. The decoder of the MISAE reconstructs the spatial correlation features to generate a second reconstructed variable that is sensitive to spatial correlation information. S6. Based on the first and second reconstructed variables, construct corresponding SPE monitoring statistics to characterize the operating status of the energy production, consumption, transfer and storage process under the two feature dimensions. Then, perform decision-level fusion of the two monitoring statistics through a Bayesian fusion algorithm, and comprehensively utilize the state information of the spatiotemporal dual dimensions to output the final monitoring result of energy consumption anomalies in the steel production process.
2. The method according to claim 1, characterized in that, In step S2, the spatial correlation strength between variables in the real-time energy monitoring data is calculated using a mutual information algorithm, specifically including: S21. Represent the real-time energy monitoring data as a time series. , subscript Indicates the first The nth time series, that is, the nth time series One variable, This represents the time series, i.e., the number of variables. Given the length of the time series, calculate the mutual information values between all variables as follows: ； in, and There are two variables. yes right mutual information value, For joint probability distribution, , It represents a marginal probability distribution; S22. Normalize each mutual information value as follows: ； in, It is the minimum of all mutual information values. It is the maximum value of all mutual information values. It is the normalized mutual information value; S23. Sum all the mutual information values of each variable corresponding to the other variables to obtain the spatial correlation strength of that variable, calculated as follows: ； in, It is a variable The sum of mutual information values represents the implicit spatial correlation of the variable. The larger the value, the more variable it represents. The stronger the spatial correlation with other variables.
3. The method according to claim 2, characterized in that, In step S2, domain expert knowledge is incorporated to quantify the weighted score of each variable's impact on energy consumption indicators, specifically including: S24. Combining the knowledge of domain experts, select the core variables that directly drive the change of a certain energy consumption indicator, and denote them as fundamental variables. This serves as the benchmark variable for weight calculation; S25. Use the mutual information algorithm to quantify the degree of nonlinear correlation between the remaining variables and the fundamental variable, assuming the set of fundamental variables is... The remaining set of variables to be calculated is Then the variable and The mutual information value is calculated as follows: ； Iterate through all variables to be calculated to obtain a set of mutual information values for each variable; S26. Normalize the sum of mutual information values for each variable to obtain the influence weight score of each variable on the energy consumption index. The higher the score, the greater the influence of the variable on the energy consumption index.
4. The method according to claim 3, characterized in that, The SAISAE processing procedure specifically includes: S41. The SAISAE encoder extracts global temporal features of multivariate time series. : ； in, For SAISAE encoders; S42. Input the weighted score of the impact of each variable on the energy consumption index into the trained SAISAE, and calculate as follows: ； in, Characteristics representing the rating, The influence weight score of the variable on the energy consumption index is given. The influence weight score is for each variable and cannot be directly integrated with the extracted time features. Therefore, the influence weight score is used as the input of the trained SAISAE to obtain the score features. S43. The extracted temporal features are processed using a self-attention mechanism, and the influence weight scores are fused together to calculate the following: ； in, For attention weights, For querying, by The retrieval intent, carrying the current features, is obtained through linear transformation. For key, also by The linear transformation yields the identification information carrying each feature. The dimension representing the time feature; S44. Then, the energy consumption weighted time feature is obtained through dot product, calculated as follows: ； in, For energy consumption weighted time characteristics, For value, by The content information carrying each feature is obtained through linear transformation; S45. The SAISAE decoder reconstructs the energy consumption weighted time features to generate reconstructed variables that are sensitive to the energy consumption indicators of interest, calculated as follows: ； in, For the first reconstructed variable, This is the decoder for SAISAE.
5. The method according to claim 4, characterized in that, The loss function for SAISAE training is: ； in Representing variables exist The value at time, and , and This refers to the encoder and decoder functions of a stacked autoencoder in SAISAE.
6. The method according to claim 5, characterized in that, The MISAE processing procedure specifically includes: S51, The encoder of the MISAE independently extracts variables. Spatial characteristics: ； in, Represents spatial characteristics, Represents the corresponding variable in MISAE The encoder function of a single-channel stacked autoencoder; S52. Construct a mutual information adjacency matrix based on the spatial correlation strength among the variables. The calculation is as follows: ； S53. Then use the mutual information matrix Weighted fusion of features for each variable: ； in, This represents the spatial association feature; S54. Then, the MISAE decoder reconstructs the spatial correlation features to generate the second reconstructed variable: ； in, For the MISAE decoder, This is the reconstructed output data for MISAE.
7. The method according to claim 6, characterized in that, The loss function for MISAE training is a weighted combination of reconstruction loss and entropy loss, defined as follows: ； The reconstruction loss is: ； The entropy loss is: ； Because the mutual information adjacency matrix still contains many redundant spatial connections, an entropy loss constraint is added to make it sparse, retaining only the important and key spatial connections. The greater the entropy loss, the sparser the matrix, retaining only the key spatial associations and improving the targeting of spatial feature extraction. This represents a hyperparameter used to balance the reconstruction task and the sparsification task.
8. The method according to claim 7, characterized in that, The SAISAE and MISAE are trained together using a joint feature extraction training strategy with mutually exclusive optimization objectives. The total loss function for the joint training is: ； Among them, input data After passing through SAISAE and MISAE, the reconstructed output data is obtained. and Assuming that SAISAE and MISAE have successfully extracted the temporal and spatial features of the input data, then it is considered that... It does not contain spatial features. It does not contain time features, and then... Input SAISAE to obtain reconstructed output data. ,Will Input MISAE to obtain reconstructed output data ,because Lacking time characteristics Lacking spatial features, SAISAE and MISAE cannot rely on these incomplete features to accurately reconstruct the original input data. This will significantly increase the reconstruction error. Therefore, by encouraging this reconstruction error in the loss function, we can inversely constrain the two models to focus on the extraction of temporal and spatial features respectively, thereby avoiding feature redundancy and mutual interference, and achieving effective separation and purity improvement of the two types of features.
9. The method according to claim 8, characterized in that, S6 specifically includes: S61. Based on the first reconstructed variable and the second reconstructed variable, construct the corresponding SPE monitoring statistics respectively, wherein... Squared Prediction Error (SPE): ； ； Among them, input data After passing through SAISAE and MISAE, the reconstructed output data is obtained. and ; This indicates the degree of deviation of energy consumption fluctuations from historical normal patterns over a time dimension. The larger the value, the more significant the deviation between the time-series energy consumption characteristics of the current energy system and the historical normal pattern, and the higher the possibility of anomalies. This indicates the degree of deviation between the energy consumption correlation of each device / link in the spatial dimension and the historical normal pattern. The larger the value, the more significant the deviation between the current energy consumption coordination relationship between devices and the historical normal pattern, and the higher the possibility of local energy consumption anomalies or device-related failures. S62. A decision-level fusion strategy is adopted to integrate monitoring information from both spatiotemporal dimensions and reduce the risk of misjudgment from a single dimension. This is achieved through Bayesian fusion of these two dimensions. The monitoring statistics are as follows: ； in, and express Posterior probability of energy consumption anomalies in monitoring statistics; S63, Control Limits Set ,in It is a preset confidence level. During monitoring, if Exceeding the control limit indicates an abnormal energy consumption.
10. A data space-based energy consumption anomaly monitoring system for steel production processes, characterized in that, The system includes: The acquisition and preprocessing module is used to acquire and preprocess real-time energy monitoring data and domain expert knowledge in the steel production process. The computational quantization module is used to calculate the spatial correlation strength between variables in the real-time energy monitoring data through the mutual information algorithm, clarify the variable dependencies, and combine domain expert knowledge to quantify the weight score of the impact of each variable on the energy consumption index, so as to achieve deep integration of multi-source heterogeneous data and domain knowledge. The construction module is used to construct a three-dimensional multi-layer data space that integrates time dimension, spatial dimension and energy consumption index dimension based on the spatial correlation strength and influence weight score. The time dimension carries the temporal evolution characteristics of the data, the spatial dimension carries the correlation of different devices and links, and the energy consumption index dimension carries the core evaluation indicators of energy consumption. Furthermore, an independent data space is constructed for each energy consumption index, and these independent data spaces for different indicators are then stacked and integrated to form a multi-layer data space including multiple index dimensions, thereby simultaneously carrying the spatiotemporal correlation information corresponding to multiple energy consumption indicators. A multi-channel stacked autoencoder (SAISAE) is used to extract temporal data from the temporal dimension of the data space. The SAISAE is trained and combined with a self-attention mechanism and the influence weight score. The encoder of the SAISAE integrates the self-attention mechanism and the influence weight score, embeds the influence weight score into the attention weight calculation, guides the model to focus on the temporal features that are key to the performance indicators, and obtains energy consumption weighted time features carrying energy consumption correlation information. The decoder of the SAISAE reconstructs the energy consumption weighted time features and generates a first reconstructed variable that is sensitive to the energy consumption indicators of interest. A single-channel stacked autoencoder (MISAE) is used to input each variable in the data space individually into a single-channel stacked autoencoder MISAE that incorporates a mutual information adjacency matrix. The encoder of the MISAE independently extracts the spatial features of each variable and combines them with the mutual information adjacency matrix constructed from the spatial correlation strength between the variables to obtain spatial correlation features that characterize the collaborative correlation characteristics of each variable. The decoder of the MISAE reconstructs the spatial correlation features to generate a second reconstructed variable that is sensitive to spatial correlation information. A fusion output module is constructed to build corresponding SPE monitoring statistics based on the first and second reconstructed variables, respectively, to characterize the operating status of the energy production, consumption, transfer and storage process under the two feature dimensions. The two types of monitoring statistics are then fused at the decision level using a Bayesian fusion algorithm. By comprehensively utilizing the state information in both the spatiotemporal dimensions, the final monitoring results of energy consumption anomalies in the steel production process are output.