Industrial process soft-sensing method, device, storage medium and computer equipment

Abstract

The present disclosure relates to the technical field of industrial measurement, and provides an industrial process soft measurement method, device, storage medium and computer equipment, the method comprising: acquiring a historical process data set of a coal-fired boiler, and training a dynamic measurement prediction model based on the historical process data set to obtain a trained dynamic measurement prediction model; acquiring real-time process data of the coal-fired boiler and a working condition weight of a previous moment corresponding to the real-time process data; and inputting the real-time process data and the working condition weight of the previous moment corresponding to the real-time process data into the trained dynamic measurement prediction model to obtain a real-time nitrogen oxide concentration estimation value. In this embodiment, the working condition weight of the previous moment is introduced as model input, so that the dynamic measurement prediction model can effectively capture the dynamic correlation and evolution law of the working condition state in time sequence, thereby effectively improving the accuracy and reliability of the real-time soft measurement estimation value of the nitrogen oxide concentration of the coal-fired boiler.

Description

Technical Field

[0001] This disclosure relates to the field of industrial measurement technology, and more specifically, to a method, apparatus, storage medium, and computer device for soft measurement of industrial processes. Background Technology

[0002] In industrial process control, accurate monitoring of key quality variables is crucial for ensuring product quality, improving process safety, and achieving efficient operation. However, due to limitations such as high cost, complex operating conditions, and potential safety risks associated with direct measurement technologies, important parameters such as component conversion rate and gasification efficiency are often difficult to obtain in real time. To address this challenge, soft sensing technology, as a model-based virtual sensing method, is widely used. It constructs mathematical models and utilizes easily measurable auxiliary variables (such as temperature, pressure, and raw material composition) to achieve online estimation and closed-loop control of key quality variables.

[0003] In actual industrial processes, due to fluctuations in raw materials and adjustments in load, process data often exhibits multi-condition characteristics, meaning that the statistical distribution of data changes significantly with different conditions. Related soft sensor models, which assume the data follows a single distribution, suffer a severe decline in predictive performance when faced with multiple conditions. To address this, researchers have employed global modeling methods such as Gaussian mixture variational autoencoders to represent different conditions using mixture distributions. However, these methods have some drawbacks. For example, when performing online estimation, they fail to consider and utilize the dynamic temporal correlation of condition weights, leading to insufficient adaptive capability in dynamically changing production processes and limited predictive stability and accuracy. Summary of the Invention

[0004] This disclosure provides at least one industrial process soft measurement method, apparatus, storage medium, and computer equipment. By introducing the operating condition weights of the previous moment as model input, the dynamic measurement prediction model can effectively capture the dynamic correlation and evolution of operating conditions over time, thereby effectively improving the accuracy and reliability of real-time soft measurement estimates.

[0005] This disclosure provides an industrial process soft sensing method, including:

[0006] A historical process dataset of a coal-fired boiler is obtained. The historical process dataset includes multiple historical process data arranged in chronological order and nitrogen oxide concentration data corresponding to each historical process data. The historical process data includes coal feed rate, primary air volume, secondary air volume, furnace temperature, and furnace pressure.

[0007] The dynamic measurement prediction model is trained based on the historical process dataset to obtain the trained dynamic measurement prediction model; wherein, the dynamic measurement prediction model includes an encoder, a working condition allocation module and a regressor;

[0008] Obtain real-time process data of the coal-fired boiler and the operating condition weight of the previous moment corresponding to the real-time process data; and input the real-time process data and the operating condition weight of the previous moment corresponding to the real-time process data into the trained dynamic measurement and prediction model.

[0009] The encoder determines the current moment process feature vector of the real-time process data; and, based on the current moment process feature vector, the operating condition weight of the previous moment and the operating condition allocation module, determines the current moment operating condition weight; and, based on the current moment process feature vector, the current moment operating condition weight and the regressor, determines the estimated real-time nitrogen oxide concentration.

[0010] In some possible embodiments, training the dynamic measurement prediction model based on the historical process dataset includes:

[0011] Based on the historical process dataset, various working conditions and corresponding working condition parameter information are determined to obtain a working condition parameter set;

[0012] A model training dataset is constructed based on the historical process dataset; wherein, the model training dataset includes multiple model training subsets, each of which corresponds to a time series segment, including input process data arranged in chronological order and nitrogen oxide concentration data corresponding to the input process data;

[0013] The dynamic measurement prediction model is trained based on the model training dataset and the operating condition parameter set.

[0014] In some possible embodiments, the operating condition parameter information includes a mean vector and a covariance matrix; the step of determining multiple operating conditions and corresponding operating condition parameter information based on the historical process dataset to obtain an operating condition parameter set includes:

[0015] Based on the clustering analysis results of the historical process dataset, multiple operating conditions are identified; wherein, the multiple operating conditions include high load conditions, stable operation conditions, and low load conditions;

[0016] For each working condition, data samples that conform to the range of working condition parameters are selected from the historical process dataset to obtain a subset of historical process data corresponding to the working condition;

[0017] For each working condition, based on a subset of historical process data for that working condition, the mean vector and covariance matrix of that working condition are calculated as the working condition parameter information corresponding to that working condition.

[0018] In some possible embodiments, the dynamic measurement prediction model further includes a decoder; training the dynamic measurement prediction model based on the model training dataset and the operating condition parameter set includes:

[0019] Initialize the initial load weights for the first time step;

[0020] Iterate through the input data in each model training subset in chronological order. For each time step:

[0021] The input process data of the current time step is input into the encoder to obtain the input process data feature vector of the current time step;

[0022] When the current time step is the first time step, the training working condition weights for the current time step are determined based on the input process data feature vector of the current time step, the initial working condition weights, and the working condition allocation module.

[0023] If the current time step is not the first time step, the training working condition weights for the current time step are determined based on the input process data feature vector of the current time step, the working condition weights of the previous time step, and the working condition allocation module.

[0024] The training condition weights of the current time step and the feature vector of the input process data of the current time step are input into the regressor to determine the predicted nitrogen oxide concentration estimate corresponding to the input process data of the current time step.

[0025] Based on the decoder and the feature vector of the input process data at the current time step, the reconstruction process data corresponding to the input process data at the current time step is determined;

[0026] The target loss of the dynamic measurement prediction model is determined based on a preset loss function, and the model parameters of the dynamic measurement prediction model are adjusted based on the target loss until the training result meets the preset requirements, thus obtaining the trained dynamic measurement prediction model.

[0027] In some possible embodiments, determining the target loss of the dynamic measurement prediction model based on a preset loss function includes:

[0028] For each working condition, a target Gaussian distribution is defined based on the mean vector and covariance matrix corresponding to the working condition; the constraint loss of the dynamic measurement prediction model is calculated based on the spatial distance between the feature vector of the input process data at each time step in each model training subset and the target Gaussian distribution corresponding to each working condition.

[0029] Based on the predicted nitrogen oxide concentration estimates and nitrogen oxide concentration data at each time step in the training subset of each model, the regression loss is calculated;

[0030] Based on the input process data and reconstruction process data at each time step in the training subset of each model, the reconstruction loss is calculated;

[0031] The target loss is determined based on the constraint loss, the regression loss, and the reconstruction loss.

[0032] In some possible embodiments, determining the operating condition weight at the current moment includes:

[0033] The operating condition weight at the current moment is stored as the operating condition weight of the previous moment for the next prediction moment.

[0034] In some possible embodiments, determining the real-time nitrogen oxide concentration estimate includes:

[0035] When the estimated real-time nitrogen oxide concentration does not meet the preset industrial standard, an adjustment suggestion for the operating parameters of the coal-fired boiler is determined and output based on the current operating condition weight and the estimated real-time nitrogen oxide concentration.

[0036] This disclosure provides an industrial process soft measurement device, comprising:

[0037] The data acquisition module is used to acquire historical process datasets of a coal-fired boiler. The historical process datasets include multiple historical process data arranged in chronological order and nitrogen oxide concentration data corresponding to each historical process data. The historical process data includes coal feed rate, primary air volume, secondary air volume, furnace temperature, and furnace pressure.

[0038] The model training module is used to train the dynamic measurement prediction model based on the historical process dataset to obtain the trained dynamic measurement prediction model; wherein, the dynamic measurement prediction model includes an encoder, a working condition allocation module, and a regressor;

[0039] The data input module is used to acquire real-time process data of the coal-fired boiler and the operating condition weight of the previous moment corresponding to the real-time process data; and input the real-time process data and the operating condition weight of the previous moment corresponding to the real-time process data into the trained dynamic measurement and prediction model.

[0040] The model execution module is configured to determine the current moment process feature vector of the real-time process data through the encoder; and to determine the current moment operating weight based on the current moment process feature vector, the operating weight of the previous moment and the operating weight allocation module; and to determine the real-time nitrogen oxide concentration estimate based on the current moment process feature vector, the current moment operating weight and the regressor.

[0041] In some possible embodiments, the model training module is specifically used for:

[0042] Based on the historical process dataset, various working conditions and corresponding working condition parameter information are determined to obtain a working condition parameter set;

[0043] A model training dataset is constructed based on the historical process dataset; wherein, the model training dataset includes multiple model training subsets, each of which corresponds to a time series segment, including input process data arranged in chronological order and nitrogen oxide concentration data corresponding to the input process data;

[0044] The dynamic measurement prediction model is trained based on the model training dataset and the operating condition parameter set.

[0045] In some possible embodiments, the operating condition parameter information includes a mean vector and a covariance matrix; the model training module is specifically used for:

[0046] Based on the clustering analysis results of the historical process dataset, multiple operating conditions are identified; wherein, the multiple operating conditions include high load conditions, stable operation conditions, and low load conditions;

[0047] For each working condition, data samples that conform to the range of working condition parameters are selected from the historical process dataset to obtain a subset of historical process data corresponding to the working condition;

[0048] For each working condition, based on a subset of historical process data for that working condition, the mean vector and covariance matrix of that working condition are calculated as the working condition parameter information corresponding to that working condition.

[0049] In some possible embodiments, the dynamic measurement prediction model further includes a decoder; the model training module is specifically used for:

[0050] Initialize the initial load weights for the first time step;

[0051] Iterate through the input data in each model training subset in chronological order. For each time step:

[0052] The input process data of the current time step is input into the encoder to obtain the input process data feature vector of the current time step;

[0053] When the current time step is the first time step, the training working condition weights for the current time step are determined based on the input process data feature vector of the current time step, the initial working condition weights, and the working condition allocation module.

[0054] If the current time step is not the first time step, the training working condition weights for the current time step are determined based on the input process data feature vector of the current time step, the working condition weights of the previous time step, and the working condition allocation module.

[0055] The training condition weights of the current time step and the feature vector of the input process data of the current time step are input into the regressor to determine the predicted nitrogen oxide concentration estimate corresponding to the input process data of the current time step.

[0056] Based on the decoder and the feature vector of the input process data at the current time step, the reconstruction process data corresponding to the input process data at the current time step is determined;

[0057] The target loss of the dynamic measurement prediction model is determined based on a preset loss function, and the model parameters of the dynamic measurement prediction model are adjusted based on the target loss until the training result meets the preset requirements, thus obtaining the trained dynamic measurement prediction model.

[0058] In some possible embodiments, the model training module is specifically used for:

[0059] For each working condition, a target Gaussian distribution is defined based on the mean vector and covariance matrix corresponding to the working condition; the constraint loss of the dynamic measurement prediction model is calculated based on the spatial distance between the feature vector of the input process data at each time step in each model training subset and the target Gaussian distribution corresponding to each working condition.

[0060] Based on the predicted nitrogen oxide concentration estimates and nitrogen oxide concentration data at each time step in the training subset of each model, the regression loss is calculated;

[0061] Based on the input process data and reconstruction process data at each time step in the training subset of each model, the reconstruction loss is calculated;

[0062] The target loss is determined based on the constraint loss, the regression loss, and the reconstruction loss.

[0063] In some possible embodiments, the model execution module is further configured to:

[0064] The operating condition weight at the current moment is stored as the operating condition weight of the previous moment for the next prediction moment.

[0065] In some possible embodiments, the model execution module is further configured to:

[0066] When the estimated real-time nitrogen oxide concentration does not meet the preset industrial standard, an adjustment suggestion for the operating parameters of the coal-fired boiler is determined and output based on the current operating condition weight and the estimated real-time nitrogen oxide concentration.

[0067] This disclosure provides a computer device including a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the industrial process soft measurement method as described in any of the above possible embodiments.

[0068] This disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the industrial process soft measurement method as described in any of the possible embodiments above.

[0069] The industrial process soft measurement method, apparatus, storage medium, and computer equipment provided in this disclosure construct a dynamic measurement prediction model including an encoder, a working condition allocation module, and a regressor. By incorporating the working condition weights from the previous moment and real-time process data as model inputs, the trained model can effectively perceive and learn the dynamic correlation and evolution patterns of working conditions over time. In this way, by combining historical data training with real-time data and integrating working condition weights, the impact of the dynamic changes in working conditions during industrial processes on the measurement results is fully considered. Addressing the multi-working-condition operation characteristics of coal-fired boilers, the method utilizes easily measurable process variables (coal feed rate, air volume, temperature, and pressure) to achieve a more accurate and stable estimation of the key parameter—nitrogen oxide concentration—which is difficult to monitor online.

[0070] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0071] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings referenced in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. Those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0072] Figure 1 A flowchart of an industrial process soft measurement method provided by an embodiment of this disclosure is shown;

[0073] Figure 2 A flowchart of a dynamic measurement prediction model training method provided by an embodiment of this disclosure is shown;

[0074] Figure 3 A flowchart of a dynamic measurement prediction model operation method provided by an embodiment of this disclosure is shown;

[0075] Figure 4 This diagram illustrates the structure of a dynamic measurement prediction model provided in an embodiment of the present disclosure.

[0076] Figure 5 A flowchart of a target loss calculation method provided by an embodiment of this disclosure is shown;

[0077] Figure 6 A schematic diagram of the structure of an industrial process soft measurement device provided in an embodiment of this disclosure is shown;

[0078] Figure 7 A schematic diagram of the structure of a computer device provided in an embodiment of this disclosure is shown. Detailed Implementation

[0079] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. The components of the embodiments of this disclosure described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely represents selected embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0080] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0081] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0082] In industrial processes, accurate monitoring of key quality variables is crucial for ensuring product quality, meeting user needs, and achieving safe process operation. However, due to high costs, complex operating conditions, and potential safety risks, the real-time direct measurement of key parameters such as component conversion rate and gasification efficiency faces significant technical bottlenecks. To overcome these limitations, soft sensing technology, as a virtual sensing solution, has been widely applied in the industrial field. It derives the dynamic trends of key quality variables by constructing mathematical models based on easily measurable auxiliary variables (such as temperature, pressure, and raw material composition), thereby achieving online estimation and closed-loop control of process parameters. Among current mainstream soft sensing methods, data-driven modeling techniques have become a research hotspot due to their advantages such as not relying on complex mechanistic models and adapting to nonlinear relationships. These methods use machine learning algorithms to uncover implicit correlations between process variables, with significant progress, particularly in neural network applications. Structures such as autoencoders (AEs) exhibit superior performance in industrial process scenarios due to their ability to extract complex features, noise immunity, and scalable architecture.

[0083] However, in actual industrial processes, due to changes in various operating conditions such as raw material fluctuations, catalyst activity variations, and load adjustments, production processes often exhibit multiple stable or transitional states, resulting in significant multi-condition characteristics in the collected process data. Specifically, data from different operating conditions show significant differences in statistical distribution (such as mean and variance). Traditional soft sensor models are typically trained based on the assumption that the data follows a single distribution. When faced with multi-condition data, the model struggles to learn the common characteristics of all conditions, leading to a severe "condition drift" problem. This means that the model's predictive performance drops sharply or even fails completely under certain conditions, greatly limiting its reliability in complex operating environments.

[0084] Research has revealed that current research on modeling multi-condition data primarily follows two technical approaches: local modeling strategies and global modeling strategies. Local modeling strategies employ a multi-model architecture, constructing dedicated models for different conditions to extract condition-specific features. While this method effectively characterizes the independent features of each condition, its inherent limitations restrict industrial applications: on the one hand, the multi-model architecture significantly increases system complexity and maintenance costs; on the other hand, during condition transitions, model switching lag can easily lead to prediction instability, making it difficult to meet the real-time requirements of dynamic production scenarios.

[0085] In contrast, global modeling strategies represent the distribution of multi-condition data uniformly through a single model. Their advantages are twofold: first, a unified feature extraction framework effectively reduces modeling complexity; second, cross-condition feature fusion can be achieved within a shared feature space, enhancing the model's ability to perceive the global data structure. However, in practical industrial applications, factors such as raw material fluctuations and adjustments to operating parameters can cause dynamic changes in the weight of different operating condition data on process objectives. Existing global modeling strategies do not consider such operating condition dependencies, leading to insufficient adaptability of the model in complex production scenarios.

[0086] Based on the above research, this disclosure provides an industrial process soft measurement method, apparatus, storage medium, and computer equipment. Specifically, firstly, a historical process dataset of a coal-fired boiler is acquired. The historical process dataset includes multiple historical process data arranged in chronological order and nitrogen oxide concentration data corresponding to each historical process data. The historical process data includes coal feed rate, primary air volume, secondary air volume, furnace temperature, and furnace pressure. Secondly, a dynamic measurement prediction model is trained based on the historical process dataset to obtain a trained dynamic measurement prediction model. The dynamic measurement prediction model includes an encoder, a working condition allocation module, and a regressor. Finally, real-time process data of the coal-fired boiler and the working condition weights corresponding to the real-time process data at the previous moment are acquired. The real-time process data and the working condition weights corresponding to the real-time process data at the previous moment are then input into the trained dynamic measurement prediction model. Here, the operation in this model is as follows: based on real-time process data, the process feature vector at the current moment is determined by the encoder; and based on the process feature vector at the current moment, the operating condition weight at the previous moment, and the operating condition allocation module, the operating condition weight at the current moment is determined; and based on the process feature vector at the current moment, the operating condition weight at the current moment, and the regressor, the estimated value of real-time nitrogen oxide concentration is determined.

[0087] In this embodiment, a dynamic measurement and prediction model is constructed, comprising an encoder, a working condition allocation module, and a regressor. The working condition weights from the previous moment and real-time process data are used as input to the model. This training allows the model to effectively perceive and learn the dynamic correlation and evolution of working conditions over time. By combining historical data training with real-time data and incorporating working condition weights, the impact of dynamic changes in working conditions during industrial processes on measurement results is fully considered. Addressing the multi-working-condition operation characteristics of coal-fired boilers, easily measurable process variables (coal feed rate, air volume, temperature, and pressure) are used to achieve a more accurate and stable estimation of the key parameter—nitrogen oxide concentration—which is difficult to monitor online.

[0088] To facilitate understanding of this embodiment, the executing entity of the industrial process soft measurement method provided in this disclosure will first be described in detail. The executing entity of the industrial process soft measurement method provided in this disclosure is a computer device. This computer device can be a terminal device or a server. The terminal device can also be a mobile device, user terminal, terminal, handheld device, computing device, vehicle-mounted device, wearable device, etc. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud storage, big data, and artificial intelligence platforms. Optionally, this method can also be applied to an implementation environment composed of computer devices and servers.

[0089] The industrial process soft measurement method provided in the embodiments of this application will be described in detail below with reference to the accompanying drawings. See also Figure 1 The diagram shows a flowchart of an industrial process soft measurement method provided in an embodiment of this disclosure. The method includes the following steps S101 to S104:

[0090] S101, Obtain the historical process dataset of coal-fired boilers.

[0091] Here, the historical process dataset refers to a collection of process data recorded from the industrial process control system of a coal-fired boiler, arranged in chronological order. It typically includes two types of variables: first, auxiliary variables that are easily measured in real time, including coal feed rate, primary air volume, secondary air volume, furnace temperature, and furnace pressure; second, key quality variables that are difficult to measure in real time but directly related to environmental compliance and operational evaluation, namely nitrogen oxide concentration data. This mainly refers to the instantaneous or periodic average value of nitrogen oxide concentration obtained through extractive or in-situ online monitoring instruments installed on the boiler outlet flue, or through laboratory testing of periodically collected flue gas samples. Due to limitations in analysis cycle or sampling frequency, the measured value of this variable lags significantly by minutes or even hours relative to the rapidly changing combustion process. Therefore, the historical process dataset includes multiple historical process data arranged in chronological order and the corresponding nitrogen oxide concentration data for each historical process data. Among them, coal quantity is the amount of fuel fed into the furnace per unit time; primary air volume refers to the oxygen required for transporting pulverized coal and supplying it for initial combustion, and is a key parameter for organizing the flow and stable combustion within the furnace; secondary air volume refers to the air supplied from the top of the burner, used to supplement the air required for combustion and enhance the airflow disturbance within the furnace to achieve complete fuel combustion and control the generation of nitrogen oxides; furnace temperature reflects the ignition conditions of the pulverized coal airflow and the overall thermal intensity within the furnace, and can affect combustion efficiency and pollutant generation rate; furnace pressure represents the ventilation status within the furnace, and its stability is a prerequisite for ensuring safe combustion and economical operation.

[0092] The industrial process soft measurement method proposed in this disclosure possesses excellent versatility and scalability, and its core architecture can be seamlessly adapted to various industrial process scenarios. In specific applications, the variables in the historical process dataset will be defined accordingly based on the actual industrial object. For example, in a petrochemical catalytic cracking unit, these auxiliary variables can be reactor temperature, reactor pressure, feed flow rate, catalyst circulation rate, etc.; the key quality variable can be the gasoline yield in the product distribution of the catalytic cracking unit; in a wastewater treatment bioreactor, auxiliary variables can be dissolved oxygen concentration, pH value, sludge concentration, influent flow rate, etc., and the key quality variable is the effluent chemical oxygen demand concentration; furthermore, in complex processes such as metallurgy and pharmaceuticals, process mechanism-derived variables such as thermal efficiency indicators and reaction accumulation can also be introduced. Therefore, the scope of protection of this method is not limited to coal-fired boilers or any specific application instance, but covers the implementation of the core method in different industrial scenarios.

[0093] In some possible embodiments, after obtaining historical data, data preprocessing can be performed, including handling missing values, removing outliers, and performing standardization or normalization operations to eliminate the impact of differences in the units of different variables on model training. For example, long-term missing variables caused by sensor failure can be directly removed from the input features; for partial null values ​​caused by momentary communication interruptions, the moving average of data from adjacent time points can be used to fill them. Outlier handling can combine the normal operating range set by process knowledge with data-driven statistical methods, such as first using a limiting method to remove samples that exceed the process specification range, and then combining 3 The criteria further remove outliers that significantly deviate from the center of the data distribution. Data standardization typically employs the z-score method, which involves subtracting the mean from the value of each variable and then dividing by the standard deviation, ensuring that the processed data conforms to a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0094] S102, the dynamic measurement prediction model is trained based on the historical process dataset to obtain the trained dynamic measurement prediction model.

[0095] Specifically, after obtaining historical data, a dynamic measurement prediction model can be trained based on the historical process dataset. This model can be used to estimate key quality variables (i.e., real-time nitrogen oxide concentration estimates) from time-series data with multiple operating conditions. Here, the dynamic measurement prediction model can include an encoder, an operating condition allocation module, a regressor, and a decoder. The encoder is responsible for compressing the high-dimensional input process data into low-dimensional feature vectors, thereby extracting key information from the data. The operating condition allocation module dynamically calculates the probability of belonging to each operating condition at the current moment, i.e., the operating condition weight, based on the data characteristics at the current moment and the operating condition weights at the previous moment. An operating condition is a relatively stable production state in an industrial process, defined by a series of process parameters, such as the high-load operating condition, stable operation condition, and low-load operating condition of a boiler. The regressor calculates the predicted value of the key quality variable, i.e., the real-time nitrogen oxide concentration estimate, based on the feature vector and the operating condition weights. The decoder attempts to reconstruct the original input process data from the feature vectors. This reconstruction process, as an auxiliary task, constrains the model to learn more representative features.

[0096] For example, refer to Figure 2 The training process of the dynamic measurement prediction model is described in detail below. When training the dynamic measurement prediction model based on historical process datasets, the following steps S201~S203 may be included:

[0097] S201, Based on the historical process dataset, determine multiple working conditions and the working condition parameter information corresponding to each working condition to obtain a working condition parameter set.

[0098] Here, the determination of the number of operating conditions can be based on the results of cluster analysis of historical data. For example, K-means or Gaussian mixture model clustering algorithms can be used to group data samples with similar statistical characteristics into one class, with each class corresponding to a potential operating condition, such as high-load operating condition, stable operation operating condition, and low-load operating condition. At the same time, knowledge of the process mechanism can be combined to attach a clear physical label, i.e., operating condition parameter information, to each operating condition.

[0099] Specifically, before determining the operating parameters for each operating condition, it is also necessary to obtain the parameter range corresponding to each operating condition. For example, for a coal-fired boiler, the parameter range corresponding to the high-load operating condition can be defined as boiler load ≥ 500MW, excess air coefficient 1.2-1.3, and furnace temperature 950-1050℃; the load corresponding to the stable operation operating condition is 300-500MW, excess air coefficient 1.3-1.4, and furnace temperature 850-950℃; the parameter range corresponding to the low-load combustion operating condition is boiler load ≤ 300MW, excess air coefficient 1.4-1.5, and furnace temperature 750-850℃.

[0100] Furthermore, after determining the parameter range for each working condition, all data samples falling entirely within the corresponding parameter range are selected from the historical dataset, forming a data subset for each working condition. Based on the data subset corresponding to each working condition, the mean vector and covariance matrix of each working condition can be calculated as the working condition parameter information for each working condition. Specifically, for any given working condition k, its working condition parameter information may include a mean vector. and a covariance matrix Mean vector This represents the typical central or average values ​​of all key process variables (such as coal feed rate, air volume, temperature, and pressure) under a stable operating condition, reflecting the baseline state of operation under that condition. For example, under high-load conditions, the mean values ​​of coal feed rate and furnace temperature are usually significantly higher than the corresponding mean values ​​under low-load conditions. Covariance matrix This indicates the amplitude of fluctuations of each process variable around its mean under that operating condition, as well as the cooperative relationship between different variables. The diagonal elements represent the variance of each variable, characterizing the intensity of its own fluctuation, while the off-diagonal elements represent the covariance between each pair of variables, revealing the degree of linear correlation between variables. For example, under stable operating conditions, the variance of each process variable is usually small, indicating a relatively stable operating state; however, during transitional phases such as load increases or decreases, even if classified under the same operating condition, the variance of the process variables may increase significantly. Thus, the aforementioned multiple operating conditions and the operating condition parameter information corresponding to each operating condition in the form of mean and covariance can collectively constitute a parameter set describing the probability distribution of multi-operating condition data.

[0101] S202, Construct a model training dataset based on the historical process dataset.

[0102] Understandably, when constructing a model training dataset, the entire historical dataset can be divided into consecutive, potentially overlapping time windows in chronological order. Each window constitutes a model training subset, and each model training subset corresponds to a time series segment, including the input process data arranged in chronological order and the nitrogen oxide concentration data corresponding to the input process data.

[0103] S203, The dynamic measurement prediction model is trained based on the model training dataset and the operating condition parameter set.

[0104] Specifically, refer to Figure 3As shown, since the core of the dynamic measurement and prediction model lies in its ability to capture the time-series dependence and dynamic evolution characteristics of industrial processes, and since there is no available historical operating condition information as input for the first time step, the initial operating condition weights for the first time step can be initialized when training the dynamic measurement and prediction model. These weights can be set as a uniformly distributed vector, representing that the process is in each operating condition with equal probability when there is no prior information. For example, when K=3, the initial weight vector is [1 / 3, 1 / 3, 1 / 3]; alternatively, they can be set based on historical statistical information. Then, the training process iterates through the input process data in each model training subset in chronological order. For each time step, the operations described in steps S301~S305 are executed sequentially.

[0105] S301, input the input process data of the current time step to the encoder to obtain the input process data feature vector of the current time step.

[0106] Understandably, the encoder is mainly responsible for performing nonlinear transformations and feature extraction on the input data, mapping the high-dimensional, potentially noisy, raw process data to a low-dimensional, dense latent representation space containing core information. Its output is the feature vector of the input process data at the current time step, which integrates information about the current process state.

[0107] S302, when the current time step is the first time step, the training working condition weights for the current time step are determined based on the input process data feature vector of the current time step, the initial working condition weights, and the working condition allocation module.

[0108] Here, when processing the start time of each model training subset, i.e., when the current time step is the first time step, the model inputs the feature vector of the input process data of the current time step together with the pre-initialized initial working condition weights into the working condition allocation module.

[0109] Specifically, the working condition allocation module can be constructed using a neural network, such as a combination of fully connected layers and the Softmax activation function. Its function is to dynamically calculate a new probability distribution vector based on the current data features, describing the probability distribution of the process at the current moment belonging to various working conditions; this is the training working condition weight for the current time step. The sum of all elements in this weight vector is 1, and the magnitude of each element directly reflects the degree of matching between the current process state and the working condition corresponding to that element.

[0110] Here, the working condition allocation module can concatenate the input process data features extracted by the encoder at the current time step with the working condition weight distribution at the previous time step, and then perform a nonlinear transformation on the concatenated features through a multilayer perceptron, and finally output the training working condition weights at the current time step through the Softmax activation function.

[0111] S303, if the current time step is not the first time step, determine the training condition weights for the current time step based on the input process data feature vector of the current time step, the condition weights of the previous time step, and the condition allocation module.

[0112] Here, when dealing with non-initial moments in the model training subset, i.e., when the current time step is not the first time step, the model inputs the feature vector of the input process data at the current time step and the working condition weights calculated from the previous time step to the working condition allocation module. At this time, the working condition allocation module also integrates two aspects of information: one is the input process data reflecting the current instantaneous state, and the other is the working condition weights from the previous moment reflecting the historical operating state, and then calculates and determines the training working condition weights for the current time step.

[0113] S304, the training condition weights of the current time step and the feature vector of the input process data of the current time step are input to the regressor to determine the predicted nitrogen oxide concentration estimate corresponding to the input process data of the current time step.

[0114] Specifically, when determining the predicted nitrogen oxide concentration estimate corresponding to the input process data at the current time step, the training operating condition weights for the current time step obtained in step S303 and the feature vector of the input process data at the current time step obtained in step S301 can be fused or concatenated, and then input together into the regressor. Here, the regressor is the core component of the model for predicting the final quality variables. It is mainly responsible for learning the complex nonlinear mapping relationship between the feature vector fused with the operating condition weights and the key quality variables, and outputting a specific numerical value, namely, the predicted nitrogen oxide concentration estimate corresponding to the input process data at the current time step.

[0115] S305, based on the decoder and the feature vector of the input process data at the current time step, determine the reconstruction process data corresponding to the input process data at the current time step.

[0116] Understandably, the decoder's structure is roughly symmetrical to the encoder's. Its task is to perform the inverse process of encoding, that is, to attempt to recover the original high-dimensional input process data from the low-dimensional feature vectors as much as possible. Its output is the reconstructed process data corresponding to the input process data at the current time step. This reconstruction process itself is not the model's final goal, but rather serves as an unsupervised auxiliary learning task. Its purpose is to force the encoder to learn feature vectors that contain enough original input information, thereby ensuring the richness and representativeness of the extracted features, preventing the model from overfitting to a single prediction task, and improving the model's generalization ability.

[0117] Here, after completing the aforementioned forward computation for one time step, the model saves the currently calculated operating condition weights as the "operating condition weights of the previous time step" needed to process the data of the next time step. This process is repeated until the entire training subset has been processed. Subsequently, based on the predicted nitrogen oxide concentration estimates, reconstructed process data, actual nitrogen oxide concentration data, and the aforementioned operating condition parameter set for all time steps, the model calculates the total loss function, which includes regression loss, reconstruction loss, and constraint loss. It then updates all learnable parameters of the encoder, operating condition allocation module, regressor, and decoder using the backpropagation algorithm and gradient descent optimizer. By iteratively executing this process multiple times on the entire training dataset, the model gradually learns how to accurately extract features, dynamically allocate operating condition weights, and predict key quality variables, ultimately converging to a stable performance state, completing the training.

[0118] For example, refer to Figure 4 The diagram shown is a structural schematic of a dynamic measurement prediction model proposed in this disclosure. The following explanation, using the collaborative working mechanism of the encoder, working condition allocation module, regressor, and decoder in the model as an example, is provided in conjunction with the illustrated structure:

[0119] Centered on the encoder module within the dashed box on the left of the diagram, when the input process data (i.e., time-series data collected by multi-dimensional sensors during industrial processes) enters the encoder, the encoder performs dimensionality reduction and feature extraction on the input data through a multi-layer nonlinear transformation network (represented by the fully connected layer structure shown by the matrix and circles in the diagram). Its output is the input process data feature vector Z for the current time step (represented by the column vector on the right side of the encoder in the diagram). This vector compresses the key information in the original high-dimensional data into a low-dimensional latent space, while also incorporating the current process state information.

[0120] Subsequently, the input process data feature vector Z is passed to the working condition allocation module (the unit labeled "working condition allocation module" in the figure). This module can use a neural network (such as a combination of fully connected layers and a Softmax function) to adjust the input process data feature vector Z and the working condition weights of the previous time step. (Or the initial working condition weights) are dynamically analyzed to generate the training working condition weights for the current time step. (The right-hand vector of the module in the diagram represents this). The weight values ​​of each working condition are normalized by the Softmax function to ensure that the sum of the weights is 1. The values ​​directly reflect the matching probability between the current process state and different preset working conditions.

[0121] In the regression prediction stage, as shown in the regressor section on the right side of the figure, the input process data feature vector Z and the training condition weights are... By splicing or weighted fusion (the "×" symbol in the figure indicates the information fusion operation), the information is input into the regressor (the structure of the output scalar after nonlinear transformation of the column vector on the right in the figure). The regressor learns the complex mapping relationship from the fused features to the nitrogen oxide concentration and finally outputs the predicted value (i.e. the predicted nitrogen oxide concentration estimate).

[0122] Simultaneously, as shown in the decoder path at the bottom of the figure, the input process data feature vector Z is also fed into the decoder for data reconstruction. The decoder maps the low-dimensional features back to the original data space through a symmetrical network structure (such as transposed convolution or fully connected layers), outputting the reconstructed process data. This path serves as an auxiliary task, forcing the encoder to retain key information from the original data and preventing the model from overemphasizing prediction tasks while neglecting feature generalization capabilities. For example, if the reconstructed data differs significantly from the original input, the reconstruction loss function constrains the model's optimization direction, enhancing the expressive power of the input process data feature vector Z.

[0123] Thus, through the closed-loop structure shown in the diagram (current operating condition weight) (This will be used as input for the next time step), the model realizes the temporal transmission and dynamic updating of operating condition information. The entire forward propagation process, through the synergistic effect of encoding-assignment-regression-decoding, not only achieves accurate prediction of nitrogen oxide concentration, but also ensures the robustness of the model through reconstruction constraints.

[0124] Specifically, refer to Figure 5 As shown, calculating the target loss of the dynamic measurement prediction model may include the following steps S501~S504:

[0125] S501, for each working condition, a target Gaussian distribution is defined based on the mean vector and covariance matrix corresponding to the working condition; the constraint loss of the dynamic measurement prediction model is calculated based on the spatial distance between the feature vector of the input process data at each time step in each model training subset and the target Gaussian distribution corresponding to each working condition.

[0126] Specifically, to incorporate process mechanism knowledge into the model training process and ensure that the feature distribution learned by the model is consistent with the actual physical conditions, the degree of difference between the feature vector of the input process data extracted at each time step and the representation of each working condition (i.e., the working condition parameter information corresponding to each working condition) in the working condition parameter set can be calculated. Here, spatial distance can refer to various metrics, including but not limited to Euclidean distance, Mahalanobis distance, or divergence metrics based on probability distributions, such as KL divergence. In actual calculations, the distance metric corresponding to each working condition k is usually calculated, and then a weighted sum is performed according to the working condition weight of the current time step to obtain the overall constraint loss value.

[0127] Thus, by introducing constraint loss, this disclosure can force the model to map data samples with similar physical properties to adjacent regions in the latent space, thereby enhancing the interpretability of the model and ensuring the separation of different operating conditions in the feature space.

[0128] For example, a preferred implementation of the aforementioned "spatial distance" is to calculate the KL divergence, the purpose of which is to use the physical knowledge represented by the set of working condition parameters obtained in step S201 as constraints to guide the structure of the model's potential space. The constraint loss is calculated as follows:

[0129] For each operating condition k, its parameters are directly derived from the operating condition parameter set, i.e., the mean vector. Covariance Matrix They collectively define the target Gaussian distribution. , This is represented as the k-th work condition in the set of work conditions; during training, the model infers a latent variable distribution for each work condition based on the current input (i.e., the feature vector of the input process data). Its parameters are Covariance Matrix Let z be the distribution of the input process data feature vector z under working condition k. Then, the KL divergence between the predicted distribution and the target distribution (i.e., the target Gaussian distribution for each working condition) can be calculated using the following formula, and these divergences are summed to obtain the final constraint loss, which can be expressed as:

[0130] ;

[0131] ;

[0132] in, It is the sum of the diagonal elements of the matrix. The inverse of the covariance matrix of the target working condition distribution. For working conditions The latent variable covariance matrix, For working conditions The latent variable mean, The average value under the target operating condition. As a potential spatial dimension, Let the determinant of the target operating condition covariance matrix be... For working conditions The determinant of the covariance matrix of latent variables.

[0133] S502, based on the predicted nitrogen oxide concentration estimates and nitrogen oxide concentration data at each time step in each model training subset, calculate the regression loss.

[0134] Here, the regression loss measures the difference between the model's predicted values ​​and the actual measured values. Variables used in this measurement can include mean squared error (MSE) and mean absolute error (MAE). Specifically, for each time step, the error between the predicted nitrogen oxide concentration estimate and the corresponding nitrogen oxide concentration data is calculated. Then, the errors from all time steps are aggregated, such as by taking the average. This regression loss term provides a clear direction for model optimization, directly driving the model to improve its core predictive capabilities.

[0135] S503 calculates the reconstruction loss based on the input process data and reconstruction process data at each time step in each model training subset.

[0136] Understandably, the reconstruction loss assesses the decoder's ability to recover the original input data from the feature vectors, typically using metrics such as mean squared error or cross-entropy. Specifically, for each time step, the difference between the original input data and the reconstructed data obtained through the encoder-decoder process is calculated. This loss term can serve as a regularization mechanism, ensuring that the feature vectors extracted by the encoder retain key information from the original data, preventing the model from over-focusing on a single prediction task and neglecting the overall structural features of the data, thereby improving the model's robustness and generalization ability.

[0137] S504, Based on the constraint loss, the regression loss, and the reconstruction loss, determine the target loss.

[0138] Understandably, after obtaining the constraint loss, regression loss, and reconstruction loss, they can be uniformly calculated to obtain the target loss, thus integrating the three different optimization objectives into a unified training objective. Here, a weighted summation approach can be used, assigning a specific weight coefficient to each loss term and then linearly combining them into the final target loss function. The weight coefficients reflect the importance of different training objectives in the overall optimization. For example, a larger weight can be given to the regression loss to ensure prediction accuracy, while appropriate weights can be given to the constraint loss and reconstruction loss to balance the model's physical consistency and feature representation ability. This comprehensive target loss function will serve as the final basis for optimizing model parameters, and then optimization algorithms such as gradient descent can be used to simultaneously optimize the model's performance in the three dimensions of prediction accuracy, physical plausibility, and feature learning.

[0139] S103, acquire the real-time process data of the coal-fired boiler and the operating condition weight of the previous moment corresponding to the real-time process data; and input the real-time process data and the operating condition weight of the previous moment corresponding to the real-time process data into the trained dynamic measurement and prediction model.

[0140] Understandably, real-time process data refers to data measured at the current moment, such as the current coal feed rate, air volume, and temperature, which can be obtained from sensors and data acquisition systems at the industrial site. The operating condition weights of the previous moment are vectors calculated by the model in the previous sampling period, representing the probability of the process being in each operating condition, and can be read from the system's storage unit.

[0141] Furthermore, the real-time process data and the operating condition weights from the previous moment can be input together into the dynamic measurement prediction model trained in step S102 to determine the estimated value of real-time nitrogen oxide concentration and complete the soft measurement of data.

[0142] S104, determine the current moment process feature vector of the real-time process data through the encoder; and determine the current moment operating weight based on the current moment process feature vector, the operating weight of the previous moment and the operating condition allocation module; and determine the real-time nitrogen oxide concentration estimate based on the current moment process feature vector, the current moment operating weight and the regressor.

[0143] Understandably, after real-time process data is input into the trained dynamic measurement prediction model, the model first extracts features from the current real-time data through the encoder to obtain the process feature vector at the current moment. Next, the operating condition allocation module combines this feature vector with the operating condition weights from the previous moment to calculate the operating condition weights at the current moment, reflecting the latest operating state. Subsequently, the regressor comprehensively utilizes the process feature vector and the operating condition weights at the current moment for forward calculation, ultimately outputting the real-time soft measurement estimate for the current moment, i.e., the predicted current NO. x Concentration. Here, the real-time soft-sensor estimate can be transmitted to the process control system or displayed to the operator as the output of a virtual sensor for real-time monitoring and closed-loop control. Simultaneously, the current operating condition weights can be stored and updated to the previous operating condition weights for the next prediction time, thus forming a closed-loop prediction system with time recursion characteristics, ensuring that the model can dynamically track changes in the operating conditions of the industrial process.

[0144] In some possible embodiments, after obtaining the real-time nitrogen oxide concentration estimate, it can be compared with a preset industry standard to achieve process optimization. When the real-time nitrogen oxide concentration estimate does not conform to the preset industry standard, adjustment suggestions for the operating parameters of the coal-fired boiler can be determined and output based on the current operating condition weight and the real-time nitrogen oxide concentration estimate. Specifically, when the predicted value exceeds the allowable range, targeted operational guidance can be generated based on the current dominant operating condition (i.e., the operating condition with the highest weight) and the magnitude of the real-time nitrogen oxide concentration estimate, combined with a preset process knowledge base. For example, when the predicted NO...x When the concentration exceeds the emission standard and the furnace is currently operating at high load, it is advisable to appropriately reduce the furnace temperature or increase the secondary air volume; or, in chemical processes, when the predicted product composition deviates from the target range, it is advisable to adjust the corresponding feed ratio or reaction conditions.

[0145] The industrial process soft measurement method, apparatus, storage medium, and computer equipment provided in this disclosure construct a dynamic measurement prediction model including an encoder, a working condition allocation module, and a regressor. By incorporating the working condition weights from the previous moment and real-time process data as model inputs, the trained model can effectively perceive and learn the dynamic correlation and evolution patterns of working conditions over time. In this way, by combining historical data training with real-time data and integrating working condition weights, the impact of the dynamic changes in working conditions during industrial processes on the measurement results is fully considered. Addressing the multi-working-condition operation characteristics of coal-fired boilers, the method utilizes easily measurable process variables (coal feed rate, air volume, temperature, and pressure) to achieve a more accurate and stable estimation of the key parameter—nitrogen oxide concentration—which is difficult to monitor online.

[0146] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0147] Based on the same inventive concept, this disclosure also provides an industrial process soft measurement device corresponding to the industrial process soft measurement method. Since the principle of the device in this disclosure for solving the problem is similar to that of the industrial process soft measurement method described above, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0148] Reference Figure 6 The diagram shown is a schematic of an industrial process soft measurement device 600 provided in an embodiment of this disclosure. The device includes:

[0149] The data acquisition module 601 is used to acquire a historical process dataset of a coal-fired boiler. The historical process dataset includes multiple historical process data arranged in chronological order and nitrogen oxide concentration data corresponding to each historical process data. The historical process data includes coal feed rate, primary air volume, secondary air volume, furnace temperature, and furnace pressure.

[0150] The model training module 602 is used to train the dynamic measurement prediction model based on the historical process dataset to obtain the trained dynamic measurement prediction model; wherein, the dynamic measurement prediction model includes an encoder, a working condition allocation module and a regressor;

[0151] The data input module 603 is used to acquire the real-time process data of the coal-fired boiler and the operating condition weight of the previous moment corresponding to the real-time process data; and input the real-time process data and the operating condition weight of the previous moment corresponding to the real-time process data into the trained dynamic measurement and prediction model.

[0152] The model execution module 604 is configured to determine the current moment process feature vector of the real-time process data through the encoder; and to determine the current moment operating weight based on the current moment process feature vector, the operating weight of the previous moment and the operating weight allocation module; and to determine the real-time nitrogen oxide concentration estimate based on the current moment process feature vector, the current moment operating weight and the regressor.

[0153] In some possible embodiments, the model training module 602 is specifically used for:

[0154] Based on the historical process dataset, various working conditions and corresponding working condition parameter information are determined to obtain a working condition parameter set;

[0155] A model training dataset is constructed based on the historical process dataset; wherein, the model training dataset includes multiple model training subsets, each of which corresponds to a time series segment, including input process data arranged in chronological order and nitrogen oxide concentration data corresponding to the input process data;

[0156] The dynamic measurement prediction model is trained based on the model training dataset and the operating condition parameter set.

[0157] In some possible embodiments, the operating condition parameter information includes a mean vector and a covariance matrix; the model training module 602 is specifically used for:

[0158] Based on the clustering analysis results of the historical process dataset, multiple operating conditions are identified; wherein, the multiple operating conditions include high load conditions, stable operation conditions, and low load conditions;

[0159] For each working condition, data samples that conform to the range of working condition parameters are selected from the historical process dataset to obtain a subset of historical process data corresponding to the working condition;

[0160] For each working condition, based on a subset of historical process data for that working condition, the mean vector and covariance matrix of that working condition are calculated as the working condition parameter information corresponding to that working condition.

[0161] In some possible embodiments, the dynamic measurement prediction model further includes a decoder; the model training module 602 is specifically used for:

[0162] Initialize the initial load weights for the first time step;

[0163] Iterate through the input data in each model training subset in chronological order. For each time step:

[0164] The input process data of the current time step is input into the encoder to obtain the input process data feature vector of the current time step;

[0165] When the current time step is the first time step, the training working condition weights for the current time step are determined based on the input process data feature vector of the current time step, the initial working condition weights, and the working condition allocation module.

[0166] If the current time step is not the first time step, the training working condition weights for the current time step are determined based on the input process data feature vector of the current time step, the working condition weights of the previous time step, and the working condition allocation module.

[0167] The training condition weights of the current time step and the feature vector of the input process data of the current time step are input into the regressor to determine the predicted nitrogen oxide concentration estimate corresponding to the input process data of the current time step.

[0168] Based on the decoder and the feature vector of the input process data at the current time step, the reconstruction process data corresponding to the input process data at the current time step is determined;

[0169] The target loss of the dynamic measurement prediction model is determined based on a preset loss function, and the model parameters of the dynamic measurement prediction model are adjusted based on the target loss until the training result meets the preset requirements, thus obtaining the trained dynamic measurement prediction model.

[0170] In some possible embodiments, the model training module 602 is specifically used for:

[0171] For each working condition, a target Gaussian distribution is defined based on the mean vector and covariance matrix corresponding to the working condition; the constraint loss of the dynamic measurement prediction model is calculated based on the spatial distance between the feature vector of the input process data at each time step in each model training subset and the target Gaussian distribution corresponding to each working condition.

[0172] Based on the predicted nitrogen oxide concentration estimates and nitrogen oxide concentration data at each time step in the training subset of each model, the regression loss is calculated;

[0173] Based on the input process data and reconstruction process data at each time step in the training subset of each model, the reconstruction loss is calculated;

[0174] The target loss is determined based on the constraint loss, the regression loss, and the reconstruction loss.

[0175] In some possible embodiments, the model execution module 604 is further configured to:

[0176] The operating condition weight at the current moment is stored as the operating condition weight of the previous moment for the next prediction moment.

[0177] In some possible embodiments, the model execution module 604 is further configured to:

[0178] When the estimated real-time nitrogen oxide concentration does not meet the preset industrial standard, an adjustment suggestion for the operating parameters of the coal-fired boiler is determined and output based on the current operating condition weight and the estimated real-time nitrogen oxide concentration.

[0179] Based on the same technical concept, this disclosure also provides a computer device. (See also...) Figure 7 The diagram shows the structure of a computer device 700 provided in this embodiment of the present disclosure, including a processor 701, a memory 702, and a bus 703. The memory 702 stores execution instructions and includes a main memory 7021 and an external memory 7022. The main memory 7021, also called internal memory, is used to temporarily store computational data in the processor 701, as well as data exchanged with external memory 7022 such as a hard disk. The processor 701 exchanges data with the external memory 7022 through the main memory 7021.

[0180] In this embodiment, the memory 702 is specifically used to store application code that executes the solution of this application, and its execution is controlled by the processor 701. That is, when the computer device 700 is running, the processor 701 communicates with the memory 702 through the bus 703, so that the processor 701 executes the application code stored in the memory 702, and then executes the method described in any of the foregoing embodiments.

[0181] The memory 702 may be, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc.

[0182] Processor 701 may be an integrated circuit chip with signal processing capabilities. The aforementioned processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor.

[0183] It is understood that the structures illustrated in the embodiments of this application do not constitute a specific limitation on the computer device 700. In other embodiments of this application, the computer device 700 may include more or fewer components than illustrated, or combine some components, or split some components, or have different component arrangements. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0184] This disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the industrial process soft measurement method described in the above-described method embodiments. The storage medium can be a volatile or non-volatile computer-readable storage medium.

[0185] This disclosure also provides a computer program product carrying program code. The program code includes instructions that can be used to execute the steps of the industrial process soft measurement method described in the above method embodiments. For details, please refer to the above method embodiments, which will not be repeated here.

[0186] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium; in another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0187] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this disclosure, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division; in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces; the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms.

[0188] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0189] In addition, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0190] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0191] Finally, it should be noted that the above-described embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A soft measurement method for industrial processes, characterized in that, include: A historical process dataset of a coal-fired boiler is obtained. The historical process dataset includes multiple historical process data arranged in chronological order and nitrogen oxide concentration data corresponding to each historical process data. The historical process data includes coal feed rate, primary air volume, secondary air volume, furnace temperature, and furnace pressure. The dynamic measurement prediction model is trained based on the historical process dataset to obtain the trained dynamic measurement prediction model; wherein, the dynamic measurement prediction model includes an encoder, a working condition allocation module and a regressor; Obtain real-time process data of the coal-fired boiler and the operating condition weight of the previous moment corresponding to the real-time process data; and input the real-time process data and the operating condition weight of the previous moment corresponding to the real-time process data into the trained dynamic measurement and prediction model. The encoder determines the current moment process feature vector of the real-time process data; and, based on the current moment process feature vector, the operating condition weight of the previous moment and the operating condition allocation module, determines the current moment operating condition weight; and, based on the current moment process feature vector, the current moment operating condition weight and the regressor, determines the estimated real-time nitrogen oxide concentration.

2. The method according to claim 1, characterized in that, The training of the dynamic measurement prediction model based on the historical process dataset includes: Based on the historical process dataset, various working conditions and corresponding working condition parameter information are determined to obtain a working condition parameter set; A model training dataset is constructed based on the historical process dataset; wherein, the model training dataset includes multiple model training subsets, each of which corresponds to a time series segment, including input process data arranged in chronological order and nitrogen oxide concentration data corresponding to the input process data; The dynamic measurement prediction model is trained based on the model training dataset and the operating condition parameter set.

3. The method according to claim 2, characterized in that, The operating condition parameter information includes a mean vector and a covariance matrix; the determination of multiple operating conditions and corresponding operating condition parameter information based on the historical process dataset yields an operating condition parameter set, including: Based on the clustering analysis results of the historical process dataset, multiple operating conditions are identified; wherein, the multiple operating conditions include high load conditions, stable operation conditions, and low load conditions; For each working condition, data samples that conform to the range of working condition parameters are selected from the historical process dataset to obtain a subset of historical process data corresponding to the working condition; For each working condition, based on a subset of historical process data for that working condition, the mean vector and covariance matrix of that working condition are calculated as the working condition parameter information corresponding to that working condition.

4. The method according to claim 3, characterized in that, The dynamic measurement prediction model further includes a decoder; training the dynamic measurement prediction model based on the model training dataset and the operating condition parameter set includes: Initialize the initial load weights for the first time step; Iterate through the input data in each model training subset in chronological order. For each time step: The input process data of the current time step is input into the encoder to obtain the input process data feature vector of the current time step; When the current time step is the first time step, the training working condition weights for the current time step are determined based on the input process data feature vector of the current time step, the initial working condition weights, and the working condition allocation module. If the current time step is not the first time step, the training working condition weights for the current time step are determined based on the input process data feature vector of the current time step, the working condition weights of the previous time step, and the working condition allocation module. The training condition weights of the current time step and the feature vector of the input process data of the current time step are input into the regressor to determine the predicted nitrogen oxide concentration estimate corresponding to the input process data of the current time step. Based on the decoder and the feature vector of the input process data at the current time step, the reconstruction process data corresponding to the input process data at the current time step is determined; The target loss of the dynamic measurement prediction model is determined based on a preset loss function, and the model parameters of the dynamic measurement prediction model are adjusted based on the target loss until the training result meets the preset requirements, thus obtaining the trained dynamic measurement prediction model.

5. The method according to claim 4, characterized in that, The determination of the target loss of the dynamic measurement prediction model based on a preset loss function includes: For each working condition, a target Gaussian distribution is defined based on the mean vector and covariance matrix corresponding to the working condition; the constraint loss of the dynamic measurement prediction model is calculated based on the spatial distance between the feature vector of the input process data at each time step in each model training subset and the target Gaussian distribution corresponding to each working condition. Based on the predicted nitrogen oxide concentration estimates and nitrogen oxide concentration data at each time step in the training subset of each model, the regression loss is calculated; Based on the input process data and reconstruction process data at each time step in the training subset of each model, the reconstruction loss is calculated; The target loss is determined based on the constraint loss, the regression loss, and the reconstruction loss.

6. The method according to claim 1, characterized in that, After determining the operating condition weight at the current moment, the process includes: The operating condition weight at the current moment is stored as the operating condition weight of the previous moment for the next prediction moment.

7. The method according to claim 6, characterized in that, After determining the estimated real-time nitrogen oxide concentration, the process includes: When the estimated real-time nitrogen oxide concentration does not meet the preset industrial standard, an adjustment suggestion for the operating parameters of the coal-fired boiler is determined and output based on the current operating condition weight and the estimated real-time nitrogen oxide concentration.

8. A soft measuring device for industrial processes, characterized in that, include: The data acquisition module is used to acquire historical process datasets of a coal-fired boiler. The historical process datasets include multiple historical process data arranged in chronological order and nitrogen oxide concentration data corresponding to each historical process data. The historical process data includes coal feed rate, primary air volume, secondary air volume, furnace temperature, and furnace pressure. The model training module is used to train the dynamic measurement prediction model based on the historical process dataset to obtain the trained dynamic measurement prediction model; wherein, the dynamic measurement prediction model includes an encoder, a working condition allocation module, and a regressor; The data input module is used to acquire real-time process data of the coal-fired boiler and the operating condition weight of the previous moment corresponding to the real-time process data; and input the real-time process data and the operating condition weight of the previous moment corresponding to the real-time process data into the trained dynamic measurement and prediction model. The model execution module is configured to determine the current moment process feature vector of the real-time process data through the encoder; and to determine the current moment operating weight based on the current moment process feature vector, the operating weight of the previous moment and the operating weight allocation module; and to determine the real-time nitrogen oxide concentration estimate based on the current moment process feature vector, the current moment operating weight and the regressor.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

10. A computer device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

Images