Intelligent monitoring method for convection heating surface in biomass bulk material combustion process

Abstract

The invention relates to the technical field of boilers, in particular to an intelligent monitoring method for a convection heating surface in a biomass bulk material combustion process, which comprises the following steps of: acquiring temperature and pressure of each preset monitoring position of the convection heating surface to form a real-time sample, and performing current classification on all real-time samples by utilizing a target support vector machine model; obtaining real-time samples on the interval boundary and in the interval boundary, and if the target support vector machine model needs model parameter updating, obtaining an adjustment deviation judgment coefficient and an adjustment coefficient of a penalty parameter according to the real-time samples on the interval boundary and in the interval boundary to obtain a new penalty parameter; according to the new penalty parameters, model parameter updating is carried out on the target support vector machine model, a final classification result of final classification of all real-time samples is obtained, according to the final classification result, the ash deposition state on the convection heating surface is monitored, and the model parameters are dynamically adjusted so as to be better suitable for real-time sample classification.

Description

Technical Field

[0001] The present invention relates to the technical field of boilers, and in particular to an intelligent monitoring method for convection heating surfaces in a biomass bulk material combustion process. Background Art

[0002] During the combustion of bulk biomass, the convection heating surface is a crucial component of the boiler, and its performance directly impacts the boiler's thermal efficiency and emission quality. Biomass fuels (such as wood chips, straw, and rice husks) are characterized by high volatility, low ash melting points, and high alkali metal content. Combustion easily produces solid fly ash particles and other deposits, and ash deposits adhere to the convection heating surface. The thickness of ash deposits on the convection heating surface can affect boiler performance, corrode the surface, shorten equipment life, and ensure the safe operation of coal-fired units. Therefore, monitoring ash accumulation on the convection heating surface is crucial to promptly identify and address ash accumulation issues and ensure boiler performance.

[0003] In existing technology, a support vector machine model is used to analyze the characteristics of soot accumulation on the convection heating surface using monitoring data (such as temperature and pressure) to determine whether soot blowing is required, thereby performing soot blowing operations. Due to the hysteresis of the support vector machine model, a trained support vector machine model will always use fixed model parameters for model classification. However, when biomass fuel is changed, sudden soot accumulation may occur, causing deviations in the model classification results and leading to incorrect soot blowing operations.

[0004] Therefore, how to enhance the robustness and generalization of the support vector machine model to ensure the accuracy of dust accumulation feature analysis on convection heating surfaces has become an urgent problem to be solved. Summary of the Invention

[0005] In view of this, an embodiment of the present invention provides an intelligent monitoring method for the convective heating surface of a biomass bulk combustion process to solve the problem of how to enhance the robustness and generalization of a support vector machine model to ensure the accuracy of the ash accumulation feature analysis of the convective heating surface.

[0006] An embodiment of the present invention provides a method for intelligently monitoring a convective heating surface during a biomass bulk combustion process, the method comprising the following steps: Acquire the temperature and pressure of each preset monitoring position of the convection heating surface to form a real-time sample, and use the target support vector machine model to perform current classification on all real-time samples to obtain support vectors, real-time samples on the interval boundary, and real-time samples within the interval boundary. The target support vector machine model refers to the support vector machine model after the last final classification of the current classification. Determining whether the target support vector machine model requires model parameter update based on the support vector, and if the target support vector machine model requires model parameter update, obtaining an adjustment bias determination coefficient for a penalty parameter based on a probability that each of the real-time samples is noise and real-time samples on and within the interval boundary; Obtaining an adjustment coefficient for a penalty parameter based on a position difference between a sample closest to the decision boundary within the interval boundary of the current classification and a sample closest to the decision boundary within the interval boundary of the previous final classification, and obtaining a new penalty parameter based on the adjustment bias determination coefficient and the adjustment coefficient; Based on all real-time samples and new penalty parameters, the model parameters of the target support vector machine model are updated to obtain an updated support vector machine model. The updated support vector machine model is used to perform final classification on all real-time samples to obtain a final classification result. The dust accumulation status on the convection heating surface is monitored based on the final classification result.

[0007] Preferably, the support vector includes normal samples and dust accumulation samples, and the step of determining whether the target support vector machine model needs to update model parameters based on the support vector includes: Obtain the ratio values and quantity differences of normal samples and dust samples in the support vector of the current classification, as well as the ratio values and quantity differences of normal samples and dust samples in the support vectors of a preset number of historical classifications before the current classification, to form a ratio value sequence and a quantity difference sequence; Performing first-order difference processing on the proportion value sequence and the quantity difference value sequence respectively to obtain a proportion difference value sequence and a quantity difference value sequence respectively, calculating the mean of all elements except the last element in the proportion difference value sequence, and recording it as the proportion difference value threshold; calculating the mean of all elements except the last element in the quantity difference value sequence, and recording it as the quantity difference value threshold; If the last element in the proportion difference value sequence is greater than the proportion difference value threshold, and the last element in the quantity difference value sequence is greater than the quantity difference value threshold, it is determined that the target support vector machine model requires model parameter update.

[0008] Preferably, obtaining the adjustment bias determination coefficient of the penalty parameter based on the probability that each real-time sample belongs to noise and the real-time samples on and within the interval boundary includes: Based on the preset number of historical samples during historical classification, obtaining a probability that each of the real-time samples belongs to noise, marking the noise samples according to the probability that each of the real-time samples belongs to noise, and counting the number of noise samples in the real-time samples on and within the interval boundary; For any real-time sample on or within the interval boundary, obtain the Euclidean distance between the real-time sample and each real-time sample on or within the interval boundary, and obtain a mean Euclidean distance corresponding to the real-time sample; calculate an average of the mean Euclidean distances based on the mean Euclidean distances corresponding to each real-time sample on or within the interval boundary, and record the difference between a constant 1 and the reciprocal of the mean as the sample dispersion; The proportion of the number of noise samples in the total number of real-time samples on and within the interval boundary is calculated, and the adjustment bias determination coefficient of the penalty parameter is obtained according to the proportion and the mean of the sample dispersion.

[0009] Preferably, obtaining the probability that each real-time sample belongs to noise based on the preset number of historical samples during historical classification includes: For any historical sample in any historical classification, obtain the straight-line distance from the historical sample to the decision boundary of the historical classification, and use the inverse of the straight-line distance as the influence coefficient of the historical sample; obtain the influence coefficient of each support vector in the historical classification, and use the minimum influence coefficient as the influence coefficient threshold. If the influence coefficient of any historical sample is greater than or equal to the influence coefficient threshold, use the historical sample as the target historical sample; Obtain the target historical samples for the preset number of historical classifications, perform density clustering on all target historical samples and all real-time samples to obtain at least one cluster center, obtain the minimum Euclidean distance for any real-time sample based on the Euclidean distance between any real-time sample and each of the cluster centers, construct a circular area with the any real-time sample as the center and a preset multiple of the minimum Euclidean distance as the radius, and count the number of samples within the circular area; Obtain a difference between the reciprocal of the minimum Euclidean distance and a constant 1, obtain the reciprocal of the sum of the number of samples and the constant 1, perform weighted summation on the difference and the reciprocal of the sum, and obtain a probability that any real-time sample belongs to noise.

[0010] Preferably, the adjustment coefficient of the penalty parameter is obtained based on the position difference between the sample closest to the decision boundary within the interval boundary of the current classification and the sample closest to the decision boundary within the interval boundary of the last final classification, including: In the interval boundary of the current classification, calculate the sample closest to the decision boundary in the upper interval and the sample closest to the decision boundary in the lower interval, which are recorded as upper samples and lower samples respectively, and obtain the minimum Euclidean distance between the upper samples and the lower samples; Obtain the upper sample and the lower sample closest to the decision boundary within the interval boundary of the last final classification to obtain the minimum reference Euclidean distance, calculate the reciprocal of the absolute value of the difference between the minimum reference Euclidean distance and the minimum Euclidean distance, and use the difference between the constant 1 and the reciprocal as the adjustment coefficient of the penalty parameter.

[0011] Preferably, obtaining a new penalty parameter according to the adjustment bias determination coefficient and the adjustment coefficient includes: Obtaining a penalty parameter of the target support vector machine model, obtaining the product of the penalty parameter and the adjustment coefficient, and recording it as an adjustment value; obtaining a vertical distance between the interval boundary of the current classification and the decision boundary, calculating the difference between a constant 1 and the reciprocal of the vertical distance, and recording the sum of a preset value and the difference as an adjustment bias determination coefficient threshold; If the adjustment bias determination coefficient is greater than or equal to the adjustment bias determination coefficient threshold, the difference between the penalty parameter of the target support vector machine model and the adjustment value is used as the new penalty parameter; if the adjustment bias determination coefficient is less than the adjustment bias determination coefficient threshold, the sum of the penalty parameter of the target support vector machine model and the adjustment value is used as the new penalty parameter.

[0012] Preferably, updating the model parameters of the target support vector machine model according to all real-time samples and the new penalty parameters to obtain an updated support vector machine model includes: Cluster all real-time samples to obtain at least one cluster. For any cluster, calculate the distance between the cluster center of the cluster and each target historical sample. Obtain the target historical sample corresponding to the minimum distance as the reference sample. Based on the classification result of the reference sample, obtain the sample classification label of each real-time sample in the cluster. Obtain the sample classification label of each of the real-time samples, replace the penalty parameters in the target support vector machine model with the new penalty parameters, and then update other model parameters in the target support vector machine model based on each of the real-time samples and the corresponding sample classification label to obtain an updated support vector machine model.

[0013] Preferably, monitoring the dust accumulation state on the convection heating surface according to the final classification result includes: The final classification result includes normal samples and dust accumulation samples. The distance between each dust accumulation sample and the decision boundary of the final classification is calculated to obtain a distance mean. The real-time dust accumulation degree is obtained based on the difference between the reciprocal of the distance mean and a constant 1. The number of preset monitoring locations involved in the dust accumulation samples is obtained, and the proportion of the number in the total number of preset monitoring locations is obtained, which is recorded as the dust accumulation range. If the dust accumulation range is smaller than the preset dust accumulation range threshold, and the real-time dust accumulation level is greater than or equal to the preset dust accumulation level threshold, then only the local soot blowing operation is performed on the area where the preset monitoring position involved in the dust accumulation sample is located; If the dust accumulation range is greater than or equal to a preset dust accumulation range threshold, and the real-time dust accumulation level is greater than or equal to a preset dust accumulation level threshold, an overall soot blowing operation is performed on the convection heating surface.

[0014] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: The present invention obtains the temperature and pressure of each preset monitoring position of the convection heating surface to form a real-time sample, uses the target support vector machine model to perform current classification on all real-time samples, and obtains support vectors, real-time samples on and within the interval boundary, and the target support vector machine model refers to the support vector machine model after the last final classification of the current classification; according to the support vector, it is judged whether the target support vector machine model needs to update the model parameters. If the target support vector machine model needs to update the model parameters, then according to the probability that each real-time sample belongs to noise, and the real-time samples on and within the interval boundary, the penalty parameter is obtained. Adjust the bias determination coefficient; obtain the penalty parameter adjustment coefficient based on the position difference between the sample closest to the decision boundary within the interval boundary of the current classification and the sample closest to the decision boundary within the interval boundary of the previous final classification, and obtain a new penalty parameter based on the adjusted bias determination coefficient and the adjustment coefficient; update the model parameters of the target support vector machine model based on all real-time samples and the new penalty parameter to obtain an updated support vector machine model, use the updated support vector machine model to perform a final classification on all real-time samples to obtain a final classification result, and monitor the dust accumulation status on the convection heating surface based on the final classification result. In this case, based on the previous classification result of the support vector machine model and combined with the result of the current classification of the support vector machine model, the model parameters of the support vector machine model are dynamically adjusted and updated to obtain the updated support vector machine model, thereby improving the robustness and generalization ability of the support vector machine model, so as to better adapt to the classification of real-time samples and ensure the accuracy of dust accumulation judgment based on the real-time classification result. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, 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 only 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.

[0016] Figure 1This is a flow chart of a method for intelligently monitoring a convection heating surface in a biomass bulk combustion process provided by the first embodiment of the present invention. DETAILED DESCRIPTION

[0017] The embodiments of the present disclosure are described in detail below, and examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to be used to explain the present disclosure, but should not be understood as limiting the present disclosure.

[0018] It should be noted that the terms "first," "second," and the like in the specification of the present disclosure and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure.

[0019] In order to illustrate the technical solution of the present invention, specific embodiments are provided below.

[0020] See also Figure 1 , is a method flow chart of a method for intelligently monitoring the convection heating surface of a biomass bulk combustion process provided by the first embodiment of the present invention, such as Figure 1 As shown, the method may include: Step S101, obtain the temperature and pressure of each preset monitoring position of the convection heating surface to form a real-time sample, use the target support vector machine model to perform current classification on all real-time samples, and obtain support vectors, real-time samples on and within the interval boundary. The target support vector machine model refers to the support vector machine model after the last final classification of the current classification.

[0021] The purpose of the present invention is to better monitor the ash accumulation on the convection heating surface of the boiler in real time by constructing a support vector machine model to monitor the ash accumulation level in real time. However, fixed model parameters cannot adapt to all scenarios, such as the replacement of biomass fuel. Therefore, in order to improve the robustness and generalization ability of the support vector machine model, the model parameters of the support vector machine model are adjusted and updated in real time.

[0022] Since dust accumulation on the convection heating surface occurs gradually as the boiler operates, and there is no dust accumulation at the beginning, and since dust accumulation causes the convection heating surface to absorb less heat, the cross-sectional area of the working fluid (such as water, steam, or air) is reduced, which reduces the flow rate of the working fluid on the convection heating surface and increases the flow resistance. According to the principles of fluid dynamics, increased flow resistance will inevitably lead to increased pressure. Therefore, in an embodiment of the present invention, after the boiler has been running for a period of time, temperature data of multiple sub-areas of the convection heating surface and pressure data obtained by the pressure sensor are obtained through thermocouples, and the temperature data and pressure data of each sub-area are combined into a sample, with each sub-area corresponding to one sample. The sampling frequency of the temperature data and the pressure data is consistent, and the present invention sets the sampling frequency to once per minute.

[0023] It should be noted that the number of sub-areas of the convective heating surface is related to the number of sensors set. Assuming that 10 sensors are set, the convective heating surface is divided into 10 sub-areas. One sub-area represents a preset monitoring position. Correspondingly, 10 samples can be obtained at each sampling moment to construct the initial support vector machine model.

[0024] Specifically, the embodiment of the present invention sets samples within 10 minutes to form a training set for constructing an initial support vector machine model, wherein the construction process of the initial support vector machine model is: obtaining the theoretical temperature threshold and theoretical pressure threshold of the convective heating surface when the known combustible material is burned. Since a sample has two properties of temperature and pressure, and ash accumulation causes the convective heating surface to absorb less heat and increase pressure, for any sample in the training set, if the temperature in the sample is greater than or equal to the theoretical temperature threshold, and the pressure is less than or equal to the theoretical pressure threshold, then the sample is regarded as a normal sample, that is, the normal sample has the characteristics of high temperature and low pressure. Samples in other cases are regarded as ash accumulation samples, and ash accumulation samples have the characteristics of low temperature and high pressure. In this way, the sample label setting is completed for each sample in the training set.

[0025] After the sample label of each sample in the training set is known, the support vector machine model is trained using the training set to obtain a trained support vector machine model, which is recorded as the initial support vector machine model. The training of the support vector machine model belongs to the existing technology and will not be described here. It should be noted that there is no restriction on the setting of the theoretical temperature threshold and the theoretical pressure threshold. It can be set according to the combustible material and the combustion scenario. For example, when the combustible material is straw, the temperature and pressure of the convection heating surface will be affected by many factors, such as the type of boiler, design parameters, operating conditions, etc. The following are approximate ranges for some common situations: (1) Temperature: Low temperature range. In some small or conservatively designed straw-fired boilers, the temperature of the convection heating surface may be around 200℃-400℃. For example, the convection heating surface of some small biomass hot water boilers is mainly used to heat hot water, and the operating temperature is relatively low. Medium temperature range. For medium-sized straw-fired boilers, the temperature of the convection heating surface is usually between 400℃-600℃. Such boilers may be used to supply a certain scale of industrial steam or heating. High temperature range. In large, high-efficiency straw-fired power generation boilers, the temperature of the convection heating surface may reach 600℃-800℃ or even higher. This is because the power generation boiler needs to generate high-temperature and high-pressure steam to drive the turbine to generate electricity, and the convection heating surface needs to absorb more heat to improve the steam parameters.

[0026] (2) Pressure: Low pressure range: For some small straw burning equipment, such as small steam generators or low-pressure hot water boilers, the pressure at the convection heating surface may be around 0.1MPa-1MPa; Medium pressure range: For medium-sized straw burning boilers used for industrial steam supply and other purposes, the pressure at the convection heating surface is generally between 1MPa-4MPa; High pressure range: In straw burning power generation boilers, the pressure at the convection heating surface is usually higher, which may reach 4MPa-10MPa or even higher to meet the steam pressure and temperature requirements for power generation.

[0027] After obtaining the initial support vector machine model, subsequent real-time samples collected are continuously input for real-time classification according to the initial support vector machine model. However, since different combustion materials generate different amounts of heat, when the combustion materials change, the distribution of the model classification results will change as follows: since the characteristics of the new samples may have similar characteristics to the original category, for example, the heat of wood is about 15-20 MJ / kg, and that of straw is about 14-17 MJ / kg, since the heat of straw is slightly lower than that of wood, the normal heat may be regarded as the heat of the wood in the ash-accumulated state when burning. In other words, the heat generated by different combustion materials is different. Therefore, the temperature obtained by normal combustion itself may cause the overall temperature data to be offset. The new sample may show a larger or smaller difference in the ash-accumulated sample. It may also cause the support vector originally near the decision boundary to no longer be a support vector after the addition of the new sample. At the same time, the sample change may also introduce noise or make some abnormal samples more prominent under the new sample. In this embodiment of the present invention, the model parameters of the initial support vector machine model can be adjusted and updated by collecting real-time samples, which is applicable to the scenario of replacing biomass fuel.

[0028] In an embodiment of the present invention, a real-time sample is collected once every minute, and is used to update the model parameters at different times based on the initial support vector machine model, that is, sample classification is performed once every minute. For example: when the initial support vector machine model is trained at the i-th sampling moment, after obtaining the real-time sample at the i+1-th sampling moment, the classification result of the real-time sample by the initial support vector machine model is used to update the model parameters of the initial support vector machine model, and the updated support vector machine model is used as the target support vector machine model at the i+1-th sampling moment. After obtaining the real-time sample at the i+2-th sampling moment, the classification result of the real-time sample at the i+2-th sampling moment by the target support vector machine model at the i+1-th sampling moment is used to update the model parameters of the target support vector machine model at the i+1-th sampling moment, and the updated support vector machine model is used as the target support vector machine model at the i+2-th sampling moment. And so on, the support vector machine model at each sampling moment is adaptively updated and adjusted in model parameters.

[0029] Taking any sampling moment as an example, a specific analysis of the update and adjustment of model parameters is carried out: first, the real-time samples of each preset monitoring position of the convective heating surface at any sampling moment are obtained respectively. Then, all real-time samples are currently classified using the target support vector machine model to obtain support vectors, real-time samples on and within the interval boundary, where the support vectors include normal samples and dust accumulation samples. The target support vector machine model refers to the support vector machine model after the last final classification (the last sampling moment of any sampling moment) of the current classification (that is, any sampling moment).

[0030] Step S102, judging whether the target support vector machine model needs model parameter update based on the support vector, if the target support vector machine model needs model parameter update, then obtaining the adjustment bias determination coefficient of the penalty parameter based on the probability that each real-time sample belongs to noise, and the real-time samples on and within the interval boundary.

[0031] As the dust accumulation level of dust samples increases, their sample points will move further and further away from the decision boundary during the classification process. Therefore, the support vectors of dust samples will most likely remain unchanged, while the support vectors of normal samples may change slightly. Therefore, a model parameter update judgment mechanism is set to determine whether the target support vector machine model needs to be updated. The model parameter update judgment mechanism is as follows: Obtain the ratio values and quantity differences of normal samples and dust samples in the support vector of the current classification, as well as the ratio values and quantity differences of normal samples and dust samples in the support vectors of at least three historical classifications before the current classification, to form a ratio value sequence and a quantity difference sequence; Performing first-order difference processing on the proportion value sequence and the quantity difference value sequence respectively to obtain a proportion difference value sequence and a quantity difference value sequence respectively, calculating the mean of all elements except the last element in the proportion difference value sequence, and recording it as the proportion difference value threshold; calculating the mean of all elements except the last element in the quantity difference value sequence, and recording it as the quantity difference value threshold; If the last element in the proportional difference value sequence is greater than the proportional difference value threshold, and the last element in the quantity difference value sequence is greater than the quantity difference value threshold, it is determined that the target support vector machine model requires model parameter update; otherwise, it is considered that the target support vector machine model does not need model parameter update, and the current classification result is also the final classification result of the real-time sample.

[0032] After determining that the target support vector machine model requires model parameter update, considering that the penalty parameter of the support vector machine model is set manually, while the remaining model parameters are obtained based on sample training, therefore, in an embodiment of the present invention, the penalty parameter in the target support vector machine model is updated first, and the remaining model parameters are updated and obtained according to the normal training method.

[0033] A smaller penalty parameter results in a more tolerant model, i.e., a wider buffer zone (interval boundary), while a larger penalty parameter results in a less tolerant model, i.e., a narrower buffer zone. When sample characteristics change, it is necessary to analyze the sample distribution, particularly historical and real-time samples that fall within the interval boundary before the penalty parameter update, to determine the penalty parameter adjustment bias coefficient for the current classification.

[0034] Since each time a new sample (that is, a real-time sample) is added for model classification, the number of samples is too large, which easily increases the computational load. Therefore, after each classification, it is necessary to calculate the influence coefficient of each sample and the decision boundary, and only retain samples with a large influence coefficient in the model. Therefore, in an embodiment of the present invention, for any historical sample in any historical classification, the straight-line distance from the any historical sample to the decision boundary of the any historical classification is obtained, and the inverse of the straight-line distance is used as the influence coefficient of the any historical sample; the influence coefficient of each support vector in the any historical classification is obtained, and the minimum influence coefficient is used as the influence coefficient threshold. If the influence coefficient of any historical sample is greater than or equal to the influence coefficient threshold, then the any historical sample is used as the target historical sample, that is, the sample with a large influence coefficient.

[0035] Since real-time samples may contain noise, which is caused by electromagnetic equipment in industrial environments and can easily cause electromagnetic interference to sensors, and the convection heating surface of the boiler during operation may be subject to mechanical vibration or impact, affecting the stability of the sensor, causing sample changes or interference, thereby generating noise. For noise samples, choosing a smaller penalty parameter can avoid overfitting. Therefore, based on the characteristics of noise, such as abnormal location, small number, and strong randomness, we analyze whether there are noise samples in real-time samples. The probability of each real-time sample being noise is obtained as follows: Obtain the target historical samples of the three historical classifications before the current classification as a reference, perform density clustering on all target historical samples and all real-time samples, obtain at least one cluster center, and for any real-time sample, obtain the minimum Euclidean distance based on the Euclidean distance between any real-time sample and each cluster center. , taking any of the real-time samples as the center of the circle and the minimum Euclidean distance of a preset multiple as the radius of the circle, constructing a circular area, and counting the number of samples in the circular area; Obtain a difference between the reciprocal of the minimum Euclidean distance and a constant 1, obtain the reciprocal of the sum of the number of samples and the constant 1, perform weighted summation on the difference and the reciprocal of the sum, and obtain a probability that any real-time sample belongs to noise.

[0036] In one embodiment, in order to prevent interference caused by samples near the cluster center, the circle radius cannot be the complete minimum Euclidean distance. At the same time, considering that the density of the clusters may not be too dense, the distance cannot be too close, so the preset multiple is set to , that is, the radius of the circle is , then the calculation formula for the probability that any real-time sample belongs to noise is:

[0037] in, represents the probability that any real-time sample belongs to noise, represents the first weight, 1 represents a constant, represents the minimum Euclidean distance, represents the second weight, Indicates the number of samples within the circular area.

[0038] It should be noted that the greater the distance between any real-time sample and the nearest cluster center, that is, the greater the minimum Euclidean distance, the greater the probability that any real-time sample is an isolated point and the greater the possibility of location anomaly; if the number of samples in the circular area of any real-time sample is smaller, it means that there are fewer similar samples around the real-time sample, which means that the probability that any real-time sample is noise is greater.

[0039] Preferably, since the cluster size is not fixed, that is, it is still possible to select some cluster members with high density in the circular area, it is possible to set , there is no restriction here.

[0040] Furthermore, based on experimental statistics, a probability threshold is set to 0.7. If the probability of any real-time sample belonging to noise is greater than or equal to 0.7, the real-time sample is marked as a noise sample, thereby obtaining all noise samples in the real-time sample. After obtaining the marked noise samples, the real-time samples on and within the interval boundary of the current classification are analyzed to obtain the adjustment bias determination coefficient of the penalty parameter. The specific analysis is as follows: For any real-time sample on or within the interval boundary, obtain the Euclidean distance between the real-time sample and each real-time sample on or within the interval boundary, and obtain a mean Euclidean distance corresponding to the real-time sample; calculate an average of the mean Euclidean distances based on the mean Euclidean distances corresponding to each real-time sample on or within the interval boundary, and record the difference between a constant 1 and the reciprocal of the mean as the sample dispersion; The number of noise samples in the real-time samples on and within the interval boundary is counted, and the proportion of the number of noise samples in the total number of real-time samples on and within the interval boundary is calculated. According to the proportion and the mean of the discrete degree of the samples, the adjustment bias determination coefficient of the penalty parameter is obtained.

[0041] In one embodiment, the calculation formula of the adjustment bias determination coefficient of the penalty parameter is:

[0042] in, represents the adjustment bias determination coefficient of the penalty parameter, Indicates the total number of real-time samples on and within the interval boundary, represents the number of noise samples in the real-time samples on and within the interval boundaries, It represents the Euclidean distance between the jth real-time sample and the ith real-time sample on and within the interval boundary, and 1 represents a constant.

[0043] It should be noted that Characterizes the proportion of quantity. The larger the proportion of quantity, the larger the noise sample, the greater the adjustment bias of the penalty parameter should be, and the larger the corresponding adjustment bias determination coefficient; It is used to characterize the degree of discreteness of real-time samples. The greater the degree of discreteness, the greater the adjustment bias of the penalty parameter should be. The larger the corresponding adjustment bias determination coefficient is, the smaller the penalty parameter should be.

[0044] Step S103, obtain the adjustment coefficient of the penalty parameter based on the position difference between the sample closest to the decision boundary within the interval boundary of the current classification and the sample closest to the decision boundary within the interval boundary of the last final classification, and obtain the new penalty parameter based on the adjusted bias determination coefficient and the adjustment coefficient.

[0045] Since it is necessary to separate samples of different categories as much as possible during classification, the penalty parameter can be adjusted according to the distribution difference between real-time samples. In the interval boundary of the current classification, the sample closest to the decision boundary in the upper interval and the sample closest to the decision boundary in the lower interval are calculated, which are recorded as upper samples and lower samples respectively, and the minimum Euclidean distance between the upper and lower samples is obtained; According to the method for obtaining the minimum Euclidean distance corresponding to the interval boundary of the current classification mentioned above, the minimum Euclidean distance between the upper sample and the lower sample closest to the decision boundary within the interval boundary of the previous final classification is obtained, recorded as the minimum reference Euclidean distance, and the reciprocal of the absolute value of the difference between the minimum reference Euclidean distance and the minimum Euclidean distance is calculated, and the difference between the constant 1 and the reciprocal is used as the adjustment coefficient of the penalty parameter.

[0046] In one embodiment, the calculation formula of the adjustment coefficient of the penalty parameter is:

[0047] in, represents the adjustment coefficient of the penalty parameter, Indicates the minimum Euclidean distance corresponding to the interval boundary of the current classification, Indicates the minimum reference Euclidean distance corresponding to the interval boundary of the last final classification, || indicates the absolute value sign, and 1 indicates a constant.

[0048] It should be noted that Characterizes the minimum Euclidean distance corresponding to the addition of real-time samples, Indicates the minimum Euclidean distance before adding real-time samples. If the difference between the minimum Euclidean distance before and after adding is large, it means that the penalty parameter is adjusted more, and the corresponding adjustment coefficient is larger.

[0049] Furthermore, after obtaining the adjustment bias determination coefficient and the adjustment coefficient of the penalty parameter, a new penalty parameter can be obtained according to the adjustment bias determination coefficient and the adjustment coefficient to replace the penalty parameter in the target support vector machine model. The method for obtaining the new penalty parameter is: First, obtain the penalty parameter of the target support vector machine model, obtain the product of the penalty parameter and the adjustment coefficient, and record it as the adjustment value. Then, refer to the calculation formula of the adjustment bias determination coefficient and set the adjustment bias determination coefficient threshold according to the rounding principle. The specific setting method is: obtain the vertical distance between the interval boundary of the current classification and the decision boundary, calculate the difference between the constant 1 and the inverse of the vertical distance, and record the sum of the preset value and the difference as the adjustment bias determination coefficient threshold; preferably, the adjustment bias determination coefficient threshold is ,in, Indicates vertical distance.

[0050] Finally, if the adjustment bias determination coefficient is greater than or equal to the adjustment bias determination coefficient threshold, the difference between the penalty parameter of the target support vector machine model and the adjustment value is used as the new penalty parameter; if the adjustment bias determination coefficient is less than the adjustment bias determination coefficient threshold, the sum of the penalty parameter of the target support vector machine model and the adjustment value is used as the new penalty parameter.

[0051] In one embodiment, the calculation formula of the new penalty parameter is:

[0052] in, represents the new penalty parameter, represents the penalty parameter of the target support vector machine model, represents the adjustment coefficient of the penalty parameter, represents the adjusted biased determination coefficient of the penalty parameter.

[0053] In step S104, the model parameters of the target support vector machine model are updated according to all real-time samples and the new penalty parameters to obtain an updated support vector machine model. The updated support vector machine model is used to perform a final classification on all real-time samples to obtain a final classification result. The dust accumulation state on the convection heating surface is monitored according to the final classification result.

[0054] After obtaining the new penalty parameters, the model parameters of the target support vector machine model are updated based on all real-time samples and the new penalty parameters. Since the characteristics of real-time samples may change, the classification criteria of the target support vector machine model may not be applicable to real-time samples. However, there are certain inherent differences between real-time samples. Therefore, all real-time samples are clustered to obtain at least one cluster. For any cluster, the distance between the cluster center of any cluster and each target historical sample is calculated. The target historical sample corresponding to the minimum distance is obtained and recorded as a reference sample. Based on the classification result of the reference sample, the sample classification label of each real-time sample in any cluster is obtained. Obtain the sample classification label of each of the real-time samples, replace the penalty parameters in the target support vector machine model with the new penalty parameters, and then update the other model parameters in the target support vector machine model based on each of the real-time samples and the corresponding sample classification labels to obtain an updated support vector machine model. That is, each real-time sample and the corresponding sample classification label are used as a training set, and the target support vector machine model is retrained to obtain a new support vector machine model. At this time, the model parameters in the new support vector machine model are the updated model parameters, and the new support vector machine model is also the updated support vector machine model.

[0055] After obtaining the updated support vector machine model, all real-time samples are finally classified using the updated support vector machine model to obtain the final classification results, which include normal samples and dust accumulation samples. The dust accumulation status on the convection heating surface is then monitored based on the final classification results. The specific method is as follows: Calculate the distance between each dust accumulation sample and the decision boundary of the final classification to obtain the mean distance. Calculate the real-time dust accumulation level based on the difference between the inverse of the mean distance and a constant of 1. Obtain the number of preset monitoring locations involved in the dust accumulation samples and calculate the proportion of the number in the total number of preset monitoring locations, which is recorded as the dust accumulation range. If the dust accumulation range is smaller than the preset dust accumulation range threshold, and the real-time dust accumulation level is greater than or equal to the preset dust accumulation level threshold, then only the local soot blowing operation is performed on the area where the preset monitoring position involved in the dust accumulation sample is located; If the dust accumulation range is greater than or equal to a preset dust accumulation range threshold, and the real-time dust accumulation level is greater than or equal to a preset dust accumulation level threshold, an overall soot blowing operation is performed on the convection heating surface.

[0056] In one embodiment, the calculation formula for the real-time dust accumulation level is:

[0057] in, Indicates the real-time dust accumulation level. Represents the distance mean.

[0058] It should be noted that the larger the distance mean, the farther the dust accumulation sample is from the interval boundary, which means more dust accumulation, that is, the greater the degree of dust accumulation.

[0059] The calculation formula for the dust accumulation range is:

[0060] in, Indicates the dust accumulation range, that is, the proportion of the number of preset monitoring locations involved in the dust accumulation samples in the total number of preset monitoring locations. Indicates the total number of preset monitoring locations, Indicates the number of preset monitoring locations involved in the dust accumulation sample.

[0061] It should be noted that the greater the number of preset monitoring locations involved in the dust accumulation samples, the larger the dust accumulation range.

[0062] Since the range of the real-time dust accumulation degree and dust accumulation range is [0, 1], the median value is taken as the dividing line between the local dust blowing operation and the overall dust blowing operation, and the dust accumulation range threshold and dust accumulation degree threshold are set as , there is no restriction here.

[0063] like , it means that the dust accumulation may be serious in some areas and needs to be repaired. Perform single soot blowing operation in a sub-area at a preset monitoring position (referring to soot blowing operation in one area of the boiler) to achieve local soot blowing operation. , then double blowing operation (referring to the soot cleaning operation of the entire boiler) is carried out to realize the overall soot blowing operation. In other cases, choose to observe and do not carry out soot blowing operation.

[0064] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A method for intelligently monitoring the convection heating surface during the combustion of bulk biomass, characterized in that: The method comprises: Acquire the temperature and pressure of each preset monitoring position of the convection heating surface to form a real-time sample, and use the target support vector machine model to perform current classification on all real-time samples to obtain support vectors, real-time samples on the interval boundary, and real-time samples within the interval boundary. The target support vector machine model refers to the support vector machine model after the last final classification of the current classification. Determining whether the target support vector machine model requires model parameter update based on the support vector, and if the target support vector machine model requires model parameter update, obtaining an adjustment bias determination coefficient for a penalty parameter based on a probability that each of the real-time samples is noise and real-time samples on and within the interval boundary; Obtaining an adjustment coefficient for a penalty parameter based on a position difference between a sample closest to the decision boundary within the interval boundary of the current classification and a sample closest to the decision boundary within the interval boundary of the previous final classification, and obtaining a new penalty parameter based on the adjustment bias determination coefficient and the adjustment coefficient; Based on all real-time samples and new penalty parameters, the model parameters of the target support vector machine model are updated to obtain an updated support vector machine model. The updated support vector machine model is used to perform final classification on all real-time samples to obtain a final classification result. The dust accumulation status on the convection heating surface is monitored based on the final classification result.

2. The method for intelligently monitoring the convection heating surface of a biomass bulk combustion process according to claim 1, characterized in that: The support vector includes normal samples and dust accumulation samples, and the step of determining whether the target support vector machine model needs to update model parameters according to the support vector includes: Obtain the ratio values and quantity differences of normal samples and dust samples in the support vector of the current classification, as well as the ratio values and quantity differences of normal samples and dust samples in the support vectors of a preset number of historical classifications before the current classification, to form a ratio value sequence and a quantity difference sequence; Performing first-order difference processing on the proportion value sequence and the quantity difference value sequence respectively to obtain a proportion difference value sequence and a quantity difference value sequence respectively, calculating the mean of all elements except the last element in the proportion difference value sequence, and recording it as the proportion difference value threshold; calculating the mean of all elements except the last element in the quantity difference value sequence, and recording it as the quantity difference value threshold; If the last element in the proportion difference value sequence is greater than the proportion difference value threshold, and the last element in the quantity difference value sequence is greater than the quantity difference value threshold, it is determined that the target support vector machine model requires model parameter update.

3. The method for intelligently monitoring the convection heating surface of a biomass bulk combustion process according to claim 2, characterized in that: The step of obtaining an adjustment bias determination coefficient of a penalty parameter based on the probability that each real-time sample belongs to noise and the real-time samples on and within the interval boundary comprises: Based on the preset number of historical samples during historical classification, obtaining a probability that each of the real-time samples belongs to noise, marking the noise samples according to the probability that each of the real-time samples belongs to noise, and counting the number of noise samples in the real-time samples on and within the interval boundary; For any real-time sample on or within the interval boundary, obtain the Euclidean distance between the real-time sample and each real-time sample on or within the interval boundary, and obtain a mean Euclidean distance corresponding to the real-time sample; calculate an average of the mean Euclidean distances based on the mean Euclidean distances corresponding to each real-time sample on or within the interval boundary, and record the difference between a constant 1 and the reciprocal of the mean as the sample dispersion; The proportion of the number of noise samples in the total number of real-time samples on and within the interval boundary is calculated, and the adjustment bias determination coefficient of the penalty parameter is obtained according to the proportion and the mean of the sample dispersion.

4. The method for intelligently monitoring the convection heating surface of a biomass bulk combustion process according to claim 3, characterized in that: The obtaining, based on the preset number of historical samples during historical classification, a probability that each real-time sample belongs to noise includes: For any historical sample in any historical classification, obtain the straight-line distance from the historical sample to the decision boundary of the historical classification, and use the inverse of the straight-line distance as the influence coefficient of the historical sample; obtain the influence coefficient of each support vector in the historical classification, and use the minimum influence coefficient as the influence coefficient threshold. If the influence coefficient of any historical sample is greater than or equal to the influence coefficient threshold, use the historical sample as the target historical sample; Obtain the target historical samples for the preset number of historical classifications, perform density clustering on all target historical samples and all real-time samples to obtain at least one cluster center, obtain the minimum Euclidean distance for any real-time sample based on the Euclidean distance between any real-time sample and each of the cluster centers, construct a circular area with the any real-time sample as the center and a preset multiple of the minimum Euclidean distance as the radius, and count the number of samples within the circular area; Obtain a difference between the reciprocal of the minimum Euclidean distance and a constant 1, obtain the reciprocal of the sum of the number of samples and the constant 1, perform weighted summation on the difference and the reciprocal of the sum, and obtain a probability that any real-time sample belongs to noise.

5. The method for intelligently monitoring the convection heating surface of a biomass bulk combustion process according to claim 1, characterized in that: The adjustment coefficient of the penalty parameter is obtained based on the position difference between the sample closest to the decision boundary within the interval boundary of the current classification and the sample closest to the decision boundary within the interval boundary of the previous final classification, including: In the interval boundary of the current classification, calculate the sample closest to the decision boundary in the upper interval and the sample closest to the decision boundary in the lower interval, which are recorded as upper samples and lower samples respectively, and obtain the minimum Euclidean distance between the upper samples and the lower samples; Obtain the upper sample and the lower sample closest to the decision boundary within the interval boundary of the last final classification to obtain the minimum reference Euclidean distance, calculate the reciprocal of the absolute value of the difference between the minimum reference Euclidean distance and the minimum Euclidean distance, and use the difference between the constant 1 and the reciprocal as the adjustment coefficient of the penalty parameter.

6. The method for intelligently monitoring the convection heating surface during the combustion of bulk biomass materials according to claim 1, characterized in that: The obtaining of a new penalty parameter according to the adjustment bias determination coefficient and the adjustment coefficient includes: Obtaining a penalty parameter of the target support vector machine model, obtaining the product of the penalty parameter and the adjustment coefficient, and recording it as an adjustment value; obtaining a vertical distance between the interval boundary of the current classification and the decision boundary, calculating the difference between a constant 1 and the reciprocal of the vertical distance, and recording the sum of a preset value and the difference as an adjustment bias determination coefficient threshold; If the adjustment bias determination coefficient is greater than or equal to the adjustment bias determination coefficient threshold, the difference between the penalty parameter of the target support vector machine model and the adjustment value is used as the new penalty parameter; if the adjustment bias determination coefficient is less than the adjustment bias determination coefficient threshold, the sum of the penalty parameter of the target support vector machine model and the adjustment value is used as the new penalty parameter.

7. The method for intelligently monitoring the convection heating surface during the combustion of bulk biomass materials according to claim 4, characterized in that: The method of updating the model parameters of the target support vector machine model according to all real-time samples and the new penalty parameters to obtain an updated support vector machine model includes: Cluster all real-time samples to obtain at least one cluster. For any cluster, calculate the distance between the cluster center of the cluster and each target historical sample. Obtain the target historical sample corresponding to the minimum distance as the reference sample. Based on the classification result of the reference sample, obtain the sample classification label of each real-time sample in the cluster. Obtain the sample classification label of each of the real-time samples, replace the penalty parameters in the target support vector machine model with the new penalty parameters, and then update other model parameters in the target support vector machine model based on each of the real-time samples and the corresponding sample classification label to obtain an updated support vector machine model.

8. The method for intelligently monitoring the convection heating surface of a biomass bulk combustion process according to claim 1, characterized in that: The monitoring of the dust accumulation state on the convection heating surface according to the final classification result includes: The final classification result includes normal samples and dust accumulation samples. The distance between each dust accumulation sample and the decision boundary of the final classification is calculated to obtain a distance mean. The real-time dust accumulation degree is obtained based on the difference between the reciprocal of the distance mean and a constant 1. The number of preset monitoring locations involved in the dust accumulation samples is obtained, and the proportion of the number in the total number of preset monitoring locations is obtained, which is recorded as the dust accumulation range. If the dust accumulation range is smaller than the preset dust accumulation range threshold, and the real-time dust accumulation level is greater than or equal to the preset dust accumulation level threshold, then only the local soot blowing operation is performed on the area where the preset monitoring position involved in the dust accumulation sample is located; If the dust accumulation range is greater than or equal to a preset dust accumulation range threshold, and the real-time dust accumulation level is greater than or equal to a preset dust accumulation level threshold, an overall soot blowing operation is performed on the convection heating surface.

Images