Lightweight active online modeling method fusing mechanism and dynamic memory

Abstract

Disclosed in the present invention is a lightweight active online modeling method fusing mechanism and dynamic memory, which comprises the following steps: S1, data preprocessing and model initialization; S2, online data representativeness evaluation based on dynamic memory of process data; S3, online data informativeness evaluation based on dynamic memory of label data; S4, online monitoring of a mechanism state; S5, parallel hybrid active sampling policy; and S6, online modeling of a soft sensor model. The present invention separately designs memory policies for process data and label data and only reserves high-value data, thereby satisfying a lightweight hardware requirement; in addition, the present invention designs a time delay-based mechanism state online monitoring method to detect a real concept drift of the process, and further provides the parallel hybrid active sampling policy for detecting both a real concept drift and a false concept drift; in combination with dynamic memory and a kernel incremental regression method, the computation amount for model updating and the hardware overhead for edge deployment are reduced.

Description

A lightweight active online modeling method integrating mechanism and dynamic memory Technical Field

[0001] The present invention relates to the technical field of industrial soft measurement, and in particular to a lightweight active online modeling method integrating mechanism and dynamic memory. Background Art

[0002] Soft sensing technology has rapidly developed and has been widely applied in fields such as environmental protection, industry, and agriculture. In particular, it has been extensively researched in the perception of complex and difficult-to-measure data in industrial processes. These methods have provided effective data supplements for industries such as coal preparation, metallurgy, and petrochemicals. However, with the practical application of soft sensing technology in industrial sites, new challenges have gradually emerged. Influenced by factors such as raw material properties, equipment consumption, and operating conditions, data from actual industrial operations often exhibits dynamic and time-varying characteristics. This makes it difficult for online data to meet the independent and identically distributed (IID) requirements of soft sensing models, leading to inaccuracy in soft sensing models. Furthermore, current soft sensing technologies do not fully consider the limited storage and computing power in real-world industrial systems. This is particularly true in the context of Industry 4.0, as industrial systems transition towards end-cloud-edge collaboration. The question of whether models can be deployed at the edge is crucial.

[0003] Summary of the Invention

[0004] In order to overcome the deficiencies of the above-mentioned prior art, the present invention provides a lightweight active online modeling method that integrates mechanism and dynamic memory. The purpose is to: propose a dynamic memory method for online data to solve the problem of how to selectively cache infinite data in the data stream; propose an online monitoring method for mechanism status that can discover the true concept drift of the process; propose a parallel hybrid active sampling strategy to simultaneously detect true concept drift and false concept drift to achieve high-value sampling of online data; combine dynamic memory with incremental kernel regression method to propose a lightweight online soft measurement model to reduce the computational overhead of model updates. In order to achieve the above-mentioned technical objectives, the present invention adopts the following technical solutions:

[0005] A lightweight active online modeling method integrating mechanism and dynamic memory includes the following steps:

[0006] S1: Data preprocessing and model initialization:

[0007] S2: Online data representativeness evaluation based on dynamic memory of process data;

[0008] S3: Online data informativeness evaluation based on dynamic memory of labeled data;

[0009] S4: online monitoring of mechanism status;

[0010] S5: Parallel hybrid active sampling strategy;

[0011] S6: Online modeling of soft sensor model.

[0012] Compared with the prior art, the present invention has the following beneficial effects:

[0013] This paper considers the dynamic, time-varying characteristics of industrial processes, as well as industrial field sampling and hardware conditions, and proposes a lightweight, active online modeling method that integrates mechanisms and dynamic memory. By designing memory strategies for both process data and label data, only high-value data is retained, requiring only lightweight data storage and reducing hardware requirements. A time-delayed online monitoring method for mechanism states is designed to detect true concept drift in the process. Furthermore, a parallel hybrid active sampling strategy is proposed to simultaneously detect both true and false concept drift and sample high-value data. Combining dynamic memory with kernel incremental regression, a lightweight online soft-sensing model is proposed, reducing the computational complexity of model updates and the hardware overhead of edge deployment. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0015] FIG1 is a flow chart of the method of the present invention.

[0016] FIG2 is a process flow chart of a typical heavy medium coal preparation process according to an embodiment.

[0017] Figure 3 shows the results of soft sensor modeling. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] As shown in Figure 1, a lightweight active online modeling method that integrates mechanism and dynamic memory includes the following steps:

[0020] S1: Data preprocessing and model initialization:

[0021] Step S1 specifically includes:

[0022] S11: Perform mean filtering on the collected labeled data to obtain the initial data set L old ;

[0023] S12: Use After filtering, the initial data set L old Normalization is performed, where μ and σ represent the mean and variance of the data respectively, L new represents the data after normalization;

[0024] S13: Randomly extract some data without replacement from the normalized data set as the test set data, then randomly extract some data without replacement to simulate unlabeled data stream data, and the rest is used as the training set for initialization of the method.

[0025] S2: To achieve lightweight modeling, only fixed-budget process data is cached. To this end, Ebbinghaus's human memory law is introduced for dynamic memory of process data. Based on the distance and time decay relationship between online data and cached data, the cached data memory strength is efficiently updated. Data is updated based on the memory strength. Furthermore, the cached process data is used to calculate the local density of online data, enabling online data representativeness evaluation.

[0026] Step S2 specifically includes:

[0027] S21: The representativeness of online data is evaluated using process data in dynamic memory, based on the formula Calculate the local density ld(x c ), to achieve representative evaluation of online data; among them, mindis(x i ) represents x i The minimum distance to other data, dis(x c ,x i ) represents x c ,x i The distance between c Represents new data, x i Represents the data in the memory window, mw represents the memory window, and I is the discriminant function;

[0028] S22: Read the data stream online and design a cache strategy for process data based on the Ebbinghaus memory law. Memory strength is equivalent to the cache value of the data. Memory strength decays over time and is also affected by the enhanced memory reactivation of online data. When the remembered data is correlated with the online data and the minimum distance between the remembered data is greater than the distance to the new data, it is considered that there is effective memory between the data. The memory strength of the data with effective memory is enhanced. If there is no correlation, the memory strength of the remembered data decays exponentially over time. The decay factor is as follows:

[0029] f x represents the attenuation factor of x, λ x Represents the effective memory quantity of x; the memory strength decays over time, is inversely proportional to the decay factor, and is proportional to the memory frequency, as shown in the following formula:

[0030] s x represents the memory strength, t represents the current time, τ x represents the last valid memory time, and e represents the exponential function;

[0031] S23: When the window reaches the upper limit of data memory, the new data replaces the data with the minimum memory strength to realize dynamic memory update of data.

[0032] S3: Online Data Informativeness Evaluation Based on Dynamic Memory of Labeled Data: Similarly, to achieve lightweight modeling, only a fixed budget of labeled data is cached. The weights of cached data and online data are calculated based on data informativeness and forgetting factors. Labeled data is dynamically cached based on the data weights. The kernel method is then used to calculate the similarity between online highly representative data and the existing cached labeled data to achieve data informativeness evaluation.

[0033] Step S3 specifically includes:

[0034] S31: The informativeness of online data is evaluated based on the similarity between the feature spaces of online data and cached data. The calculation formula is as follows:

[0035] Among them, φ is the feature vector that maps the input d to a certain high-dimensional space, a represents the linear correlation coefficient, i represents the data index, and d i Indicates cache data, d i ∈D, D represents the data matrix, d m+1 represents online data, and m represents the number of data in the memory window. The larger δ is, the lower the similarity between the new data and the cached data is, and the more informative the data is.

[0036] S32: Determine a by minimizing δ. The minimization of δ is expressed as:

[0037] Among them, K = k (D, D) is the kernel matrix, k = k (D, d m+1 ) is the kernel vector, k=k(d m+1 , d m+1 ) is the kernel value, and the property of the kernel value represented by the inner product of the eigenvector is minimized to obtain the optimal coefficient vector α=K -1 k, and then we get: δ=k(d m+1 ,d m+1 )-k T a

[0038] Based on δ, decide whether to actively sample online data and add it to the cached data;

[0039] S33: When new tag data is added, the caching of the new data can be divided into two cases:

[0040] S331: If the tag data memory window has not reached the upper limit, the new data is directly added to the memory window, and the cached data δ is updated for subsequent memory evaluation. The calculation formula is as follows:

[0041] Among them, δ i Indicates the information degree of the i-th data, k(d i ,d i ) means d i own core value, Represents the linear correlation coefficient after adding new data, Indicates the linear correlation coefficient before adding new data, k i,m+1 Represents the core value of the i-th data and the new data, Represents the kernel vector before adding new data, Represents the kernel matrix before adding new data, k m+1 Represents the kernel vector between cached data and new data, k m+1 Indicates the newly added data's own verification value;

[0042] S332: If the tag data memory window has reached the upper limit, the data with the smallest δ in the memory data is deleted to free up memory for the new data. δ is calculated based on the following formula:

[0043] Among them, δ i represents the information degree of the i-th data, j represents the index of the replaced data, k m+1 Indicates the new data core value, row j [A * ] represents the j-th row of the given matrix A*;

[0044] S34: To simulate the decay of data information over time, a time-related forgetting factor is added to update δ. The calculation formula is as follows:

[0045] Among them, λ represents the forgetting factor, t represents the current time, and t irepresents the time when the data is added to the memory, h represents the forgetting strength; δ i =λ i δ i .

[0046] S4: Online monitoring of mechanism status: using the maximum relevant information to identify the time delay between variables, and then monitoring whether the mechanism changes based on the online time delay fluctuations;

[0047] Step S4 specifically includes:

[0048] S41: Analyze the process mechanism and take the process variable with the smallest time lag as the base variable. Take the continuous time series of l length of the base variable at time i in real time, and take the continuous time series of l length of the remaining process variables at time i.

[0049] S42: Calculate the maximum information coefficient between the remaining process sequence and the base sequence under different time lags. The calculation formula is as follows:

[0050] Where I(x, y) is the mutual information between variables x and y, x and y are two series of data, P(x) and P(y) are the probability density functions of x and y, P(x, y) is the joint probability density function, and B(n) is the 0.5 power of the total sample size n. The time delay corresponding to the lag sequence with the largest maximum information coefficient is the time delay between the remaining process variables and the base variable.

[0051] S43: The time-lag relationship between the process variable and the base variable is evaluated and monitored online. When a change in the time-lag relationship is detected, it is considered that the mechanism state has changed and the process has actually drifted.

[0052] S5: Parallel hybrid active sampling strategy: Parallel monitoring mechanism and data distribution changes, detecting true concept drift and false data drift, and realizing active sampling of online data samples:

[0053] Step S5 specifically includes:

[0054] S51: Real-time evaluation of online data representativeness indicators. Based on threshold judgment, highly representative data is selected to further evaluate the data informativeness. If the informativeness also meets the high informativeness characteristic, that is, when the cache is not full, the online data δ is greater than the preset threshold, and when the cache reaches the upper limit, the online data δ is greater than the minimum δ in the cache, then the active sampling mechanism is triggered;

[0055] S52: Monitor the process mechanism status in parallel and in real time. When the time lag change exceeds a preset threshold, it is considered that the mechanism has changed, and the active sampling mechanism is triggered.

[0056] S6: Online modeling of soft measurement models: Combining the Gaussian process regression soft measurement model with the incremental kernel regression method, incremental learning is used to obtain data from active sampling. Only some parameters need to be updated to achieve online modeling of the soft measurement model.

[0057] Step S6 specifically includes:

[0058] S61: The prediction formula of the original Gaussian process regression is as follows: y * =K(X * ,X)K(X,X) -1 y K(X * ,X * )=K(X * ,X)K(X,X) -1 K(X,X * )

[0059] Among them, X represents the process variable of the existing data, X* represents the process variable of the data to be predicted, y is the output of the existing data, y* is the output to be predicted, and K represents the covariance matrix;

[0060] S62: When the label data memory does not reach the upper limit, the covariance matrix The incremental update is calculated as follows:

[0061] The covariance matrix is ​​directly expanded, and the inverse matrix calculation formula of the expanded covariance matrix is ​​as follows:

[0062] Implement incremental learning of Gaussian process regression models that only require local parameter calculations;

[0063] S63: When new data is added, the label data memory has reached the upper limit, then the minimum δ data in the memory is replaced with the new data, and the covariance matrix The incremental update is calculated as follows:

[0064] j is the index of the replaced data, k m+1 Represents the kernel vector between cached data and new data, k m+1 Represents the kernel value of the newly added data itself. The inverse matrix of the new covariance matrix is ​​calculated based on the following formula:

[0065] Implementing incremental learning of Gaussian process regression models that only requires local parameter computation.

[0066] The heavy medium coal preparation process is a typical chemical production process and is currently one of the most direct and effective coal preparation methods. A typical heavy medium coal preparation process is shown in Figure 2. It primarily includes a coal-medium mixing drum, a heavy medium cyclone, a de-medium screen, a magnetic separator, an iron ore bin, a qualified medium drum, and several actuators and instruments. First, deslimed and dewatered raw coal is conveyed by conveyor to the coal-medium mixing drum for thorough mixing with the heavy medium, forming a mixed slurry that is then pumped into the heavy medium cyclone. Then, under the combined effects of gravity and centrifugal force, the mixed slurry in the heavy medium cyclone is separated into overflow slurry and underflow slurry. The slurry, which is lighter than the medium, floats at the top of the heavy medium cyclone, while the heavier slurry settles at the bottom. The overflow slurry and underflow slurry are then transported to the de-medium treatment process to form clean coal product and tailings waste, respectively. The remaining slurry is recycled. The medium to be recovered is filtered by a magnetic separator and sent to a qualified medium barrel, where it is mixed with magnetite and a certain amount of water to maintain the density of the heavy medium within a certain range, ensuring that the recovered heavy medium can be directly reused.

[0067] Ash content in coal preparation is an important process indicator for evaluating coal quality. It is mainly affected by coal mine feed rate, heavy medium density, feed pressure, etc. Based on this, a soft measurement model for heavy medium coal preparation is established.

[0068] 200 pieces of labeled data are collected for initialization of the labeled data memory window and soft measurement model. Then, the data stream is read online, and the data representativeness is evaluated based on the process data dynamic memory window. The informativeness of highly representative data is further calculated based on the labeled data dynamic memory window. At the same time, the mechanism state time delay is monitored online in parallel based on the maximum information coefficient. Finally, active sampling of high-value samples is achieved based on a parallel hybrid active sampling strategy, and the labeled memory window is updated. The soft measurement model is incrementally updated online to improve modeling accuracy.

[0069] The soft measurement modeling results are shown in FIG3 . It can be seen that the method of the present invention has a significantly faster decrease in modeling error, lower error, and higher efficiency compared to the traditional periodic sampling incremental modeling. The periodic sampling method requires a large number of additional data points to approach the effect of the method of the present invention.

Claims

1. A lightweight active online modeling method integrating mechanism and dynamic memory, characterized in that: The steps include: S1: data preprocessing and model initialization; S2: Online data representativeness evaluation based on dynamic memory of process data; S3: Evaluation of online data informativeness based on dynamic memory of labeled data; S4: online monitoring of mechanism status; S5: Parallel hybrid active sampling strategy; S6: Online modeling of soft sensor model.

2. The lightweight active online modeling method integrating mechanism and dynamic memory according to claim 1, characterized in that: Step S1 specifically includes: S11: Perform mean filtering on the collected labeled data to obtain the initial data set L old ; S12: Use After filtering, the initial data set L old Normalization is performed, where μ and σ represent the mean and variance of the data respectively, L new Represents the normalized data; S13: Randomly extract some data without replacement from the normalized data set as the test set data, and then randomly extract some data without replacement to simulate the unlabeled data stream data, and the rest is used as the training set for initializing the method.

3. The lightweight active online modeling method integrating mechanism and dynamic memory according to claim 1, characterized in that: Step S2 specifically includes: S21: The representativeness of online data is evaluated using process data in dynamic memory, based on the formula Calculate the local density ld(x) of the online data and the data in the memory window c ), to achieve representative evaluation of online data; among them, mindis(x i ) represents x i The minimum distance to other data, dis(x c ,x i ) represents x c ,x i The distance between c represents new data, x i represents the data in the memory window, mw represents the memory window, and I is the discriminant function; S22: Read the data stream online, design the cache data cache strategy based on the Ebbinghaus memory law, and equate the memory strength with the cache value of the data. The memory strength decays over time and is also affected by the online data's enhanced memory reactivation. When the memorized data is correlated with the online data, and the minimum distance between the memorized data is greater than the distance to the new data, it is considered that there is effective memory between the data, and the memory strength of the data with effective memory is enhanced. If there is no correlation, the memory strength of the memorized data decays exponentially over time, and the decay factor is as follows: f x represents the attenuation factor of x, λ x Represents the effective memory quantity of x; the memory strength decays over time, is inversely proportional to the decay factor, and is proportional to the memory frequency, and the formula is as follows: s x represents the memory strength, t represents the current time, τ x represents the last valid memory time, and e represents the exponential function; S23: When the window reaches the upper limit of data memory, new data replaces the data with the minimum memory strength to realize dynamic memory update of data.

4. The lightweight active online modeling method integrating mechanism and dynamic memory according to claim 1, characterized in that: Step S3 specifically includes: S31: The informativeness of online data is evaluated based on the similarity between the feature space of online data and cached data. The calculation formula is as follows: Among them, φ is the feature vector that maps the input d to a high-dimensional space, a represents the linear correlation coefficient, i represents the data index, and d i Indicates cache data, d i ∈D, D represents the data matrix, d m+1 represents online data, and m represents the number of data in the memory window; the larger δ is, the lower the similarity between the new data and the cached data is, and the more informative the data is; S32: a is determined by minimizing δ, and the minimization of δ is expressed as: Among them, K = k (D, D) is the kernel matrix, k = k (D, d m+1 ) is the kernel vector, k=k(d m+1 , d m+1 ) is the kernel value, and the property of the kernel value represented by the inner product of the eigenvector is minimized to obtain the optimal coefficient vector α=K -1 k, and then we get: δ=k(d m+1 ,d m+1 )-k T a Based on δ, decide whether to actively sample online data and add it to cache data; S33: When new tag data is added, the cache of the new data can be divided into two cases: S331: If the tag data memory window has not reached the upper limit, the newly added data is directly added to the memory window, and the cache data δ is updated for subsequent memory evaluation. The calculation formula is as follows: Among them, δ i represents the information degree of the i-th data, k(d i ,d i ) means d i own core value, Represents the linear correlation coefficient after adding new data, represents the linear correlation coefficient before adding new data, k i,m+1 Represents the core value of the i-th data and the new data, represents the kernel vector before adding new data, represents the kernel matrix before adding new data, k m+1 Represents the kernel vector between cached data and new data, k m+1 Indicates the value of the newly added data itself; S332: If the tag data memory window has reached the upper limit, the data with the smallest δ in the memory data is deleted to free up memory for the new data. δ is calculated based on the following formula: Among them, δ i represents the information degree of the i-th data, j represents the index of the replaced data, k m+1 Indicates the new data core value, row j [A * ] represents the j-th row of the given matrix A*; S34: To simulate the decay of data information over time, a time-related forgetting factor is added to update δ. The calculation formula is as follows: Among them, λ represents the forgetting factor, t represents the current time, and t i represents the time when the data is added to the memory, and h represents the forgetting intensity; d i =λ i d i 。 5. The lightweight active online modeling method integrating mechanism and dynamic memory according to claim 1, characterized in that: Step S4 specifically includes: S41: Analyze the process mechanism and take the process variable with the smaller time lag as the base variable, take the l-length continuous time series of the base variable at time i in real time, and take the l-length continuous time series of the remaining process variables at time i; S42: Calculate the maximum information coefficient between the remaining process sequence and the base sequence under different time lags. The calculation formula is as follows: Where I(x, y) is the mutual information value between variables x and y, x and y are two sequence data, P(x) and P(y) are the probability density functions of x and y, P(x, y) is the joint probability density function, B(n) is the 0.5 power of the total sample size n; the time delay corresponding to the time lag sequence with the largest maximum information coefficient is the time delay between the remaining process variables and the base variable; S43: The time-lag relationship between the process variable and the base variable is evaluated and monitored online. When a change in the time-lag relationship is detected, it is considered that the mechanism state has changed and a real drift has occurred in the process.

6. The lightweight active online modeling method integrating mechanism and dynamic memory according to claim 1, characterized in that: Step S5 specifically includes: S51: Real-time evaluation of online data representativeness indicators, based on threshold judgment, filter out highly representative data to further evaluate data informativeness. If the informativeness also meets the high informativeness feature, that is, when the cache is not full, the online data δ is greater than the preset threshold, and when the cache reaches the upper limit, the online data δ is greater than the minimum δ in the cache, then the active sampling mechanism is triggered; S52: Monitor the process mechanism status in parallel and in real time. When the time delay changes beyond a preset threshold, it is considered that the mechanism has changed, triggering an active sampling mechanism.

7. The lightweight active online modeling method integrating mechanism and dynamic memory according to claim 1, characterized in that: Step S6 specifically includes: S61: The prediction formula of the original Gaussian process regression is as follows: y * =K(X * ,X)K(X,X) -1 y K(X * ,X * )=K(X * ,X)K(X,X) -1 K(X,X * ) Among them, X represents the process variable with existing data, X* represents the process variable with data to be predicted, y represents the output with existing data, y* represents the output to be predicted, and K represents the covariance matrix; S62: When the label data memory does not reach the upper limit, the covariance matrix The incremental update is calculated as follows: The covariance matrix is ​​directly expanded, and the inverse matrix calculation formula of the expanded covariance matrix is ​​as follows: Implement incremental learning of Gaussian process regression models that only require local parameter calculations; S63: When new data is added, the label data memory has reached the upper limit, then the minimum delta data in the memory is replaced with the new data, and the covariance matrix The incremental update is calculated as follows: j is the index of the replaced data, k m+1 Represents the kernel vector between cached data and new data, k m+1 Represents the kernel value of the newly added data itself. The inverse matrix of the new covariance matrix is ​​calculated based on the following formula: Implementing incremental learning of Gaussian process regression models that only requires local parameter computation.

Images