Dust concentration measurement method based on coupling compensation and characteristic matching

Abstract

The application discloses a charge type dust concentration measuring method based on coupling compensation and characteristic matching, relates to the technical field of dust concentration measurement, and obtains an original signal matrix and an original environment matrix of each period through high-frequency sampling, carries out layered filtering on the original signal matrix, filters out abnormal windows by using the Leight criterion, and recombines a pure signal matrix through a wavelet threshold denoising algorithm, synchronously eliminates abnormal windows in the original environment matrix, and recombines a pure environment matrix; a heuristic neural network is used to quantize the influence of temperature and humidity on charge mobility, generate a correction signal matrix, calculate a signal characteristic vector, and then perform cosine matching on a concentration model library to select a concentration model, and calculate the average value of the dust concentration in the period to complete accurate measurement.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of dust concentration measurement, and in particular to a charge type dust concentration measurement method based on coupling compensation and characteristic matching. BACKGROUND

[0002] Dust concentration measurement in the pipeline is a core link of industrial production safety control and environmental quality monitoring, and the accuracy of the measurement result directly determines the rationality of production process adjustment, the safety of equipment operation and the compliance of environmental protection emission. If the dust concentration data is inaccurate, it may cause equipment wear or blockage due to excessive dust accumulation, so it is urgent to support industrial production and environmental management with reliable dust concentration measurement technology in the pipeline.

[0003] The existing dust measurement technology has the following defects: firstly, it does not quantify the influence of temperature and humidity on charge mobility based on charge induction principle, lacks coupling compensation mechanism of environmental factors into the measurement system, and causes distortion of charge signal when the environment fluctuates, so the measurement accuracy is disturbed; secondly, it does not build a characteristic matching system that adapts to different dust types, and cannot select the corresponding concentration model through accurate matching of signal characteristics and standard characteristics, so it is difficult to adapt to the differentiated characteristics of single or mixed dust, resulting in large deviation between the measurement result and the actual concentration. SUMMARY

[0004] The present application proposes a charge type dust concentration measurement method based on coupling compensation and characteristic matching to solve the problems in the prior art. Pure signal matrix is obtained by high-frequency sampling and hierarchical filtering, coupling compensation is realized by using heuristic neural network, and dust concentration is calculated by cosine matching, which effectively improves the measurement accuracy and adapts to different dust types and environmental fluctuation scenarios.

[0005] The technical solution for achieving the present application is as follows:

[0006] The charge type dust concentration measurement method based on coupling compensation and characteristic matching comprises the following steps:

[0007] In each cycle, the peak value, main frequency, signal duration, temperature and humidity of the continuous window are obtained by high-frequency sampling, the signal vector and environment vector are generated and arranged in sequence to obtain the original signal matrix and the original environment matrix of each cycle;

[0008] The original signal matrix of each cycle is subjected to hierarchical filtering, the Leight criterion of reference peak value is used to filter out abnormal windows, and the wavelet threshold denoising algorithm is used to recombine to generate the pure signal matrix of each cycle, and the abnormal windows in the original environment matrix are simultaneously removed, and the pure environment matrix of each cycle is recombined;

[0009] The pure signal matrix of each cycle and the pure environment matrix are input into the heuristic neural network to mine the time sequence correlation and multi-modal fusion of the pure signal matrix and the pure environment matrix of each cycle, the influence of temperature and humidity on charge mobility is quantified based on the principle of charge induction and used as the initialization basis and physical constraint condition of the neuron weight in the hidden layer, and the correction signal matrix of each cycle is generated through forward propagation mapping;

[0010] The dynamic frequency entropy, peak variation coefficient, peak skewness, peak duration product variation coefficient and frequency duration ratio skewness of the correction signal matrix of each cycle are calculated to construct a signal feature vector, the standard feature vector corresponding to the concentration model with the highest cosine similarity in the concentration model library is selected through cosine matching, the dust concentration of all windows in the correction signal matrix of each cycle is calculated, and the average value is taken to obtain the dust concentration of each cycle.

[0011] Further, at the start time of the i-th cycle M times of window-based high-frequency sampling are performed at a sampling frequency, in the m-th sampling, the maximum voltage value of the signal in the sensing window is induced to obtain the peak value of the m-th sampling in the i-th cycle The frequency value with the most concentrated energy in the signal in the analysis extraction window is analyzed to obtain the main frequency of the m-th sampling in the i-th cycle The duration of the signal in the recording window being greater than the signal threshold value is recorded to obtain the signal duration τ of the m-th sampling in the i-th cycle i,m The signal vector x of the m-th sampling in the i-th cycle is generated i,m The temperature T i,m and the humidity H i,m at the left frame time in the synchronous measurement window are measured to generate the environment vector of the m-th sampling in the i-th cycle The signal vector and the environment vector of M samplings are spliced according to the time sequence to obtain the original signal matrix X of the i-th cycle i and the original environment matrix wherein the window width is equal to the inverse of the sampling frequency, the left frame time of the window of the m-th sampling in the i-th cycle is equal to the start time of the i-th cycle the product of the sampling number m minus 1 and the window width, the right frame time of the m-th sampling window is equal to the left frame time of the m+1-th sampling window, i is the cycle number, i=1, 2, 3, …, and M is the total number of samplings.

[0012] Specifically, the peak values of all windows are extracted from the original signal matrix X of the i-th cycle i to splice to obtain the peak vector of the i-th cycle The average peak mean value of the i-th cycle and the peak standard deviation of the i-th cycle Based on the Leighton criterion, the reasonable lower limit and reasonable upper limit of the peak value for the i-th period are set as the average peak value of the i-th period, respectively. Reduce the peak standard deviation by 3 times and 3 times the peak standard deviation Window values ​​that are less than the reasonable lower peak value limit or greater than the reasonable upper peak value limit are marked as abnormal windows, and their corresponding peak values, main frequencies, and signal durations are removed to generate the filtered signal matrix for the i-th period.

[0013] Furthermore, the wavelet thresholding denoising algorithm includes the following steps:

[0014] From the filtered signal matrix of the i-th period Extract the peak values ​​of the window sampled M′ times, and concatenate them to generate the filtered peak vector of the i-th period. Then, a 5-level progressive decomposition is performed using the db4 wavelet basis to generate the low-frequency coefficient vector a of the 5th level in the i-th period. i,5 And the high-frequency coefficient vector of the 5th layer in the i-th cycle, M′ is the total number of remaining samples;

[0015] The absolute median difference MAD(d) of the first layer in the i-th period is... i,1 Dividing by the standardized coefficient corresponding to the median of the normal distribution yields the noise standard deviation σ for the i-th period. i Wherein, the high-frequency coefficient vector d of the first layer in the i-th period is... i,1 The vector d of each high-frequency coefficient in the first layer of the i-th period. i,1 median(d) i,1 The absolute deviation of each high-frequency coefficient is obtained by subtracting the values ​​and taking their absolute values. The median of the absolute deviations is the absolute median difference (MAD) of the first layer in the i-th period. i,1 ), median(·) represents the median operation;

[0016] The noise standard deviation σ of the i-th period i Multiplying by the square root of the natural logarithm of twice the total number of remaining samples M′ yields the Donoh threshold ξ for the i-th period. i Soft thresholding is then applied to any high-frequency coefficient d in the high-frequency coefficient vector of the i-th cycle of the 5th layer, replacing the high-frequency coefficient d with a noise-reduced high-frequency coefficient d. ξ When the absolute value of the high-frequency coefficient d is greater than the Donoh threshold ξ of the i-th period i At that time, the high-frequency coefficient d of noise reduction ξ It equals the absolute value of the high-frequency coefficient d and the Donohue threshold ξ of the i-th period. i The difference is multiplied by the sign function value of the high-frequency coefficient d; in other cases, the noise reduction high-frequency coefficient d... ξIf the value is 0, generate the noise reduction high-frequency coefficient vector of the 5th layer in the i-th cycle, where the sign function values ​​corresponding to negative numbers, 0 and positive numbers are -1, 0 and 1 respectively;

[0017] Using the db4 wavelet basis to analyze the low-frequency coefficient vector a of the 5th layer in the i-th period i,5 The denoised high-frequency coefficient vector of the 5th layer in the i-th period is reconstructed by inverse transformation to generate the clean signal matrix of the i-th period.

[0018] Furthermore, when using the db4 wavelet basis for progressive decomposition, the filtered peak vector of the i-th period is... Consider the low-frequency coefficient vector a of the 0th layer in the i-th period as i,0 During each layer decomposition, low-pass and high-pass filters are used to convolve with the low-frequency coefficient vector of the current layer in the i-th period, and simultaneously downsampled by a factor of 2 to generate the low-frequency and high-frequency coefficient vectors of the next layer in the i-th period, until the low-frequency coefficient vector a of the 5th layer in the i-th period is obtained. i,5 With high-frequency coefficient vector d i,5 When using the db4 wavelet basis for inverse transform reconstruction, the low-frequency coefficient vector and the noise-reduced high-frequency coefficient vector of the current layer in the i-th period are simultaneously upsampled by a factor of 2 and then transposed and convolved with the low-pass filter and high-pass filter respectively. The summation yields the low-frequency coefficient vector of the previous layer in the i-th period, and so on until the 0th layer, thus obtaining the clean signal matrix of the i-th period.

[0019] Furthermore, heuristic neural networks include multimodal temporal modules, hidden modules, and output modules;

[0020] The multimodal timing module normalizes the pure signal matrix of the i-th period. and pure environment matrix Generate the normalized signal matrix of the i-th period and normalized environment matrix Extracting the normalized signal matrix of the i-th period using one-dimensional convolution The temporal correlation in the transpose is used to generate the temporal feature matrix of the i-th period. With the normalized environment matrix of the i-th period The transpose of the tensor is concatenated and mapped through a fully connected layer to generate the input tensor of the i-th period;

[0021] The hidden module includes a first hidden layer and a second hidden layer, the first hidden layer includes 32 signal neurons, 16 temperature neurons and 16 humidity neurons, the input tensor of the i-th period is linearly modulated with the first weight matrix constructed by 64 neurons to generate the first output vector of the i-th period after ReLU function mapping, the second hidden layer includes 16 temperature neurons and 16 humidity neurons, the first output vector of the i-th period is linearly modulated with the second weight matrix constructed by 32 neurons to generate the second output vector of the i-th period after ReLU function mapping;

[0022] The output module linearly modulates the second output vector of the i-th period with the third linear weight to generate the correction signal matrix of the i-th period The correction signal matrix of the i-th period The correction signal matrix of the i-th period includes the correction peak value, the correction main frequency and the correction signal duration in the window of M' times of sampling.

[0023] Further, the hidden module is designed based on the charge induction principle, and a linear combination of a difference between a temperature T and a standard temperature T std and a difference between a humidity H and a standard humidity H std is calculated, a power value obtained by taking the linear combination as an index with a natural constant as a base is equal to a ratio of a charge mobility μ to a standard charge mobility μ std , temperature migration factor α1 and humidity migration factor α2 in the linear combination are determined according to charge determination experimental data fitting, all initial weights of the temperature neurons are set to 0.1 times of the temperature migration factor α1, all initial weights of the humidity neurons are set to 0.1 times of the humidity migration factor α2, a physical constraint is set, and a difference between a weight w involved in the temperature neurons and the humidity neurons and a corresponding initial weight w o must be less than or equal to 0.2 times of the initial weight w o , that is, 0.8w o ≤w≤1.2w o , wherein the initial weight w o is determined according to whether the weight w is derived from the temperature neurons or the humidity neurons.

[0024] Further, pre-training of the heuristic neural network needs to collect a data set through a dust concentration experiment, an original signal matrix is sampled under working conditions formed by different temperature and humidity combinations, the dust concentration of each sampling is calculated by using the technical solution of the application, the real dust concentration of each time is obtained by a weighing method, and the error between the dust concentration and the real dust concentration is minimized as an optimization objective, an Adam optimizer is used to adjust network parameters involved in the heuristic neural network, after each round of update of the network parameters, the weights w in the temperature neurons and the humidity neurons that do not satisfy the physical constraint are linearly scaled and adjusted to obtain new weights w after adjustment.new When the optimization target converges, the heuristic neural network pre-training is completed.

[0025] Further, the dynamic frequency entropy of the i th cycle is calculated The steps include:

[0026] The i th cycle of the correction signal matrix The M' times of the sampling of the correction main frequency is extracted and the i th cycle of the correction main frequency vector is constructed The silhouette coefficient is used to measure the clustering effect, and the clustering number N of K-Means is set from 3 to 8 in turn, and the i th cycle of the correction main frequency vector is clustered, and the best clustering number N closest to 1 is selected * , and the corresponding clustering center is determined;

[0027] The minimum correction main frequency and the maximum correction main frequency in the i th cycle of the correction main frequency vector are selected as the lower limit and the upper limit of the interval, respectively, the midpoint of the adjacent clustering centers is taken as the interval division point, and the interval is divided into dynamic sub-intervals with the number equal to the best clustering number N * .

[0028] The number of the correction main frequencies falling in each dynamic sub-interval is counted and divided by the total number of the remaining samples M', to obtain the energy proportion of each dynamic sub-interval, the center correction main frequency of each dynamic sub-interval is confirmed and divided by the sum of the center correction main frequencies of all dynamic sub-intervals, to obtain the frequency weight of each dynamic sub-interval, the product of the frequency weight, the energy proportion and the logarithm value of the energy proportion of each dynamic sub-interval is accumulated and taken as the opposite number, to obtain the dynamic frequency entropy of the i th cycle

[0029] Further, the peak value coefficient of variation of the i th cycle and the peak value skewness The steps include:

[0030] The M' times of the sampling of the correction peak value is extracted from the i th cycle of the correction signal matrix and the i th cycle of the correction peak value vector is constructed The i th cycle of the correction peak value vector is divided into continuous K correction peak value groups according to the order, and if it cannot be divided evenly, the excess correction peak value is added to the K correction peak value group;

[0031] The peak value group mean and the peak value group standard deviation of each correction peak value group are calculated, and the sum of the peak value group standard deviations of the K correction peak value groups is divided by the sum of the peak value group means to obtain the peak value coefficient of variation of the i th cycle

[0032] calculating the corrected peak value vector of the i-th cycle the corrected peak value mean of the i-th cycle the corrected peak value standard deviation of the i-th cycle and performing Z-Score normalization to obtain the normalized corrected peak value vector of the i-th cycle summing the 3rd power of each normalized corrected peak value in the normalized corrected peak value vector of the i-th cycle and multiplying by the correction ratio λ(M) related to the total number of remaining samples M' to obtain the peak value skewness of the i-th cycle wherein the correction ratio λ(M) is equal to the total number of remaining samples M' divided by the product of the remainder of the total number of remaining samples M' minus 1 and the remainder of the total number of remaining samples M' minus 2.

[0033] Further, calculating the peak duration product coefficient of variation of the i-th cycle and the frequency duration ratio skewness of the i-th cycle comprising the following steps:

[0034] extracting the corrected signal duration of M' samples from the corrected signal matrix of the i-th cycle and constructing the corrected signal duration vector of the i-th cycle

[0035] performing element-wise multiplication between the corrected peak value vector of the i-th cycle and the corrected signal duration vector of the i-th cycle to obtain the peak duration product vector of the i-th cycle, performing element-wise division between the corrected main frequency vector of the i-th cycle and the corrected signal duration vector of the i-th cycle to obtain the frequency duration ratio vector of the i-th cycle;

[0036] dividing the peak duration product vector of the i-th cycle into K1 consecutive peak duration product groups according to order, and if there is a remainder, the remainder is incorporated into the K1th peak duration product group;

[0037] calculating the product group mean and the product group standard deviation of each peak duration product group, and dividing the sum of the product group standard deviations of the K1 peak duration product groups by the sum of the product group means to obtain the peak duration product coefficient of variation of the i-th cycle

[0038] calculating the ratio mean and the ratio standard deviation of the frequency duration ratio vector of the i-th cycle and performing Z-Score normalization to obtain the normalized frequency duration ratio vector of the i-th cycle, summing the 3rd power of each normalized frequency duration ratio in the normalized frequency duration ratio vector of the i-th cycle and multiplying by the correction ratio λ(M) related to the total number of remaining samples M' to obtain the frequency duration ratio skewness of the i-th cycle

[0039] Furthermore, the concentration model library needs to be pre-built, clearly defining the dust type. For each dust type, scenarios with different dust concentrations are created under standard operating conditions. The original signal matrix is ​​sampled multiple times for different periods at each dust concentration. The original signal matrix for each period is processed to generate a corrected signal matrix. Labels are added using the dust concentration. The signal feature vector for each period is calculated. The mean of the signal feature vector for each period is used as the standard feature vector for each dust type at that dust concentration. A BP neural network model is trained based on the corrected signal matrix and labels for each dust type and used as the concentration model for each dust type. The corresponding storage formats for dust type, standard feature vectors for different dust concentrations, and concentration models are established, and the concentration model library is constructed.

[0040] Compared with existing technologies, this invention addresses the lack of quantification of the impact of temperature and humidity on charge mobility by incorporating the principle of charge induction into a heuristic neural network design. This quantifies the influence of temperature and humidity on charge mobility, using this as the initialization basis for the weights of temperature and humidity neurons in the hidden layer of the neural network. Physical constraints are set to ensure that weight adjustments conform to this principle, achieving coupling compensation between temperature / humidity and charge signals and reducing interference from environmental fluctuations on measurements. Furthermore, a dust type characteristic matching system is established. Addressing the inability of existing technologies to adapt to different dust types, a concentration model library containing both single and mixed dust types is pre-constructed. By calculating the signal feature vector of the correction signal matrix and performing cosine matching with the standard feature vector in the model library, the concentration model corresponding to the highest similarity is selected, achieving characteristic matching for different dust types and improving measurement adaptability. Attached Figure Description

[0041] Figure 1 Flowchart of a charge-based dust concentration measurement method based on coupling compensation and characteristic matching;

[0042] Figure 2 Here is a flowchart of the wavelet thresholding noise reduction algorithm;

[0043] Figure 3 The flowchart shows the progressive decomposition process based on the db4 wavelet basis.

[0044] Figure 4 This is a diagram of a heuristic neural network model. Detailed Implementation

[0045] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0046] like Figure 1 As shown, a specific embodiment of the present invention discloses a charge-based dust concentration measurement method based on coupling compensation and characteristic matching, comprising the following steps:

[0047] At the starting moment of each cycle, high-frequency sampling is performed at a sampling frequency, and each sampling obtains the peak value, main frequency, signal duration, temperature and humidity within the window to generate a signal vector and an environment vector, and the signal vector and the environment vector are arranged in time sequence respectively at equal intervals in each cycle to generate an original signal matrix and an original environment matrix of each cycle;

[0048] The original signal matrix of each cycle is subjected to hierarchical filtering, the abnormal signal vector corresponding to the abnormal window is filtered out by using the Leight criterion of the reference peak value, and denoising processing is performed through a wavelet threshold denoising algorithm, and a pure signal matrix of each cycle is recombined, and the environment vector corresponding to the abnormal window in the original environment matrix of each cycle is synchronously removed to recombine a pure environment matrix of each cycle, wherein the window of the pure signal matrix and the pure environment matrix of each cycle is kept synchronous in time sequence;

[0049] The pure signal matrix and the pure environment matrix of each cycle are cooperatively input into a heuristic neural network, the pure signal matrix and the pure environment matrix of each cycle are mined in time sequence and multi-modal fusion, based on the principle of charge induction, the influence of temperature and humidity on charge mobility is quantified and used as the initialization basis and physical constraint condition of the neuron weight in the hidden layer, and a corrected signal matrix of each cycle is generated through forward propagation mapping;

[0050] Based on the corrected signal matrix of each cycle, the dynamic frequency entropy, peak value variation coefficient, peak value skewness, peak value duration product variation coefficient and frequency duration ratio skewness of each cycle are calculated, a signal feature vector of each cycle is constructed, cosine matching is performed between the signal feature vector and a standard feature vector in a concentration model library, a concentration model corresponding to the standard feature vector with the highest cosine similarity is selected, the dust concentration of all windows in the corrected signal matrix of each cycle is calculated, and the average value is taken to obtain the dust concentration of each cycle.

[0051] Further, an embedded controller is used to perform periodic scheduling on the alternating current charge induction sensor and the temperature and humidity sensor, at the starting moment of the i-th cycle The embedded controller sends a high-frequency sampling instruction to the alternating current charge induction sensor and the temperature and humidity sensor through an SPI communication interface, the alternating current charge induction sensor receives the high-frequency sampling instruction and performs M times of window-based high-frequency sampling at a sampling frequency, the alternating current charge induction sensor includes a peak value holding circuit, a frequency detection circuit and a comparison timing circuit, the peak value holding circuit captures the maximum voltage value of the signal within the window of the m-th sampling to obtain the peak value of the m-th sampling in the i-th cycle The frequency detection circuit analyzes the spectral characteristics of the signal within the window of the m-th sampling, extracts the frequency value with the most concentrated energy, and obtains the main frequency of the m-th sampling in the i-th cycle The comparison timing circuit records the duration of the signal greater than the signal threshold within the window of the m-th sampling, and obtains the signal duration τ of the m-th sampling in the i-th cyclei,m , the signal vector x of the i th cycle m th sampling is generated i,m , the temperature T and humidity H of the left frame time in the window of the i th cycle m th sampling are measured i,m i,m , the environmental vector of the i th cycle m th sampling is generated The signal vector and the environmental vector obtained by M times of sampling are spliced according to the time sequence order to obtain the original signal matrix X of the i th cycle i and the original environmental matrix Specifically as follows:

[0052]

[0053] Wherein, the window width is equal to the sampling interval, which is equal to the reciprocal of the sampling frequency, the left frame time of the window of the i th cycle m th sampling is equal to the starting time of the i th cycle The right frame time of the window of the m th sampling is equal to the left frame time of the window of the m+1 th sampling, it is worth noting that the right frame time of the window of the M th sampling of the i th cycle is exactly equal to the starting time of the i+1 th cycle That is, according to the sampling frequency, the period length between the starting time of the i th cycle and the starting time of the i+1 th cycle can be divided into M sampling windows with equal interval, i is the cycle number, i = 1, 2, 3, …, M is the total number of sampling, in this embodiment, the embedded controller uses STM32F407 industrial single chip microcomputer, the cycle length and the sampling frequency are set to 10 seconds and 50 hertz respectively, that is, the sampling interval is 20 milliseconds, the total number of sampling M = 50.

[0054] Specifically, for the original signal matrix X of the i th cycle i , according to the window time sequence, the peak value of the M times of sampling window is extracted, and spliced as the peak vector of the i th cycle The average peak mean value of the i th cycle is calculated And the peak standard deviation of the i th cycle According to the Leight criterion, the reasonable peak lower limit and the reasonable peak upper limit of the i th cycle are set to be equal to the average peak mean value of the i th cycle minus 3 times the peak standard deviation and the average peak mean value of the i th cycle plus 3 times the peak standard deviation The windows in the peak vector of the i th cycle that satisfy less than the reasonable peak lower limit or greater than the reasonable peak upper limit are marked as abnormal windows, and the original signal matrix X of the i th cycle i ​The peak value, main frequency, and signal duration corresponding to the abnormal window are filtered out to generate the filtered signal matrix for the i-th period.

[0055] like Figure 2 As shown, the wavelet thresholding denoising algorithm further includes the following steps:

[0056] Obtain the filtered signal matrix of the i-th period generated by the Leighton criterion. Then, extract the peak values ​​of the window sampled M′ times, and concatenate them column-wise to form the filtered peak vector for the i-th period. M′ represents the total number of remaining samples;

[0057] The filtered peak vector of the i-th period is obtained using the db4 wavelet basis. Perform a 5-layer progressive decomposition to obtain the filtered peak vector of the i-th cycle. Different frequency components are assigned to different layers to generate the low-frequency coefficient vector a of the 5th layer in the i-th period. i,5 and the high-frequency coefficient vector d of the i-th cycle of the 5th layer i,5 d i,4 d i,3 d i,2 and d i,1 , where d i,j Let j be the high-frequency coefficient vector of the j-th layer in the i-th period, where j∈{1,2,3,4,5}. The smaller the layer number j, the higher the frequency.

[0058] Since the noise is distributed in the high-frequency region, the high-frequency coefficient vector d of the first layer in the i-th period is... i,1 The noise proportion is the highest, based on the high-frequency coefficient vector d of the first layer of the i-th period. i,1 Estimate the noise standard deviation σ for the i-th period i The details are as follows:

[0059]

[0060] Among them, MAD(d i,1 ) represents the absolute median difference of the first layer in the i-th period, which reflects the high-frequency coefficient vector d of the first layer in the i-th period. i,1 The degree of dispersion is determined by confirming the high-frequency coefficient vector d of the first layer in the i-th period. i,1 median(d) i,1 ), and the high-frequency coefficient vector d of the first layer in the i-th period. i,1 Subtract the median(d) from each high-frequency coefficient. i,1 And take the absolute value to obtain the median(d) for each high-frequency coefficient. i,1 The absolute deviation of the first layer in the i-th period is determined by the median of the absolute deviation.i,1 ), 0.6745 is the normalized coefficient corresponding to the median of normal distribution, and median(·) represents the median operation;

[0061] The noise standard deviation σ i of the i th period is determined dynamically i The i th period of the Donoho threshold ξ i is equal to the noise standard deviation σ i of the i th period multiplied by the square root of the natural logarithm value of 2 times the total number of remaining samples M' The i th period of the Donoho threshold ξ i is related to the total number of remaining samples M', and can adaptively balance noise suppression and signal detail preservation;

[0062] The soft threshold processing is performed on any high frequency coefficient d in the high frequency coefficient vector of the 5 th layer of the i th period. When the absolute value of the high frequency coefficient d is greater than the i th period of the Donoho threshold ξ i , the difference between the absolute value of the high frequency coefficient d and the i th period of the Donoho threshold ξ i is multiplied by the sign function value of the high frequency coefficient d to obtain the corresponding denoising high frequency coefficient d ξ = sign(d)·(|d|-ξ i ), when the absolute value of the high frequency coefficient d is less than or equal to the i th period of the Donoho threshold ξ i , the corresponding denoising high frequency coefficient d ξ is set to 0, that is, the denoising high frequency coefficient d ξ is as follows:

[0063]

[0064] Wherein, sign(·) is a sign function, the sign function values of negative numbers, 0 and positive numbers are-1, 0 and 1 respectively, the high frequency coefficient d belongs to the set composed of the high frequency coefficient vector of the 5 th layer of the i th period, and after the soft threshold processing, the high frequency coefficient vector of the 5 th layer of the i th period is converted into the denoising high frequency coefficient vector of the 5 th layer of the i th period and Wherein, is the denoising high frequency coefficient vector of the j th layer of the i th period;

[0065] Since the low frequency coefficient vector a i,5 of the 5 th layer of the i th period includes the main component of the effective signal, it is directly retained, and the low frequency coefficient vector a i,5 of the 5 th layer of the i th period is reconstructed by inverse transforming the denoising high frequency coefficient vector of the 5 th layer of the i th period with the db4 wavelet base, and restored to the filter signal matrix of the i th period Using the same dimensions, the remaining total number of samples M′ is used to renumber the clean signal matrix of the i-th period after noise reduction.

[0066] like Figure 3 As shown, further, the db4 wavelet basis is an orthogonal wavelet with fourth-order vanishing moments. When using the db4 wavelet basis for progressive decomposition, the filtered peak vector of the i-th period is... Consider the low-frequency coefficient vector a of the 0th layer in the i-th period as i,0 Using low-pass and high-pass filters respectively, and the low-frequency coefficient vector a of the 0th layer in the i-th period, i,0 Convolution is performed simultaneously with a 2x downsampling to preserve even index values, generating the low-frequency coefficient vector a of the first layer in the i-th period. i,1 With high-frequency coefficient vector d i,1 The low-frequency coefficient vector a of the first layer in the i-th period i,1 The vector a is then convolved again with the low-pass filter and the high-frequency filter, and then downsampled by 2 times to generate the low-frequency coefficient vector a of the second layer in the i-th period. i,2 and high-frequency coefficient vector d i,2 The low-frequency coefficient vector of the i-th period is continuously decomposed until the low-frequency coefficient vector a of the 5th layer of the i-th period is obtained. i,5 With high-frequency coefficient vector d i,5 The progressive decomposition is completed. Similarly, when using the db4 wavelet basis for inverse transform reconstruction, the low-frequency coefficient vector a of the 5th layer in the i-th period is... i,5 Perform a 2x upsampling and transpose the result with the low-pass filter to convolve the denoising high-frequency coefficient vector of the 5th layer in the i-th period. Perform a 2x upsampling and transpose the result into a high-pass filter, then sum the results to obtain the low-frequency coefficient vector a of the 4th layer in the i-th period. i,5 Following the same approach, the inverse transformation is performed step by step to lower levels to reconstruct the signal up to level 0, generating a pure signal matrix for the i-th period.

[0067] like Figure 4 As shown, the heuristic neural network further includes a multimodal temporal module, a hidden module, and an output module;

[0068] The multimodal timing module obtains the pure signal matrix for the i-th period. and pure environment matrix For the remaining M′ samples, minimum-maximum normalization is performed on the peak value, main frequency, signal duration, temperature, and humidity to generate the normalized signal matrix for the i-th period. and normalized environment matrix Since the normalized signal matrix of the i-th period is filtered by the Leyte criterion There may be a time interval between adjacent sampling windows that is an integer multiple of the window width, and the time interval between any two adjacent sampling windows may not be the same. The normalized signal matrix of the i-th period is extracted using a one-dimensional convolution with a kernel size of 3, a stride of 1, and 16 output channels. The temporal correlation between adjacent sampling windows of different durations in the transpose is used to generate a temporal feature matrix of dimension 16×M′ for the i-th period. With the normalized environment matrix of the i-th period The transposes of the tensors are concatenated and mapped through a fully connected layer to generate an input tensor of dimension 64×M′ for the i-th period.

[0069] The hidden module includes a first hidden layer and a second hidden layer. The first hidden layer includes 64 neurons, wherein the ratio of signal neurons, temperature neurons, and humidity neurons is 2:1:1. When the input tensor of the i-th cycle is input into the first hidden layer, the input tensor of the i-th cycle is linearly modulated with a first weight matrix of dimension 64×64 constructed by the 64 neurons and mapped by the ReLU function to generate a first output vector of dimension 64×1 for the i-th cycle. The second hidden layer includes 16 temperature neurons and 16 humidity neurons. The first output vector of the i-th cycle is linearly modulated with a second weight matrix of dimension 32×64 constructed by the 32 neurons and mapped by the ReLU function to generate a second output vector of dimension 32×1 for the i-th cycle.

[0070] The output module uses a third linear weight of dimension M′×3×32 to linearly modulate the second output vector of the i-th period, generating a correction signal matrix of dimension M′×3 for the i-th period. The correction signal matrix of the i-th period This includes the correction peak value, correction main frequency, and correction signal duration within the window of M' samplings.

[0071] Furthermore, the hidden module is designed based on the principle of charge induction. The charge mobility μ exhibits an exponential correlation with the linear transformation of temperature T and humidity H, as detailed below:

[0072] μ = μ std ·exp[-α1·(TT std )+α2·(HH std )],

[0073] Where, μ std T std and H std These represent the standard charge mobility, standard temperature, and standard humidity under standard operating conditions, with standard temperature T being the standard charge mobility, standard temperature, and standard humidity, respectively. std The temperature is 25 degrees Celsius, and the standard humidity is H. std It is 0.5, standard temperature Tstd and standard humidity H std The standard working condition is jointly constructed, and a1 and a2 are temperature migration factor and humidity migration factor respectively, all initial weights in the weight vector of the temperature neuron in the hidden module are set to 0.1 times of the temperature migration factor a1, all initial weights in the weight vector of the humidity neuron are set to 0.1 times of the humidity migration factor a2, and the physical constraint is set, which requires that the difference between the weight w involved in the temperature neuron and the humidity neuron and the corresponding initial weight w o must be less than or equal to 0.2 times of the initial weight w o , that is, 0.8w o ≤w≤1.2w o , wherein the initial weight w o is determined to be 0.1 times of the temperature migration factor a1 or 0.1 times of the humidity migration factor a2 according to whether the weight w is derived from the temperature neuron or the humidity neuron.

[0074] Further, the pre-training of the heuristic neural network needs to collect a data set through a dust concentration experiment, and an original signal matrix is obtained by sampling under working conditions formed by different temperature and humidity combinations using an AC charge induction sensor of the same type. The dust concentration of each sampling is calculated by using the technical solution of the present application, and the real dust concentration of each time is obtained synchronously by weighing method. The error between the dust concentration and the real dust concentration is minimized as the optimization objective, and the network parameters involved in the multi-modal time sequence module, the hidden module and the output module in the heuristic neural network are adjusted using the Adam optimizer. After each round of update of the Adam optimizer, the weights w in the temperature neuron and the humidity neuron that do not satisfy the physical constraint are linearly scaled and adjusted to obtain the new weights w new after adjustment, and the linear scaling and adjustment are as follows:

[0075]

[0076] The network parameter adjustment of the heuristic neural network always meets the requirements of the charge induction principle through the physical constraint. When the optimization objective converges, the pre-training of the heuristic neural network is completed.

[0077] Further, the dynamic frequency entropy of the i-th cycle is calculated as , which includes the following steps:

[0078] The M' times of sampling of the corrected main frequency are extracted from the corrected signal matrix of the i-th cycle , and the corrected main frequency vector of the i-th cycle is constructed The contour coefficient is used to measure the clustering effect, and the clustering number N of K-Means is set from 3 to 8 in turn, and the corrected main frequency vector of the i-th cycle is clustered Perform clustering and select the optimal number of clusters N whose silhouette coefficient is closest to 1. * And determine the corresponding cluster centers;

[0079] Select the corrected main frequency vector of the i-th cycle. The minimum and maximum corrected dominant frequencies are used as the lower and upper limits of the interval, respectively. The midpoint of adjacent cluster centers is used as the interval dividing point, and the interval is divided into N intervals, which is equal to the optimal number of clusters. * The dynamic sub-intervals;

[0080] The number of frequencies falling within the corrected dominant frequency in each dynamic sub-interval is counted and divided by the remaining total number of samples M′ to obtain the energy proportion of each dynamic sub-interval. The center corrected dominant frequency of each dynamic sub-interval is identified and divided by the sum of the center corrected dominant frequencies of all dynamic sub-intervals to obtain the frequency weight of each dynamic sub-interval. The frequency weight, energy proportion, and the logarithm of the energy proportion of each dynamic sub-interval are summed and their negative values ​​are taken to obtain the dynamic frequency entropy of the i-th period. The dynamic frequency entropy of the i-th period It reflects the frequency distribution of the signal in the i-th period.

[0081] Furthermore, the peak variation coefficient of the i-th period is calculated. and peak skewness Includes the following steps:

[0082] From the correction signal matrix of the i-th period Extract the correction peak values ​​from the M′ samples and construct the correction peak vector for the i-th period. The correction peak vector of the i-th period The peaks are divided into K consecutive corrected peak groups according to the order. If the remaining total number of samples M′ cannot be divided evenly by the total number of peak groups K, the excess corrected peaks are merged into the Kth corrected peak group.

[0083] Calculate the peak group mean and peak group standard deviation for each correction peak group separately. Divide the sum of the peak group standard deviations of the K correction peak groups by the sum of the peak group means to obtain the peak variation coefficient for the i-th period. Peak coefficient of variation in the i-th period It reflects the dispersion of the peak value in the i-th period;

[0084] Calculate the correction peak vector of the i-th cycle. Mean peak value of correction Standard deviation of the peak correction Z-score standardization is then performed to obtain the normalized correction peak vector for the i-th cycle. The normalized correction peak vector of the i-th period summing up the 3rd power of each normalized corrected peak value and multiplying it by a correction ratio λ(M) related to the total number of remaining samples M' to obtain the peak skewness of the i-th cycle the peak skewness of the i-th cycle reflecting the dust mass, wherein the correction ratio λ(M) is equal to the total number of remaining samples M' divided by the product of the remainder of the total number of remaining samples M' minus 1 and the total number of remaining samples M' minus 2.

[0085] Further, the peak duration product coefficient of variation of the i-th cycle is calculated and the frequency-duration ratio skewness comprising the following steps:

[0086] extracting the corrected signal duration of M' samples from the corrected signal matrix of the i-th cycle and constructing the corrected signal duration vector of the i-th cycle

[0087] performing element-wise multiplication of the corrected peak vector of the i-th cycle and the corrected signal duration vector to obtain the peak duration product vector of the i-th cycle, performing element-wise division of the corrected dominant frequency vector of the i-th cycle and the corrected signal duration vector to obtain the frequency-duration ratio vector of the i-th cycle;

[0088] dividing the peak duration product vector of the i-th cycle into K1 consecutive peak duration product groups according to the order, and if the total number of duration product groups K1 cannot be evenly divided by the total number of remaining samples M', then the excess peak duration products are incorporated into the K1th peak duration product group.

[0089] calculating the product group mean and the product group standard deviation of each peak duration product group, and dividing the sum of the product group standard deviations of the K1 peak duration product groups by the sum of the product group means to obtain the peak duration product coefficient of variation of the i-th cycle the peak duration product coefficient of variation of the i-th cycle reflecting the difference in particle mass in the i-th cycle;

[0090] calculating the ratio mean and the ratio standard deviation of the frequency-duration ratio vector of the i-th cycle and performing Z-Score standardization to obtain the normalized frequency-duration ratio vector of the i-th cycle, summing up the 3rd power of each normalized frequency-duration ratio in the normalized frequency-duration ratio vector of the i-th cycle and multiplying it by a correction ratio λ(M) related to the total number of remaining samples M' to obtain the frequency-duration ratio skewness of the i-th cycle the frequency-duration ratio skewness of the i-th cycle reflecting the stability of particle motion.

[0091] Furthermore, the concentration model library needs to be pre-built, clearly defining common dust types, including single dust and mixed dust, as the dust types to be adapted in the concentration model library. For each dust type, scenarios with different dust concentrations are created under standard operating conditions. The original signal matrix is ​​sampled multiple times for different periods at each dust concentration. Since the scenario is under standard operating conditions, there is no need to use heuristic neural networks for correction. The pure signal matrix of each period is obtained based on hierarchical filtering, which is regarded as the correction signal matrix of each period. The dust concentration is used to attach labels, and the signal feature vector of each period is calculated. The average value of the signal feature vector of each period is taken to obtain the standard feature vector of each dust type at the dust concentration. Based on the correction signal matrix sampled for each dust type and the corresponding labels, a training set is constructed to train the BP neural network model. The trained BP neural network model can output the dust concentration according to the correction signal matrix. The trained BP neural network model is used as the concentration model corresponding to each dust type. The corresponding storage format of dust type, standard feature vector of different dust concentrations and concentration model is established to construct the concentration model library.

[0092] This invention discloses a charge-based dust concentration measurement method based on coupling compensation and characteristic matching. In each cycle, high-frequency sampling acquires the peak value, dominant frequency, signal duration, temperature, and humidity of a continuous window, generating an original signal matrix and an original environment matrix for each cycle. The original signal matrix undergoes hierarchical filtering, using the Leyte criterion to remove abnormal windows and then reconstructing a clean signal matrix using a wavelet threshold denoising algorithm. Simultaneously, abnormal windows in the original environment matrix are removed to obtain a clean environment matrix, effectively improving signal purity. The clean signal matrix and clean environment matrix are then collaboratively input into a heuristic neural network to mine temporal correlations and perform multimodal fusion. Based on the charge induction principle, the influence of temperature and humidity on charge mobility is quantified. As the initialization basis and physical constraint for the weights of hidden layer neurons, a correction signal matrix is ​​generated through forward propagation mapping, which effectively reduces the interference of temperature and humidity fluctuations on charge signals and significantly improves the environmental adaptability of the measurement. The dynamic frequency entropy, peak coefficient of variation, peak skewness, peak-time product coefficient of variation, and frequency-time ratio skewness of the correction signal matrix are calculated to construct a signal feature vector. The corresponding concentration model is selected by cosine matching with the standard feature vector in the concentration model library. The dust concentration of all windows in the correction signal matrix of each period is calculated and the average value is taken to obtain the dust concentration of that period. This dynamically adapts to the measurement requirements of single dust and mixed dust, enhances the measurement adaptability, and significantly improves the accuracy and stability of dust concentration measurement.

[0093] The above merely describes the preferred embodiments of the present application, and the protection scope of the present application is not limited to the above-described embodiments. Any technical solution falling within the concept of the present application shall fall within the protection scope of the present application. It should be noted that, for ordinary skilled persons in the art, some improvements and refinements without departing from the principles of the present application shall also be considered as falling within the protection scope of the present application.

Claims

1. A charge-based dust concentration measurement method based on coupling compensation and characteristic matching, characterized in that, Includes the following steps: Each cycle acquires the peak value, main frequency, signal duration, temperature, and humidity of a continuous window through sampling, generating the original signal matrix and original environment matrix for each cycle; The original signal matrix for each cycle is subjected to hierarchical filtering. The Leyte criterion is used to filter out abnormal windows, and the pure signal matrix for each cycle is reconstructed using a wavelet threshold denoising algorithm. Abnormal windows in the original environment matrix are removed simultaneously, and the pure environment matrix for each cycle is obtained by reconstruction. The pure signal matrix and pure environment matrix of each cycle are input into the heuristic neural network to explore their temporal correlation and multimodal fusion. The influence of temperature and humidity on charge mobility is quantified based on the charge induction principle and used as the initialization basis and physical constraint of the weights of the hidden layer neurons. The correction signal matrix of each cycle is generated through forward propagation mapping. Calculate the signal feature vector of the correction signal matrix for each cycle, select the corresponding concentration model from the concentration model library through cosine matching, calculate the dust concentration of all windows of the correction signal matrix within the cycle and take the average value to obtain the dust concentration; The heuristic neural network includes a multimodal temporal module, a hidden module, and an output module. The hidden module is designed based on the principle of charge induction, and calculates the linear combination of the difference between temperature and standard temperature and the difference between humidity and standard humidity. The power value obtained by taking the natural constant as the base and the linear combination as the exponent is equal to the ratio of charge mobility to standard charge mobility. The temperature mobility factor and humidity mobility factor in the linear combination are determined by fitting the charge measurement experimental data. The initial weight of the temperature neuron is set to 0.1 times the temperature mobility factor, and the initial weight of the humidity neuron is set to 0.1 times the humidity mobility factor. Physical constraints are set to require that the difference between the weight of the temperature and humidity neurons and the corresponding initial weights does not exceed 0.2 times the initial weights. The wavelet thresholding denoising algorithm includes: performing a 5-level progressive decomposition on the filtered peak vector of the filtered signal matrix using a db4 wavelet basis to obtain low-frequency coefficient vectors and high-frequency coefficient vectors; based on the median of the high-frequency coefficient vector in the first level, subtracting the median from each high-frequency coefficient in the high-frequency coefficient vector in the first level and taking the absolute value to obtain the absolute deviation of each high-frequency coefficient relative to the median; confirming the median of the absolute deviation to obtain the absolute median difference of the first level, and then dividing it by the normalization coefficient corresponding to the median to obtain the noise standard deviation; multiplying the noise standard deviation by the square root of the natural logarithm of twice the total number of remaining samples to obtain the Donohue threshold; performing soft thresholding on the high-frequency coefficients, and reconstructing them using the inverse transform of the db4 wavelet basis to generate a clean signal matrix.

2. The charge-based dust concentration measurement method based on coupling compensation and characteristic matching as described in claim 1, characterized in that, Calculate the dynamic frequency entropy, peak variation coefficient, peak skewness, peak duration product variation coefficient, and frequency duration ratio skewness of the corrected signal matrix, and combine them to construct the signal feature vector.

3. The charge-based dust concentration measurement method based on coupling compensation and characteristic matching as described in claim 1, characterized in that, The Leyte criterion is used to filter out abnormal windows, which includes: extracting the peak values ​​of all windows from the original signal matrix to form a peak vector, calculating the mean of the peak values ​​and the standard deviation of the peak values; setting a reasonable lower limit for the peak values ​​as the mean of the peak values ​​minus 3 times the standard deviation of the peak values, and a reasonable upper limit for the peak values ​​as the mean of the peak values ​​plus 3 times the standard deviation of the peak values, and removing abnormal windows whose peak values ​​exceed the upper and lower limits.

4. The charge-based dust concentration measurement method based on coupling compensation and characteristic matching as described in claim 1, characterized in that, The construction of the concentration model library includes: identifying the dust type, sampling different concentration scenarios for each dust type under standard operating conditions, generating a correction signal matrix and attaching a concentration label, calculating the mean of the signal feature vector as the standard feature vector; training a BP neural network model as the corresponding concentration model, and establishing a corresponding storage format for dust type, standard feature vector and concentration model.

5. The charge-based dust concentration measurement method based on coupling compensation and characteristic matching as described in claim 2, characterized in that, The calculation of dynamic frequency entropy includes: clustering the corrected main frequency vector based on the silhouette coefficient and selecting the optimal number of clusters and cluster centers; dividing the interval according to the minimum and maximum corrected main frequencies and dividing it into dynamic sub-intervals based on the midpoint of the cluster centers; calculating the energy proportion and frequency weight of each sub-interval; and obtaining the dynamic frequency entropy by accumulating and summing the frequency weight, energy proportion, and logarithmic values ​​of the energy proportion and taking the opposite number.

6. The charge-based dust concentration measurement method based on coupling compensation and characteristic matching as described in claim 2, characterized in that, The calculation of peak variation coefficient and peak skewness includes: dividing the corrected peak vector into multiple groups, calculating the mean and standard deviation of each group, and dividing the sum of the standard deviations of all groups by the sum of the means to obtain the peak variation coefficient; calculating the mean and standard deviation of the corrected peak and performing Z-score standardization on the corrected peak vector, summing the cube of the normalized peak and multiplying by the correction ratio to obtain the peak skewness.

7. The charge-based dust concentration measurement method based on coupling compensation and characteristic matching as described in claim 2, characterized in that, The calculation of the peak duration product coefficient of variation and the frequency duration ratio skewness includes: multiplying the corrected peak value by the duration of the corrected signal element by element to obtain the peak duration product vector, and dividing the main frequency by the duration element by element to obtain the frequency duration ratio vector; calculating the peak duration product coefficient of variation in groups, and standardizing the frequency duration ratio vector to calculate the frequency duration ratio skewness.

8. The charge-based dust concentration measurement method based on coupling compensation and characteristic matching as described in claim 1, characterized in that, The pre-training of the heuristic neural network includes: sampling the original signal matrix under different temperature and humidity conditions, calculating the dust concentration and weighing it to obtain the true concentration; adjusting the network parameters using the Adam optimizer with the goal of minimizing the concentration error, verifying the physical constraints after each round of updates, scaling and adjusting the weights that do not meet the physical constraints, until the optimization objective converges.

Images