An industrial furnace working condition identification method and system

Abstract

This application relates to the field of industrial operating condition identification technology, and discloses a method and system for identifying the operating conditions of industrial furnaces and kilns. This method designs an adaptive sliding window transfer entropy as the fitness function of the optimization algorithm to accurately quantify whether the current time delay sequence is optimal. Simultaneously, it utilizes the DDS-POGIC algorithm to effectively and quickly search for the target sequence even in high-dimensional environments with numerous time delay sequences. Addressing the problems of severe multi-field and multi-phase coupling, dynamic and variable operating conditions, and severe data imbalance in large industrial furnace and kiln systems, a DWTD-SDAE operating condition identification model is constructed. This model effectively mines information from both process variables and operating condition samples, enhancing the guiding role of the operating condition target in the identification process. It can accurately and quickly identify the current operating condition even with imbalanced samples, helping operators quickly grasp operating condition fluctuations and adjust on-site operations in a timely manner.

Description

Technical Field

[0001] This application relates to the field of industrial operating condition identification technology, and in particular to a method and system for identifying the operating conditions of industrial furnaces and kilns. Background Technology

[0002] In high-energy-consuming and highly complex systems like industrial furnaces and kilns, accurate identification of operating conditions can not only effectively reduce the failure rate of industrial furnaces and kilns, but also provide a scientific basis for subsequent production scheduling. Therefore, accurate identification of the operating conditions of industrial furnaces and kilns has gradually become a widespread consensus in the field of industrial production.

[0003] Currently, methods for identifying the operating conditions of industrial furnaces and kilns are mainly divided into three categories: mechanistic model-based methods, expert system-based methods, and data-driven methods. Mechanism model-based methods primarily describe the system behavior by establishing physical, chemical, and thermodynamic process models of industrial furnaces and kilns. These models have a solid theoretical foundation, can explain complex phenomena in the process, and can predict future operating conditions by providing detailed process information. However, this type of method has limited application due to its complexity in establishment and maintenance and the need for substantial computational resources. Expert system-based methods use rule bases and inference engines to simulate the expert's decision-making process, thereby obtaining the operating condition identification results. They are highly effective when data is scarce or the model is incomplete, and the rules and decision-making process are transparent and easy to understand and modify. However, this type of method faces difficulties in knowledge updates and has limited coverage, making it difficult to handle complex and dynamically changing operating conditions. Data-driven methods build and analyze models by analyzing historical and real-time data. They often employ machine learning or deep learning algorithms, possessing strong adaptability and automated processing capabilities, and can handle complex nonlinear relationships.

[0004] Comparing the advantages and disadvantages of the above-mentioned operating condition identification methods, data-driven methods remain the most promising approach. These methods do not require in-depth understanding of the system's physical mechanisms and can automatically extract features from large amounts of data. They also have advantages in handling high-dimensional data and nonlinear relationships. Deep learning methods are increasingly being applied to industrial process monitoring and soft sensor modeling, with stacked autoencoders being one of the most widely used models. The on-site environment of industrial furnaces and kilns is complex and variable. During industrial furnace and kiln production, the time delays in process variables caused by differences in the reaction time of furnace charge transport and the temporal and spatial distribution of production units pose significant challenges to accurate operating condition identification. Stacked autoencoder models cannot effectively handle time-dependent operating condition data and are prone to overfitting when faced with complex relationships between multi-dimensional process variables and limited data samples, resulting in unsatisfactory performance in practical applications.

[0005] It is evident that existing technologies suffer from a lack of time delay information between process variables in industrial furnace and kiln data, resulting in insufficient feature extraction for industrial furnaces and kilns in situations with high dimensionality and limited abnormal operating condition samples. Summary of the Invention

[0006] This application provides a method and system for identifying the operating conditions of industrial furnaces and kilns, in order to solve the problem in the prior art that there is a lack of time delay information between process variables in industrial furnace and kiln data, and that the feature extraction of industrial furnaces and kilns is insufficient when there are high-dimensional and limited abnormal operating condition samples.

[0007] To achieve the above objectives, this application employs the following technical solution:

[0008] In a first aspect, this application provides a method for identifying the operating conditions of industrial furnaces and kilns, including:

[0009] S1: Collect raw operating data of the industrial furnace production system, calculate the time-series correlation matrix based on the raw operating data, and design an adaptive sliding window transfer entropy index based on the time-series correlation matrix.

[0010] S2: Determine the optimal time delay sequence based on the adaptive sliding window transfer entropy index;

[0011] S3: Based on the optimal time delay sequence, register the original working condition data in time sequence, and construct an industrial furnace working condition identification model based on the registered original working condition data;

[0012] S4: Real-time operating condition identification is performed based on the real-time data of the industrial furnace production system and the industrial furnace operating condition identification model.

[0013] Secondly, this application provides an industrial furnace operating condition identification system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the method described in the first aspect above.

[0014] Beneficial effects:

[0015] The industrial furnace operating condition identification method provided in this application reconstructs the time-series correlation matrix on the original data matrix based on the time-delay sequence. On this basis, an adaptive sliding window transfer entropy is designed as the fitness function of the optimization algorithm to accurately quantify whether the current time-delay sequence is optimal. Simultaneously, the DDS-POGIC algorithm is used to effectively and quickly search for the target sequence even in high-dimensional environments with numerous time-delay sequences. Addressing the problems of severe multi-field and multi-phase coupling, dynamic and variable operating conditions, and severe data imbalance in large industrial furnace systems, a DWTD-SDAE operating condition identification model is constructed. This model effectively mines information from both process variables and operating condition samples, enhancing the guiding role of the operating condition target in the operating condition identification process. The industrial furnace operating condition identification method based on DDS-POGIC and DWTD-SDAE proposed in this invention can accurately and quickly identify the current operating condition even under imbalanced sample conditions, helping operators quickly grasp the fluctuations in operating conditions and adjust on-site operations in a timely manner. Attached Figure Description

[0016] Figure 1 This is one of the flowcharts of a preferred embodiment of an industrial furnace operating condition identification method according to this application;

[0017] Figure 2 This is a second flowchart of a preferred embodiment of an industrial furnace operating condition identification method according to this application;

[0018] Figure 3 This is an adaptive sliding window transfer entropy curve of a preferred embodiment of this application;

[0019] Figure 4 This is a loss function curve of a preferred embodiment of this application;

[0020] Figure 5 This is the working condition identification confusion matrix of the preferred embodiment of this application. Detailed Implementation

[0021] The technical solution of this application is described clearly and completely below. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0022] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "a" or "one," and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "connected" or "linked," and similar terms, are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right," etc., are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship also changes accordingly.

[0023] It should be understood that current technologies suffer from a lack of time delay information between process variables in industrial furnace and kiln data, resulting in insufficient feature extraction for industrial furnaces and kilns in situations with high dimensionality and limited abnormal operating condition samples. Based on this, this application considers the lack of time delay information between process variables in industrial furnace and kiln data and the insufficient feature extraction of industrial furnaces and kilns under high-dimensional and limited abnormal operating condition samples. It proposes an industrial furnace and kiln operating condition identification method. This method is based on Dynamic Dual-Scale Particle Swarm Optimization with Genetic and Information Collaboration (DDS-POGIC) and Dual-Dimension Weighted Target-Driven Stacked Denoising Autoencoder Model (DWTD-SDAE). This method can effectively register data in time series, fully explore the effective information between industrial furnace and kiln data variables and samples, and improve the guiding role of operating condition targets in the operating condition identification operation. An industrial furnace and kiln operating condition identification model based on DDS-POGIC and DWTD-SDAE is constructed, achieving accurate identification of complex time-dependent industrial furnace and kiln operating conditions.

[0024] Please see Figures 1-2 This application provides a method for identifying the operating conditions of industrial furnaces and kilns, including:

[0025] S1: Collect raw operating data of the industrial furnace production system, calculate the time-series correlation matrix based on the raw operating data, and design an adaptive sliding window transfer entropy index based on the time-series correlation matrix.

[0026] S2: Determine the optimal time delay sequence based on the adaptive sliding window transfer entropy index;

[0027] S3: Based on the optimal time delay sequence, register the original working condition data in time sequence, and construct an industrial furnace working condition identification model based on the registered original working condition data;

[0028] S4: Real-time operating condition identification is performed based on the real-time data of the industrial furnace production system and the industrial furnace operating condition identification model.

[0029] Each piece of raw data for a specific operating condition represents a furnace condition sample and the value of a process variable at the corresponding time. Specifically, process variables are a general term encompassing temperature, flow rate, and so on.

[0030] The aforementioned industrial furnace operating condition identification method reconstructs the time-series correlation matrix based on the time-delay sequence on the original data matrix. On this basis, an adaptive sliding window transfer entropy is designed as the fitness function of the optimization algorithm to accurately quantify whether the current time-delay sequence is optimal. Simultaneously, the DDS-POGIC algorithm is utilized to effectively and quickly search for the target sequence even in high-dimensional environments with numerous time-delay sequences. Addressing the problems of severe multi-field and multi-phase coupling, dynamic and variable operating conditions, and severe data imbalance in large industrial furnace systems, a DWTD-SDAE operating condition identification model is constructed. This model effectively mines information from both process variables and operating condition samples, enhancing the guiding role of the operating condition target in the operating condition identification process. The industrial furnace operating condition identification method based on DDS-POGIC and DWTD-SDAE proposed in this invention can accurately and quickly identify the current operating condition even under imbalanced sample conditions, helping operators quickly grasp the fluctuations in operating conditions and adjust on-site operations in a timely manner.

[0031] The following is a detailed description of the specific implementation scheme of the above-mentioned industrial furnace and kiln condition identification method, using a complete example:

[0032] (1) Based on the temporal correlation matrix, an adaptive sliding window transfer entropy index is designed.

[0033] First, the industrial furnace operating condition is defined as a zero-delay reference variable. Based on the time-delay sequence, the process variables are reconstructed in time series to form a multivariate time series. The reconstructed time series correlation matrix composed of the multivariate time series and the industrial furnace operating condition is used to reflect different time-delay characteristics. Then, an adaptive sliding window transfer entropy is designed to capture the dynamic impact of process variable changes on operating condition changes, accurately quantifying the causal strength between the multivariate time series and the industrial furnace operating condition at different stages. Specifically, the following steps are included:

[0034] Step 1: Assume that in an industrial furnace production system, N sensors are installed from top to bottom, and a raw data matrix with a sample size of M is obtained by sampling at equal periods along the time axis.

[0035]

[0036] In Equation (1), the subscripts of the elements in the original time series data matrix represent process variables, and the superscripts represent the sample sampling time. Let i be the time series data of the i-th process variable, representing the process variable. A sample sequence sampled M times consecutively along the time axis. To avoid the influence of the data dimensions of industrial furnaces on the time delay characteristics during multivariate time delay estimation, the original time series data matrix is... Normalization yields X:

[0037]

[0038] X = [x1, x2, ..., x i ,...,x N (3)

[0039] In formula (2) This represents the normalized result of the i-th process variable at time t. and Let be the maximum and minimum values ​​of the i-th process variable in all data samples, respectively. These are the normalized process variables.

[0040] Assuming the multivariate correlation time delay parameter between the zero-delay reference variable (industrial furnace operating condition) and each process variable is represented by the time delay sequence Γ, and the sampling period of each process variable in the original data is denoted as T, then the time delay parameter of the i-th process variable relative to the industrial furnace operating condition needs to satisfy:

[0041] Γ=DT=[d1T,d2T,...,d i T,...,d N T](4)

[0042] In equation (4), D is the time base sequence, and d i It is the time base of the i-th process variable, and is a dimensionless integer.

[0043] For industrial furnace production processes, a time-series correlation matrix corresponding to the time-based sequence is obtained by reconstructing the original time-series data matrix in the spatiotemporal dimension. First, K data points with a sampling period of T, starting from time t, are selected from the industrial furnace operating condition y to construct a zero-delay variable sequence y':

[0044] y'=[y t ,y t+T ,...,y t+(K-1)T ] T (5)

[0045] Equation (5) should satisfy KT≥max(τ1,...,τ) i,...,τ N Furthermore, the selected data must include the process of changing operating conditions. The operating conditions of industrial furnaces are acquired in real time; therefore, changes in operating conditions during production lag behind changes in process variables. Let the process variable x... i The hysteresis coefficient is τ i Then the selected industrial furnace operating conditions correspond to the process variable time series x. i 'for:

[0046]

[0047] In the formula, d i x is the process variable i The time base value, d i =τ i / T.

[0048] According to the above reconstruction rules, based on the time delay sequence τ=[τ1,τ2,...,τ N Dividing by the sampling period yields the time base sequence, which is then used to construct the time series correlation matrix X', satisfying the following relationship:

[0049]

[0050] Step 2: Reconstruct the time-series data of industrial furnaces and kilns based on the time-delay sequences to obtain the time-series correlation matrix. Calculate the causal strength of the time-series correlation matrix under different time-delay sequences, and use the level of causal strength to reflect the degree of time-delay dependence of the reconstructed time-series correlation matrix. Therefore, an adaptive multivariate transfer entropy method based on a sliding window is proposed to dynamically track the changes in causal relationships within each window and describe the causal strength between the multivariate time series and changes in operating conditions.

[0051] First, obtain the current time-delay sequence Γ and select an appropriate embedding dimension N to construct the time-series correlation matrix. In the actual production process of industrial furnaces, the dynamic changes of process variables have a strong causal relationship in the short period before the changes in the operating conditions of the industrial furnace. Therefore, based on the characteristics of the actual production process of industrial furnaces, select a time window size ν, and obtain the time-series correlation matrix X for each time window by dividing the time series and the zero-delay variable sequence through a sliding window. w 'as follows:

[0052]

[0053] Then estimate the temporal correlation matrix X. w The joint probability of ' and marginal probability The goal is to capture the temporal correlation matrix X for each window. w The probability of 'internal x' and y' occurring simultaneously, and the distribution of x' itself, are used to estimate the overall probability distribution of the data:

[0054]

[0055] In equations (9) and (10), g represents the bandwidth using the Gaussian kernel function, and x i This represents the i-th sample within the time window. Since the industrial furnace operating conditions selected for the zero-delay reference are discrete variables and typically remain constant over long periods, a frequency estimation method is used to estimate the probability distribution of these conditions. The types of industrial furnace operating conditions are recorded as follows: The possible values ​​of the working condition are Then, within the time window, the probability of each operating condition of the industrial furnace / kiln occurring is:

[0056]

[0057] In equation (11), c(y=k) is the number of samples with condition k within each window.

[0058] Based on the results of joint and marginal probabilities, the joint and independent distributions are derived. The sliding window transfer entropy measures the degree of information transfer by the difference between the joint and independent distributions, quantifying the window X. w The causal influence of 'internal x' on 'y'. In industrial furnace systems, the computational complexity of joint distribution estimation for high-dimensional data is too high. Using local probability logarithmic differences to capture information flow trends is sufficient to meet the needs of time-series data delay analysis tasks. The transition entropy (MTE) of the nth time window. n The calculation formula is:

[0059]

[0060] The operating conditions of industrial furnaces and kilns vary unevenly, with significant fluctuations in some time windows and relatively stable conditions in others. Therefore, an adaptive approach is adopted, adjusting the window weights based on the degree of fluctuation to better capture causal relationships within industrial furnaces and kilns. Finally, the causal strength of each time window is combined to ensure that the causal analysis results reflect temporal trends and global causal relationships, yielding the adaptive sliding window transfer entropy (MTE).

[0061]

[0062] In equation (13), n is the number of sliding time windows, and c n The normalized adaptive weight values ​​for each time window.

[0063] In step (1), an adaptive sliding window transfer entropy was designed as a time delay estimation index. The size of the time window was determined by analyzing the industrial furnace production process. The transfer entropy in each window was calculated by sliding the window along the time axis. Finally, the transfer entropy of all windows was weighted and averaged to obtain the overall time delay estimation effect.

[0064] (2) Use the time delay estimation optimization algorithm to obtain the optimal time delay sequence and register the original data in time sequence.

[0065] In this step, to quickly and efficiently find the optimal time delay sequence, the DDS-POGIC optimization algorithm is proposed to search for the true time delay parameters between process variables, transforming the multivariate time delay estimation problem into an optimization problem of solving the maximum adaptive sliding window transfer entropy of a series of time series correlation matrices. Specifically, it includes the following steps:

[0066] Step 1: Due to the dynamic nature of the process flow, there are often different time delays between variables in industrial furnace and kiln operating data. Based on the mechanistic knowledge of different industrial furnace and kiln production processes, the boundaries of the time delay sequences of each variable are defined, and the process variable x is recorded. i The boundary of the delay interval is Therefore, the boundary of the time base sequence is obtained as follows: Where d i Pick Integer. The number of possible values ​​for the time delay sequence is... The large number of time-series correlation matrices corresponding to high-dimensional industrial furnace data leads to high computational complexity and easy trapping in local optima when traditional algorithm models search for time delay sequences. Therefore, an improved Particle Swarm Optimization (PSO) algorithm is proposed. By introducing a particle information sharing mechanism, the solution process for time delay sequences is further optimized on the basis of global search.

[0067] Step 2: First, initialize the position vector P of the original particle population representing the time base sequence using the Latin cubic sampling method:

[0068] P = {p i,j}

[0069]

[0070] In equation (14) π j (i) is a random permutation function of the particle number i along the process variable dimension j, u i,j ∈[0,1) is a random number, K is the total number of particles in the population, p i,j Let be the value of the position vector of the i-th particle in the j-th dimension.

[0071] Next, to balance the global search capability and local optimization capability of the particle swarm, the particle swarm is divided into two scale subgroups: large-scale particle swarm and small-scale particle swarm. Large-scale and small-scale particle swarms share optimal information to prevent the global optimum from being controlled by a single particle. The two scales of particle swarms are mainly described by particle inertia weights and particle velocities. The weight adjustment strategy and particle velocity of the large-scale particle swarm are as follows:

[0072]

[0073] In equation (15), ω and v max These represent the inertial weights and maximum velocity of the large-scale particle swarm, respectively, ω. max and ω min These are the upper and lower bounds of the weights, k and k, respectively. max The maximum number of iterations, The phase shift determines the initial state of the inertial weights, and β is the exponential decay coefficient that controls the decay rate of the inertial weights. The weight adjustment strategy and particle velocity for small-scale particle swarms are as follows:

[0074]

[0075] In equation (16), max(v) max ) and min(v max The values ​​() represent the upper and lower limits of the maximum velocity, respectively. The inertia weight of a large-scale particle swarm is a periodic fluctuation function related to the number of iterations, which helps in adaptive searching at different stages. In contrast, the inertia weight and maximum velocity of a small-scale particle swarm are monotonically decreasing functions related to the number of iterations, which helps refine the accuracy of local optima within a smaller scope.

[0076] The small-scale particle swarm is divided into multiple dynamic subgroups with a smaller number of particles to avoid premature convergence of the entire population due to a single local optimum. The update formulas for the position and velocity vectors of the large-scale and small-scale particle swarms are as follows:

[0077]

[0078] pbest in equations (17) and (18) i lbest is the best position of the i-th particle before the current iteration number, lbest is the global best position of the large-scale particle swarm, and sbest is the best position. j It is the globally optimal position of the j-th small-scale subgroup.

[0079] Step 3: Introduce an information-sharing mechanism into the particle swarm optimization algorithm. Through cooperation between two types of particle swarms, excellent solutions from different search spaces are integrated and mutually passed on and referenced, improving the algorithm's ability to explore at multiple levels and refine local optima. Specifically, when a large-scale subgroup finds the current global optimum, the algorithm moves that particle to a small-scale subgroup for a more refined search. At this point, multiple small-scale subgroups can perform parallel searches for multiple local optima, increasing the search depth for diverse solutions. Simultaneously, it ensures that the large-scale subgroup does not contain the current global optimum, thus preventing the search direction from being constrained by existing optima and maintaining greater search flexibility, helping to escape the limitations of local optima. The information-sharing mechanism is as follows:

[0080] Triggering condition: f(lbest) > f(sbest) j )

[0081] Particle migration rule: Move the globally optimal particle lbest from the large-scale subgroup S l Migrate to the j-th small-scale subgroup S s_j Then, to ensure that the size of each subgroup remains unchanged, from S s_j Randomly selected particles pbest s (i) Migration back to S l middle,

[0082] S l * =(S l U{pbest s (i)})\{lbest}

[0083] S s_j * =(S s_j U{lbest})\{pbest s (i)}(19)

[0084] The velocity formulas for updating large-scale and small-scale subgroups are as follows:

[0085] v i =ω×v i +c1·r1·(pbest i -p i )+c2·r2·(lbest new -p i (20)

[0086] v i =ω×v i +c1·r1·(pbest i -p i )+c2·r2·(lbest-p i)(twenty one)

[0087] Step 4: The global exploration capability of large-scale particle swarms is key to the information sharing mechanism, and it is further enhanced through an inertial weight control strategy after each iteration update. A trigger condition ε is set: when the average Euclidean distance between any two particles in the large-scale particle swarm is sufficiently small, it indicates that the current particle swarm is over-concentrated.

[0088]

[0089] D av <ε (22)

[0090] At this point, by adding an extra value Δω to the original inertia weight formula, we obtain the large-scale particle swarm inertia weight ω':

[0091] ω'=ω+Δω(23)

[0092] And then re-diffusion is performed on the large-scale subgroups:

[0093] v i =ω'×v i +c1·r1·(pbest i -p i )+c2·r2·(lbest-p i )

[0094] p i =p i +v i (twenty four)

[0095] Step 5: Using an information-sharing mechanism for global and local collaborative optimization effectively improves the model's convergence speed and search accuracy. However, the large solution space corresponds to a large number of iterations, and as the number of iterations increases, the jumping ability of particles decreases, weakening the global search capability. To maintain solution diversity and expand the search space of the particle swarm, a genetic operation is performed on the optimal population of individual particles after a certain number of iterations to generate new candidate solutions. Then, the current particle swarm is randomly shuffled and recombined to change the topology and jumping scale of the particles, thereby introducing more randomness into the exploration of complex multimodal problems.

[0096] Triggering condition: k%Re = 0

[0097] Parent individuals are obtained using an elite retention and random selection strategy, and offspring are generated through a simulated binary crossover mechanism.

[0098] child1 = 0.5 × [(1 + β)] q )×parent1+(1-β q )×parent2]

[0099] child2 = 0.5 × [(1-β] q )×parent1+(1+β q )×parent2]

[0100]

[0101] In equation (25) β q It is a parameter that controls the crossover strength, depending on the random number u∈[0,1] and the distribution exponent η. c Parent1 and Parent2 are two individuals drawn from the parent generation. Then, using binomial mutation, certain dimensional values ​​of the offspring individuals are altered to maintain population diversity. The particle mutation formula is:

[0102] new child i (j) = child i (j)+Δ j

[0103]

[0104] In equation (26) Δ j Let u ∈ [0,1] be the offset, and η be a uniformly distributed random number. m It is a distribution index that controls the magnitude of mutation. Then, the current particle is evaluated according to the fitness function, and the optimal position of the individual in the current large-scale and small-scale subgroups is updated. Finally, the particle swarm currently searching is randomly shuffled and reorganized, changing the neighborhood structure of the particles. This causes particles that were originally exploring at a large scale to begin a refined search, and particles that were originally searching at a refined scale to begin a large-scale leap to explore local solutions.

[0105] The adaptive sliding window transfer entropy is set as the fitness function. The global optimal particle position pbest is finally searched through the time delay estimation optimization algorithm based on DDS-POGIC. This is the optimal time base sequence D required for industrial furnace and kiln condition identification. The optimal time delay sequence is determined based on the optimal time base sequence. Specifically, the time base sequence is a dimensionless integer sequence, and the time delay sequence = time base sequence * sampling period.

[0106] In step (2), a time delay estimation optimization algorithm based on DDS-POGIC is proposed. Large-scale and small-scale particle swarms are set up and share optimal information to avoid the global optimum being controlled by a single particle. Genetic operations are introduced to update the population topology. The resulting time delay sequence provides scientific input for subsequent identification.

[0107] (3) Register the furnace data in time sequence according to the optimal time delay sequence and establish an industrial furnace operating condition identification model.

[0108] Monitoring abnormal operating conditions can be viewed as a multi-class classification problem. In industrial furnace production, the probability of abnormal operating conditions occurring is relatively low, and the number of samples that can be collected is also limited, leading to an imbalance in the operating condition samples. Furthermore, considering that interrelated process variables have different impacts on the formation of abnormal operating conditions in industrial furnaces, a method of defining dynamic factors and contribution weights in both the sample dimension and the process variable dimension can be used to improve the mining of raw data. To enhance the guiding role of the target operating condition in abstract feature extraction and the utilization of features at different layers, the operating condition identification error term can be added to the original loss function during model training to fully utilize the operating condition label information. Finally, the extracted operating condition-related abstract features are used as input to the softmax layer to obtain the industrial furnace operating condition identification results. The specific steps include:

[0109] Step 1: Treat industrial furnace operating condition identification as a multi-classification problem and construct a nonlinear model F:

[0110] y t =F(x) t (27)

[0111] In equation (27), t is a time index, and x t Let y be the input sample at time t. t This outputs the industrial furnace / kiln operating condition identification value. There are H categories recording the industrial furnace / kiln operating conditions, i.e., y t ∈{1,2,...,H}, where y t =1 indicates that the current condition is normal.

[0112] Step 2: Traditional stacked denoising autoencoder models assign the same priority to all samples and process variables without distinction. To ensure that the extracted abstract features include the dynamic relationships between samples under different operating conditions, process variables with varying degrees of correlation, and the target being identified, contribution factors are defined in both the sample and process variable dimensions of the industrial furnace operating data.

[0113] Based on the key iterative steps of the k-means algorithm, the contribution factor of each sample can be characterized to some extent by the Euclidean distance between the sample and the cluster centers of the corresponding working condition samples. Let the time-series registered industrial furnace data be {X}. h ,Y h}={(x1,y1),(x2,y2),...,(x t ,y t ),...,(x M ,y M The industrial furnace data after time-series registration is a new dataset that eliminates the effects of time delay after registering the original operating data collected initially on the time axis. Where x tLet be the data sample at time t, and M be the total number of samples. The cluster centers for different operating conditions are denoted as the query sample x. q :

[0114]

[0115] In equation (28), N represents the number of process variables. Input sample x t Euclidean distance d between the query sample and the query sample t for:

[0116]

[0117] Obtain the Euclidean distance d t Then, the sample contribution factor is set to c. t :

[0118]

[0119] In equation (30), σ is an adjustment parameter that controls the deceleration rate of the contribution factor with distance. The exponential function can smooth out dynamic factors.

[0120] Different process variables are coupled and influence each other, triggering various complex physicochemical reactions within industrial furnaces and kilns, thus affecting the formation of abnormal operating conditions. To reduce redundant information irrelevant to the operating condition identification results in the extracted abstract features, the Maximum Information Coefficient (MIC) algorithm is used to quantify the nonlinear relationship between different process variables and the operating condition output. The correlation coefficient λ between the j-th process variable and the target output is then used. j Defined as:

[0121]

[0122] In equation (31) λ j ∈(0,1], B is the optimal recommended value for the process variable. The larger the correlation coefficient, the higher the contribution weight of the corresponding process variable. The weights ωj of each process variable are:

[0123]

[0124] After obtaining the corresponding contribution factors and contribution weights in two dimensions of industrial furnace operating data, the input layer is reconstructed through encoding and decoding processes, thereby designing a new reconstruction loss function L. c-ω (W,b) is used to train the recognition model:

[0125]

[0126] Step 3: The two-dimensional weighting method emphasizes to some extent the contribution of abnormal operating condition samples and highly correlated process variables to the operating condition identification results. However, the abstract features extracted by the hidden layer still focus on the reconstruction of the input data rather than the effective interpretation of the target output. Therefore, using the hidden layer features, the identification error term with labeled data in the input is calculated, and a new pre-trained loss function L is constructed based on the reconstruction loss function. c-ω-T (W,b).

[0127] The hidden layer features of the denoising autoencoder (DAE) model are h = [h1, h2, ..., h d ] T Input h into the softmax layer to obtain the recognition output. This leads to the pre-training loss function L. c-ω-T (W,b):

[0128]

[0129] In equation (34), M is the total number of input samples. h Let L be the number of labeled samples in the input, and λ be a tradeoff coefficient used to balance the reconstruction loss term and the recognition error term in the loss function. During training, the loss function L is minimized. c-ω-T (W,b), thus obtaining the model parameter set θ={W f ,b f W g ,b g}

[0130] Step 4: Stack L DAEs layer by layer to form a deep neural network. Use the hidden layer features of the previous DAE as the input layer data of the next DAE. Then, the hidden layer features of the l-th DAE can be obtained as follows: At the same time h l It is also the input of the (l+1)th DAE. The two-dimensional weights of the input data of the lth DAE are calculated according to equations (30) and (32), resulting in... and Then the loss function L of the l-th DAE c-ω-T (θ l )for:

[0131]

[0132] In equation (35) N l It is the dimension value of the current DAE input layer.

[0133] After the DWTD-SDAE model is pre-trained, a softmax layer is added to the top of the model, and the network parameters θ are fine-tuned in reverse. The hidden layer features h of the Lth DAE... LThe working condition identification result is obtained through the softmax output layer.

[0134]

[0135] Furthermore, the hidden layer features extracted from the pre-trained model are fed into the newly added softmax layer to fine-tune the previously pre-trained model.

[0136] f in equation (36) o and θ o ={W o ,b o} represents the activation function and network parameters of the softmax layer. The optimal network parameters θ' = {W', b'; W} of the DWTD-SDAE model are obtained by minimizing the target output loss function J using the backpropagation algorithm. o ',b o Then, the final deep network used for classification is obtained through the network parameters. The target output loss function J is expressed as:

[0137]

[0138] In equation (37), M o y represents the number of samples used for model fine-tuning. t,i It is a K-dimensional one-hot encoded vector. It is a K-dimensional probability vector. Specifically, the model is fine-tuned using this loss function to obtain the final network structure parameters.

[0139] Furthermore, in the task of identifying the operating conditions of industrial furnaces and kilns, the optimal network parameters θ' of the DWTD-SDAE model are first trained using the time-registered industrial furnace and kiln operating condition data. Then, the DWTD-SDAE model is used to extract abstract features from the new industrial furnace and kiln data, and the abstract features are fed into the output layer to obtain the final operating condition identification result.

[0140] In step (3), a dual-dimensional weighted DWTD-SDAE working condition identification model was constructed. Contribution factors were defined for the samples, contribution weights were assigned to the variables, and an identification error term was added to the model loss function. This improved the guiding role of the target and the utilization of features at different levels when extracting working condition-related features, thereby enhancing the accuracy of working condition identification in complex industrial furnace systems.

[0141] In summary, this application proposes a method for identifying the operating conditions of industrial furnaces and kilns, which can effectively cope with complex time dependencies and high nonlinearity, ensuring accurate identification of industrial furnace and kiln operating conditions. Real-time data analysis and identification of potential fault risks also improve the accuracy and response speed of monitoring, providing strong support for the safe production of industrial furnaces and kilns.

[0142] In one example, using a blast furnace in an ironmaking plant as a test platform, the invented industrial furnace condition identification method based on information sharing delay optimization was applied to the identification of abnormal furnace conditions in this blast furnace. After system installation, historical data of the blast furnace is first read to find the optimal delay sequence of process variables. Figure 3 The adaptive sliding window transfer entropy change curve of this invention is shown during the optimization process. It can be seen that the DDS-POGIC-based optimization algorithm repeatedly escapes local optima and maintains a high global search capability even in the later stages of iteration. Input data is registered temporally based on the optimal delay sequence. Figure 4 The training loss function curve of the DWTD-SDAE-based working condition recognition model is shown. The loss function curve decreases significantly and fluctuates steadily in the later stages of iteration, indicating that the present invention can effectively avoid overfitting. The working condition model achieves a recognition accuracy of 99.329%, a precision of 99.331%, a recall of 99.329%, and an F1 score of 99.328%. These high performance metrics demonstrate the comprehensiveness of the present invention in furnace condition recognition tasks. Figure 5 The confusion matrix for blast furnace condition identification results is used to validate the blast furnace condition identification. Furthermore, the hit rates of the condition identification model for normal, suspended charge, and pipeline conditions are 99.219%, 100.0%, and 98.3%, respectively. Figure 5 and three The hit rate of the various furnace conditions shows that the present invention has a high hit rate for the relatively small number of suspended furnace condition samples and pipeline furnace condition samples. This indicates that the present invention can focus more attention on the fewer abnormal furnace condition samples, thereby better learning abnormal furnace condition information, thus greatly improving the accuracy of abnormal furnace condition identification and reducing the occurrence of misjudgment and missed judgment.

[0143] This application also provides an industrial furnace operating condition identification system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above method.

[0144] The industrial furnace and kiln condition identification system can implement various embodiments of the above-mentioned methods and achieve the same beneficial effects, which will not be elaborated here.

[0145] The preferred embodiments of this application have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of this application without inventive effort. Therefore, any technical solutions that can be obtained by those skilled in the art based on the concept of this application through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for identifying the operating conditions of industrial furnaces and kilns, characterized in that, include: S1: Collect raw operating data of the industrial furnace production system, calculate the time-series correlation matrix based on the raw operating data, and design an adaptive sliding window transfer entropy index based on the time-series correlation matrix. S2: Determine the optimal time delay sequence based on the adaptive sliding window transfer entropy index; S3: Based on the optimal time delay sequence, register the original working condition data in time sequence, and construct an industrial furnace working condition identification model based on the registered original working condition data; S4: Real-time operating condition identification is performed based on the real-time data of the industrial furnace production system and the industrial furnace operating condition identification model. S1 includes: S11: Collect raw operating condition data of the industrial furnace production system containing the process of changing operating conditions, construct a zero-delay variable sequence, and determine the time series of the process variables corresponding to the raw operating condition data; S12: Construct a time series correlation matrix based on the zero-delay variable sequence and the time series; estimate the probability distribution of the working conditions within each time window using a frequency estimation method based on the time series correlation matrix; and determine the adaptive sliding window transfer entropy based on the probability distribution of the working conditions within each time window. S11 includes: For industrial furnace production systems, the sample size obtained by periodic sampling along the time axis is: The original operating data is used to form the original time series data matrix. : (1) In the original time-series data matrix, the subscripts of the elements represent process variables, and the superscripts represent the sample sampling time. For the first Time series data of process variables, representing process variables Continuous sampling on the time axis The sample sequence of the next time; The original time series data matrix Normalization process is performed to obtain : (2) (3) In the formula Indicates the first The process variables are in The result of time-mapping normalization and The first The maximum and minimum values ​​of each process variable across all data samples. These are the normalized process variables; Working conditions from raw working condition data Selected from The sampling period starts at time . of These data constitute a zero-delay variable sequence. : (5) Let process variables The lag coefficient is Then the selected industrial furnace operating conditions correspond to the process variable time series. for: (6) In the formula, For process variables The time base value, ; S12 includes: According to the time delay sequence Divide by the sampling period to obtain the time base sequence, and then construct the time series correlation matrix. It satisfies the following relationship: (7) Select the size of the time window based on the actual production process characteristics of industrial furnaces and kilns. The time series and the zero-delay variable series are divided by a sliding window. Obtain the time series correlation matrix for each time window. as follows: (8) Estimating the temporal correlation matrix joint probability and marginal probability It satisfies the following relationship: (9) (10) In the formula This indicates the bandwidth using the Gaussian kernel function. Represents the first time in the time window One sample; The probability distribution of operating conditions is estimated using frequency estimation methods, and the types of operating conditions in industrial furnaces and kilns are recorded as follows: The value of the working condition is Then, within the time window, the probability of each operating condition of the industrial furnace occurring is: It satisfies the following relationship: (11) In the formula, The working conditions within each window are The number of samples; Based on the results of joint and marginal probabilities, the joint and independent distributions are derived. The sliding window transfer entropy, through the difference between the joint and independent distributions, measures the degree of information transfer and quantifies the window's characteristics. Inside right Calculate the causal effect of the first The transition entropy of a time window It satisfies the following relationship: (12) By combining the causal strengths of each time window, we obtain the adaptive sliding window transfer entropy. It satisfies the following relationship: (13) In the formula The number of sliding time windows. Normalized adaptive weight values ​​for each time window; S2 includes: S21: Record process variables The boundary of the delay interval is The boundaries of the time base sequence are obtained: ； in Pick Integer; the number of possible values ​​for the time delay sequence is ; S22: Initialize the position vectors of the original particle population representing the time base sequence using the Latin cube sampling method. : (14) In the formula, It is the number of particles In the process variable dimension Random permutation function on, It is a random number. It is the total number of particles in the population. For the first The position vector of the nth particle is at the... Values ​​on the dimension; S23: Divide the particle swarm into two scale subgroups, including a first-scale particle swarm and a second-scale particle swarm. The weight adjustment strategy and particle velocity of the first-scale particle swarm are as follows: (15) In the formula, and These represent the inertial weights and maximum velocities of the first-scale particle swarm, respectively. and These are the upper and lower limits of the weights, respectively. The maximum number of iterations, The initial state of the inertial weights is determined by the phase shift. It is the exponential decay coefficient, which controls the decay rate of the inertial weight; The weight adjustment strategy and particle velocity for the second-scale particle swarm are as follows: (16) In the formula and These represent the upper and lower limits of the maximum velocity, respectively; where the inertia weight of the first-scale particle swarm is a periodic fluctuation function related to the number of iterations; and the inertia weight and maximum velocity of the second-scale particle swarm are monotonically decreasing functions related to the number of iterations. S24: Divide the second-scale particle swarm into multiple dynamic subgroups with a particle count less than a set value. The update formulas for the position and velocity vectors of the first-scale and second-scale particle swarms are as follows: (17) (18) In the formula It is the first The optimal position of each particle before the current iteration number. It is the globally optimal position of the first-scale particle swarm. It is the first The global optimal position of each second-scale subgroup; S25: Introduce an information sharing mechanism into the particle swarm optimization algorithm. The information sharing mechanism is as follows: Triggering conditions: ; Particle migration rule: Transfer the globally optimal particle... From the first scale subgroup Migrate to the Second-scale subgroup Afterwards, from Randomly selected particles Relocation middle, (19) The rate of updating the first-scale and second-scale subgroups satisfies the following relationship: (20) (21) Set trigger conditions When the average Euclidean distance between any two particles in the first-scale particle swarm is less than a set distance threshold, an additional value is added to the original inertial weight formula. The first-scale particle swarm inertial weights are obtained. It satisfies the following relationship: (23) S26: The first-scale subgroup is re-diffused, satisfying the following relationship: (24) After a certain number of iterations, perform a genetic operation on the optimal population of the individual particles at this point, setting the trigger condition as follows: Parent individuals are obtained using an elite retention and random selection strategy, and offspring are generated through a simulated binary crossover mechanism, satisfying the following relationship: (25) In the formula, It is a parameter that controls the crossover strength and depends on random numbers. and distribution index , and They are two individuals drawn from the parent generation; The binomial mutation method is used to change certain dimensional values ​​of offspring individuals. The particle mutation formula is as follows: (26) In the formula This is the offset. It is a uniformly distributed random number. It is a distribution index that controls the magnitude of variation; S27: Set the adaptive sliding window transfer entropy as the fitness function. Evaluate the current particle based on the fitness function, update the optimal positions of individuals in each first-scale and second-scale subgroup, randomly shuffle and recombine the particle swarm currently searching, change the neighborhood structure of the particles, cause the particles that were originally exploring at the first scale to begin a refined search, and the particles that were originally searching at the refined scale to begin a first-scale jump to explore local solutions, and finally find the globally optimal particle position. The optimal time base sequence is used to identify the operating conditions of industrial furnaces and kilns, and the optimal time delay sequence is determined based on the optimal time base sequence.

2. The industrial furnace / kiln condition identification method according to claim 1, characterized in that, S3 includes: S31: Treat industrial furnace and kiln condition identification as a multi-classification problem and construct a nonlinear model. It satisfies the following relationship: (27)； In the formula For time, for Input samples of time, The output is the industrial furnace / kiln operating condition identification value; The categories for recording the operating conditions of industrial furnaces and kilns include: species, that is ,in Represents this time This is normal operating condition; S32: Assume the industrial furnace data after time-series registration is as follows: ； in, yes Time-based data samples This represents the total number of samples; the cluster centers for different working conditions are denoted as the query samples. : (28) In the formula, Number of process variables, input sample Euclidean distance between the query sample and the query sample The following relationship must be satisfied: Obtain Euclidean distance Then, the sample contribution factor was set to : (30) In the formula, It refers to adjusting parameters; The first Correlation coefficients between process variables and target output Defined as: (31) In the formula , This represents the optimal recommended value for the process variables. The larger the correlation coefficient, the higher the contribution weight of the corresponding process variable. for: (32) S33: After obtaining the contribution factor and contribution weight, the input layer is reconstructed through encoding and decoding processes, thereby designing a new reconstruction loss function. To train the recognition model: (33) S34: Utilize hidden layer features to calculate the recognition error term for labeled data in the input, and construct a new pre-trained loss function based on the reconstruction loss function. Among them, the hidden layer features of the denoising autoencoder model are: ,Will enter The layer is identified and output. This leads to the pre-training loss function. : (34) In the formula, To input the total number of samples, The number of labeled samples in the input. It is a tradeoff coefficient used to balance the reconstruction loss term and the recognition error term in the loss function, and is used to minimize the loss function during training. Obtain the model parameter set ; S35: Will A series of denoising autoencoder models are stacked layer by layer to form a deep neural network. The hidden layer features of the previous denoising autoencoder model are used as the input layer data of the next denoising autoencoder model to obtain the th... The hidden layer features of the denoising autoencoder model are: ,at the same time Also the first Inputs to a DAE; Calculate the first The two-dimensional weights of the input data of the denoising autoencoder model are obtained. and Then the first Loss function of a denoising autoencoder model for: (35) In the formula It is the dimension value of the current DAE input layer; S36: After the DWTD-SDAE model is pre-trained, Add a layer to the top of the model to fine-tune the model parameter set in reverse. , No. Hidden layer features of a DAE pass The output layer obtains the working condition identification results. : (36) In the formula, and for The activation functions and network parameters of the layers are used to minimize the target output loss function using the backpropagation (BP) algorithm. The optimal network parameters of the DWTD-SDAE model are obtained. ;Target output loss function Represented as: (37) In equation (37) This represents the number of samples used for model fine-tuning. It is One-hot encoded vector of dimension, It is A probability vector of dimension 1.

3. An industrial furnace / kiln condition identification system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method described in any one of claims 1 to 2.

Images