Dust concentration detection method based on neural network

Abstract

The neural network-based dust concentration detection method belongs to the technical field of dust concentration detection and dust explosion prevention. The method uses a dust injection type dust concentration measurement platform as an experimental environment, measures dust of different concentrations using a dust electrostatic induction sensor, records voltage signals and converts them into standard deviation sequences according to windows, and uses the standard deviation sequences as input to train a neural network algorithm to predict the standard deviation mean of unknown concentration dust. The method can continuously improve the measurement accuracy over time, so that the prediction curve slowly converges to the true state. Compared with sliding filtering and median filtering, neural network filtering can make the signal tend to be stable faster, has the advantages of high speed and high sensitivity, meets the needs of dust operation scenes, and has good popularization and application value.

Description

TECHNICAL FIELD

[0001] The present application relates to dust concentration detection technology, in particular to electrostatic induction detection dust concentration technology based on neural network method, and belongs to the technical field of dust concentration detection and dust explosion prevention. BACKGROUND

[0002] Dust, as a key industrial raw material, has a wide range of applications in various fields, such as construction, chemical industry and food industry. However, during production, manufacturing, transportation and storage, dust often disperses into the air, increasing the risk of pneumoconiosis and other related occupational diseases, affecting the health of workers; moreover, in the production and processing process, high temperature and electric spark are easily generated, and in the environment with high dust concentration, it can cause deflagration and explosion, which poses a threat to the safety of life and property of the site equipment, surrounding environment and workers. Therefore, real-time monitoring of dust concentration in production operation sites can effectively improve the safety of dust operation sites, reduce accident risks and protect the safety of personnel and property.

[0003] The current common dust detection technology in the industry mainly includes light scattering and electrostatic induction. Light scattering analyzes the remaining intensity of light after passing through the dust cloud to realize dust concentration characterization, but the detection unit is easily contaminated by dust in the environment, and needs to be frequently maintained, which is high in cost; moreover, due to excessive dependence on light penetration, it cannot work normally in high dust concentration. Electrostatic induction captures the voltage signal change caused by dust signal passing through the electrode sensing area, and comprehensively reflects the dust concentration through the continuous change characteristics of the voltage signal within a certain time, so it is suitable for a wide range of dust concentrations. Compared with light sensing, the element loss rate of electrostatic induction technology is slower, the expected service life is longer, it does not need frequent maintenance, and the investment cost is lower.

[0004] At present, the main application scenario of electrostatic induction technology is fixed pipeline. As a shielding body, the pipeline can ensure that the detection electrode is not disturbed by the outside world, and improve the accuracy and reliability of the detection effect. However, there are many open dust operation scenes, so there is a technical gap in the dust concentration measurement technology market for open scenes. Secondly, in some complex processing fields, there is often strong electromagnetic interference, which makes the voltage signal unstable, which greatly reduces the application field of electrostatic induction method for dust detection. This problem is particularly prominent in environments with unstable dust flow and large changes in dust concentration.

[0005] In recent years, the rise of neural network technology has become a hot spot of attention, and the algorithm can learn to build the corresponding relationship between the bottom state and the observation signal model, and then continuously train, improve and strengthen the generalization ability of the model. In the field of dust concentration prediction, neural network provides a new possibility for solving the problems of open space inapplicability, easy electromagnetic interference and slow speed of inductive method. Therefore, in order to further expand the application range of electrostatic induction dust concentration detection technology, the present application proposes a dust concentration detection method based on neural network, which obtains a certain scale of dust concentration data through a large number of experiments, and uses the data for model training to build a network model for generalizing and capturing the overall characteristics of dust concentration. SUMMARY

[0006] In view of the problems of inapplicability at high concentration, easy noise interference and slow measurement speed of the existing dust measurement technology, one of the purposes of the present application is to provide an electrostatic induction dust concentration detection method based on neural network algorithm, which reduces volatility, improves measurement accuracy and speed, and expands the application range of dust concentration in open space.

[0007] Previous studies have shown that the voltage change caused by dust clouds of different concentrations when passing through the electrode induction area can represent the dust concentration to a certain extent. The standard deviation of the voltage signal has the characteristics of continuity, which means that the voltage signal measurement value at the next moment will be affected by the previous state. However, using general filtering methods to process the voltage signal cannot effectively represent the concentration signal of the dust, because the occurrence of different voltage signals has time correlation, which puts certain requirements on the neural network model used.

[0008] In the field of neural network model research, the long short-term memory recursion (LSTM) model has been proved to have good performance in tasks related to time series, and is suitable for predicting the complex dynamics of the system bottom layer. Therefore, the neural network model used in the present application is mainly based on LSTM.

[0009] In the specific process of constructing the neural network, first, the standard deviation in the window is calculated for the voltage signal induced by dust of different concentrations measured multiple times, and the standard deviation and the window sequence are combined into a two-dimensional vector as the input of the neural network. According to a certain proportion, it is divided into test set and training set; the data set is divided into different sizes and batches, and is respectively introduced into the neural network constructed by the input layer, the hidden layer and the output layer. The initial weight and bias parameters are randomly set, the L2 norm loss value of the output result and the target value is taken as the loss function, the weight and bias parameters are updated by stochastic gradient descent, and the network model for predicting the mean of the standard deviation is trained. After the training is completed, the model weight file is exported, the dust data collected is predicted for concentration value, and the effect is verified by comparison with other filtering.

[0010] The technical scheme adopted by the present application is: a dust concentration detection method based on a neural network, different concentration dust data obtained through multiple repeated experiments are analyzed and learned in training to reflect sequence characteristics of different concentration dust, different feature weight combinations are used to realize identification and prediction of dust that has never been contacted, and the most possible standard deviation distribution sequence is given, and the method specifically comprises the following steps:

[0011] S1. Collect data

[0012] Different quality dust materials are divided into several groups, are uniformly scattered in an open space through a powder spraying device, and different concentration dust clouds are obtained; an electric induction sensor is placed in the dust cloud, a Yokogawa oscilloscope is used to sample and record the voltage signal, and the recorded data is processed twice, with a fixed sampling number as a time window, the standard deviation of all voltage signals in the window is calculated in sequence according to the time sequence, the standard deviation continuous change value corresponding to the concentration is obtained, and is used to construct a labeled data set, and the specific calculation method is as follows:

[0013]

[0014] Wherein, std i represents the standard deviation value, i represents the window sequence, N is the sampling number in the time window, V ij is the voltage signal of the jth sampling point in the ith window, is the average voltage signal in the ith window;

[0015] After sampling and processing all concentration dust, the data set is divided; the signal processing data corresponding to different concentration dust is randomly selected, the measurement sequence i and the standard deviation std i are combined to form a two-dimensional vector, which is used as the input of the neural network, the overall mean of the corresponding standard deviation in the process is used as the target value, and the process is divided into a training set α train and a validation set α valid according to a certain proportion;

[0016] S2. Construct a model

[0017] A network model for predicting the mean of the standard deviation is constructed, and except for the input layer, the hidden layer is composed of LSTM units, which is represented as:

[0018] a (l) =f (l) (W (l) a (l-1) +b (l) )

[0019] Wherein: the output of each hidden layer l is a (l) , the weight is W (l) , and the bias is b (l); the initial weight and bias parameters are randomly initialized, i.e. The hidden layer nodes use the ReLU activation function, and the hidden state is initialized, whose formula is as follows:

[0020] ReLU(x) = max(0, x)

[0021] When using the back propagation algorithm to update the weight and bias parameters, the L2 norm loss is used to measure the difference between the predicted value of the model and the target value, and its formula is:

[0022]

[0023] Where: L is the loss function; is the output of the model; y is the target value; ||·||2 represents the L2 norm, i.e. the Euclidean norm;

[0024] In the back propagation algorithm, the gradient of the loss function with respect to the weight and bias is calculated, and then the gradient descent algorithm is used to update the parameters; the one-step update rule of gradient descent is represented as:

[0025]

[0026] Where, θ represents the weight or bias parameter, α is the learning rate, is the gradient of the loss function with respect to the parameter;

[0027] For the L2 norm loss, the gradient is calculated by the chain rule; for the update of the weight W, the gradient is calculated as follows:

[0028]

[0029] Where: is the partial derivative of the loss function with respect to W;

[0030] S3. Model training

[0031] The training set α train is passed into the network model, and the predetermined mode is trained, iterated, and the model parameters are adjusted so that it can effectively predict the corresponding actual concentration standard deviation mean through the overall standard deviation change, and the process is basically as follows:

[0032] std i,theory = f(W theory a i +b theory )

[0033] Where: std i,theory is the standard deviation theoretical prediction value, f is the network model prediction, W theory is the model weight, a iInput for the model, b theory Bias for the model;

[0034] The overall training data is divided into different batches and imported into the model, the loss function is monitored, and when the loss function no longer decreases with the training round, the training process is ended, and the model weight parameter file is generated.

[0035] The beneficial effects of the present application relative to the prior art: the traditional dust concentration measurement method has problems such as unsuitable for high concentration dust, instability, high delay, and large noise. To solve these problems, the neural network algorithm is used to perform secondary processing on the standard deviation of the sensing signal data, and eliminate the standard deviation fluctuation of the sensing signal. When applied to an electrostatic induction sensor, the standard deviation fluctuation of the detection signal can be eliminated in a short time, and the prediction result of the system is gradually optimized, which can not only realize high-concentration dust concentration detection, but also has the advantages of high sensitivity and strong anti-interference ability, and can quickly and accurately detect dust concentration.

[0036] The above summary is only for the purpose of the description and is not intended to limit in any way. In addition to the above-described illustrative aspects, embodiments and features, further aspects, embodiments and features of the present application will be apparent from the drawings and the following detailed description. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or the prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0038] Figure 1 It is a structural schematic diagram of a dust detection experimental system.

[0039] Figure 2 It is 200g / m 3 The standard deviation of the 30μm dust sensing signal is filtered by a neural network.

[0040] Figure 3 It is 400g / m 3 The standard deviation of the 30μm dust sensing signal is filtered by a neural network.

[0041] Figure 4 It is 600g / m 3 The standard deviation of the 30μm dust sensing signal is filtered by a neural network.

[0042] Figure 5 It is 800g / m3 Concentration 30 pm dust response signal standard deviation neural network filtering processing curve.

[0043] Figure 6 1000 g / m 3 Concentration 30 pm dust response signal standard deviation neural network filtering processing curve.

[0044] Figure 7 1200 g / m 3 Concentration 30 pm dust response signal standard deviation neural network filtering processing curve.

[0045] Figure 8 1400 g / m 3 Concentration 30 pm dust response signal standard deviation neural network filtering processing curve.

[0046] Figure 9 1600 g / m 3 Concentration 30 pm dust response signal standard deviation neural network filtering processing curve.

[0047] Figure 10 200 g / m 3 Concentration 30 pm dust response signal standard deviation neural network filtering and moving average filtering and median filtering processing effect comparison chart.

[0048] Figure 11 400 g / m 3 Concentration 30 pm dust response signal standard deviation neural network filtering and moving average filtering and median filtering processing effect comparison chart.

[0049] Figure 12 600 g / m 3 Concentration 30 pm dust response signal standard deviation neural network filtering and moving average filtering and median filtering processing effect comparison chart.

[0050] Figure 13 800 g / m 3 Concentration 30 pm dust response signal standard deviation neural network filtering and moving average filtering and median filtering processing effect comparison chart.

[0051] Figure 14 1000 g / m 3 Concentration 30 pm dust response signal standard deviation neural network filtering and moving average filtering and median filtering processing effect comparison chart.

[0052] Figure 15 1200 g / m 3The concentration 30 μm dust induction signal standard deviation through neural network filtering and moving average filtering and median filtering processing effect comparison chart.

[0053] Figure 16 For 1400 g / m 3 The concentration 30 μm dust induction signal standard deviation through neural network filtering and moving average filtering and median filtering processing effect comparison chart.

[0054] Figure 17 For 1600 g / m 3 The concentration 30 μm dust induction signal standard deviation through neural network filtering and moving average filtering and median filtering processing effect comparison chart.

[0055] Figure 18 For different algorithms, the deviation degree of the prediction result of the standard deviation of the 30 μm particle size dust induction signal. DETAILED DESCRIPTION

[0056] In the following, certain example embodiments are described. As will be realized by one of ordinary skill in the art, the described embodiments can be modified in various different ways without departing from the spirit or scope thereof. Accordingly, the drawings and descriptions are to be regarded as illustrative in nature rather than restrictive.

[0057] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0058] The electrostatic induction sensor used in the measurement process mainly works as follows: under the action of external power, dust particles pass through the middle of the electrostatic detection electrode. Since the dust particles have a certain static electricity, a certain amount of induced charge is induced on the surface of the electrostatic detection electrode. As the dust particles approach the electrostatic detection electrode, the induced charge on the surface of the electrostatic detection electrode gradually increases; as the dust particles move away from the electrostatic detection electrode, the induced charge on the surface of the electrostatic detection electrode gradually decreases. Therefore, the dynamic induced charge produced by the moving dust particles first increases and then decreases, generating an alternating current signal. The volatility of the signal is positively correlated with the particle concentration. Since the dust concentration is detected, the signal needs to be analyzed and counted over a period of time, so the alternating signal in the maximum time period displayed by the oscilloscope is analyzed, and the volatility is counted. As the dust concentration increases, the corresponding standard deviation value also increases, i.e. the volatility of the alternating signal increases. Therefore, the characteristic value reflecting the signal volatility can be extracted based on the effective algorithm, and the sampler is calibrated to realize real-time detection of the dust concentration.

[0059] A neural network dust concentration prediction model is used to obtain the standard deviation of the voltage signal change data induced by different concentrations of dust measured multiple times within a certain time window and to perform secondary processing. The time series and standard deviation mean are combined into a two-dimensional vector, which is divided into a test set and a training set according to a certain proportion. The data set is imported into a neural network model constructed by an input layer, a hidden layer and an output layer, appropriate initial weights and bias parameters are set, the L2 norm loss of the difference between the output result and the target value is used as the loss function, the parameters are updated through stochastic gradient descent, and a network model for predicting the theoretical concentration of dust at the next moment is trained. The accuracy of the concentration value prediction of the newly collected dust data is verified, thereby constructing an electrostatic induction detection dust concentration technology based on a neural network filtering algorithm, eliminating the standard deviation volatility of the induction signal, applying it to an electrostatic induction sensor, eliminating the volatility of the standard deviation of the detection signal in a short time, making the processed standard deviation curve quickly converge to the target value, improving the stability of the standard deviation of the induction signal, and thus showing the advantages of high sensitivity and strong anti-interference ability, enabling fast and accurate detection of dust concentration.

[0060] Data acquisition process:

[0061] The electrostatic induction sensor is used to detect the dust flow, and the charge signal value output by the electrostatic induction sensor is recorded. The charge signal value is input into the controller for filtering processing to obtain the characteristic value of the charge induction signal, which is used to represent the actual dust concentration value. The above steps are repeated to obtain multiple sample data, and the average value is calculated to improve the measurement accuracy.

[0062] The signal processing flow of the electrostatic induction sensor mainly includes the following three parts: data preparation, model training and feature value prediction. Set the signal acquisition frequency, acquisition time and data acquisition amount, remove the noise of the collected original data, calculate the standard deviation for quantifying the fluctuation degree of the signal and reflecting the stability and change of the signal.

[0063] Data processing flow:

[0064] 1: Collect data

[0065] First, weigh different amounts of 30μm dust material, divide it into several groups, and evenly distribute it in an open space through a powder spraying device to obtain a dust cloud of different concentrations. Place the inductive sensor in the dust cloud, allowing the dust to pass through the induction area in a random motion state, causing voltage disturbance. Use a Yokogawa oscilloscope to sample and record the voltage signal. Perform secondary processing on the recorded data, using a fixed sample number as the time window, and sequentially calculating the standard deviation of all voltage signals within the window according to the time sequence to obtain the standard deviation continuous change value corresponding to the concentration, which is used to construct the labeled data set. The specific calculation method is as follows:

[0066]

[0067] where std i represents the standard deviation value, i represents the window sequence, N is the number of samples in the time window, V ij is the voltage signal of the jth sampling point in the ith window, is the average voltage signal in the ith window.

[0068] After sampling and processing all concentrations of dust, the data set is divided. Randomly select the signal processing data corresponding to different concentrations of dust, combine the measurement sequence i and the standard deviation std i to form a two-dimensional vector as the input of the neural network, and take the overall mean of the process corresponding standard deviation as the target value. According to a certain proportion, divide into training set α train and validation set α valid .

[0069] 2: Model construction

[0070] The network model for predicting the mean of the standard deviation is constructed. Except for the input layer, its hidden layer is composed of LSTM units, which can be represented as:

[0071] a (l) = f (l) (W (l) a (l-1) +b (l) )

[0072] where: the output of each hidden layer l is a (l) , the weight is W (l) , and the bias is b (l) . The initial weight and bias parameters are randomly initialized, that is The hidden layer nodes use the ReLU activation function, and the hidden state is initialized, whose formula is as follows:

[0073] ReLU(x) = max(0, x)

[0074] When using the back propagation algorithm to update the weight and bias parameters, the L2 norm loss is used to measure the difference between the predicted value of the model and the target value. Its formula is:

[0075]

[0076] where: L is the loss function. is the output of the model. y is the target value. ||·||2 represents the L2 norm, that is, the Euclidean norm.

[0077] In the backpropagation algorithm, the gradient of the loss function with respect to the weights and biases is computed, and then these parameters are updated using gradient descent or other optimization algorithms. The one-step update rule for gradient descent can be expressed as:

[0078]

[0079] where θ represents the weight or bias parameter, α is the learning rate, is the gradient of the loss function with respect to the parameter.

[0080] For the L2 norm loss, the gradient can be obtained by the chain rule. For the update of the weight W, the gradient is calculated as follows:

[0081]

[0082] where: is the partial derivative of the loss function with respect to W.

[0083] 3: Model training

[0084] The training set α train is passed into the network model, and the predetermined mode is trained, iterated, and the model parameters are adjusted so that it can effectively predict the actual concentration standard deviation mean by the overall standard deviation change. The process is basically as follows:

[0085] std i,theory = f(W theory a i +b theory )

[0086] where: std i,theory is the standard deviation theoretical prediction value, f is the network model prediction, W theory is the model weight, a i is the model input, and b theory is the model bias.

[0087] The overall training data is divided into different batches and imported into the model, and the loss function is monitored. When the loss function no longer decreases with the training round, the training process is ended, and the generated weight parameter file is obtained.

[0088] Embodiment

[0089] Using 30-micron dust particles as the research object, a dust cloud was artificially created in an open space to simulate the dust state generated in an actual production site. An electrostatic induction sensor was used to measure the AC voltage signal value in the detection electrode area caused by the flow of the dust. For the collected data, neural network filtering, median filtering, and moving average algorithms were used to perform secondary processing on the standard deviation. The differences between the three algorithms and the average standard deviation were compared to verify their effectiveness.

[0090] The data processing and network model building process is as follows:

[0091] 1. Data collection

[0092] First, different masses of 30μm dust materials were weighed and divided into several groups. These groups were then evenly dispersed in an open space using a powder spraying device to obtain dust clouds of varying concentrations. An inductive sensor was placed within the dust cloud, allowing the dust particles to pass through the sensing area in a random motion, causing voltage disturbances. The voltage signals were sampled and recorded using a Yokogawa oscilloscope. The recorded data underwent secondary processing. Using a fixed number of samples as a time window, the standard deviation of all voltage signals within the window was calculated sequentially according to the time series. The continuous variation of the standard deviation for the corresponding concentration was obtained and used to construct a labeled dataset. The specific calculation method is as follows:

[0093]

[0094] Among them, std i The standard deviation value is represented by i, which represents the window sequence, N is the number of samples within the time window, and V is the standard deviation value. ij It is the voltage signal at the j-th sampling point within the i-th window. It is the average voltage signal within the i-th window.

[0095] After sampling and processing dust at all concentrations, the dataset is divided. Signal processing data corresponding to different dust concentrations are randomly selected, and the measurement sequence i is compared with the standard deviation std. i The vectors are combined to form a two-dimensional vector, which serves as the input to the neural network. The average standard deviation of the process is used as the target value, and the vectors are divided into a training set α according to a certain ratio. train and validation set α valid .

[0096] 2. Model Building

[0097] A network model for predicting the mean standard deviation is constructed, with its hidden layers consisting of LSTM units, excluding the input and output layers, and can be represented as follows:

[0098] a (l) =f (l) (W (l) a (l-1) +b(l) )

[0099] where: the output of each hidden layer l is a (l) , the weights are W (l) , and the biases are b (l) . The initial weights and biases parameters are initialized randomly, i.e. The hidden layer nodes use the ReLU activation function, and the hidden state is initialized as follows:

[0100] ReLU(x) = max(0, x)

[0101] When updating the weights and bias parameters using the backpropagation algorithm, the L2 norm loss is used to measure the difference between the model's predicted values and the target values. The formula is:

[0102]

[0103] where: L is the loss function. is the output of the model. y is the target value. ||·||2 represents the L2 norm, i.e. the Euclidean norm.

[0104] In the backpropagation algorithm, the gradient of the loss function with respect to the weights and biases is calculated, and then gradient descent or other optimization algorithms are used to update these parameters. The one-step update rule of gradient descent can be expressed as:

[0105]

[0106] where θ represents the weight or bias parameter, α is the learning rate, is the gradient of the loss function with respect to the parameter.

[0107] For the L2 norm loss, the gradient can be obtained by the chain rule. For the update of the weights W, the gradient calculation is as follows:

[0108]

[0109] where: is the partial derivative of the loss function with respect to W.

[0110] 3. Model training

[0111] The training set α train is passed into the network model, and the predetermined mode is used for training, constantly iterating and adjusting the model parameters, so that it can effectively predict the corresponding actual concentration standard deviation mean through the overall standard deviation change, and the process is as follows:

[0112] std i,tjeory = f(W theory a i + btheory )

[0113] where std i,theory is the standard deviation theoretical prediction, f is the network model prediction, W theory is the model weight, a i is the model input, b theory is the model bias.

[0114] The overall training data is divided into different batches and imported into the model, the loss function is monitored, and when the loss function no longer decreases with the training round, the training process is ended, and the generated weight parameter file is obtained.

[0115] 4. Effect verification

[0116] The dust concentration data not used for training is predicted to verify its performance. The target dust concentration sequence is set to 200 g / m 3 , 400 g / m 3 , 600 g / m 3 , 800 g / m 3 , 1000 g / m 3 , 1200 g / m 3 , 1400 g / m 3 , 1600 g / m 3 , the purpose of this is to ensure that the concentration covers the concentration range that occurs in actual production as much as possible, and the interval is equal, which can effectively reduce the experimental error; measure the total space volume in the acrylic pipe, based on the target concentration sequence, calculate the dust mass corresponding to each concentration, and weigh it with an electronic scale. In order to control other variables except concentration to remain unchanged, the pipe pressure is adjusted before each powder spraying experiment to keep the value fixed and reduce experimental error. After the device is debugged, the concentration measurement experiment is carried out, and a portion of the weighed dust sample is evenly spread around the air pressure nozzle; open the air valve switch, and the gas is sprayed out of the nozzle under the action of pressure, the dust is lifted under the action of airflow, fills the closed space inside the acrylic pipe, forms a dust cloud, and at the same time the electromagnetic valve is excited, the upper acrylic pipe falls down, the dust cloud dissipates to the electrode sensing area of the probe, the sampling process is shown in Figure 1 At this time, the sampler starts to work, the voltage change caused by the dust particles is captured and recorded by the Kikuchi oscilloscope connected with the electrostatic induction sensor; the voltage change signal value is input into the controller for noise removal and filtering processing, the average value of the standard deviation of the voltage change signal value in each window in the total measurement period is obtained, the complete data measured is uploaded to the computer, and the neural network model is used for data processing to generate prediction data of the standard deviation average, as shown in Figures 2-9As shown, the accuracy of the neural network filtering is measured by calculating the difference between the predicted value of the standard deviation and the true average value, and the prediction effect is compared with that of the traditional median filtering and sliding filtering. The consistency of the prediction result and the actual concentration is checked, and the accuracy and generalization ability of the model are evaluated. Figures 3-17 The results of the standard deviation curves of the sensing signals of different concentrations of 30 pm dust after being processed by the neural network filtering algorithm, the sliding filtering algorithm and the median filtering algorithm are shown. As can be seen from the figure, the curve processed by the neural network filtering algorithm is close to the average value of the standard deviation after 1 s of prediction. The curves processed by the sliding filtering algorithm and the median filtering algorithm are difficult to converge to the average value of the standard deviation. In order to make this deviation more obvious, the deviation of different curves from the target value is accumulated in Figure 18 The relative deviation of the neural network filtering algorithm is 0.136623, while the relative deviations of the conventional sliding filtering algorithm and the median filtering algorithm are 0.839123 and 0.847276, respectively. It can be judged that the neural network filtering algorithm has better processing effect on the standard deviation curve of the sensing signal caused by different concentrations of dust concentration compared with the conventional sliding filtering algorithm and the median filtering algorithm, and can quickly and accurately predict the average value of the standard deviation in a large range of dust concentration, which to some extent represents the dust concentration.

Claims

1. A dust concentration detection method based on neural networks, characterized in that: By obtaining dust data of different concentrations through repeated experiments, the sequence characteristics exhibited by dust of different concentrations are analyzed and learned during training. Based on different feature weight combinations, the identification and prediction of dust that has never been encountered before is realized, and the most likely standard deviation distribution sequence is given. The specific steps include: S1. Data Collection Different masses of dust materials were divided into several groups and evenly dispersed in an open space using a powder spraying device to obtain dust clouds of different concentrations. An inductive sensor was placed within the dust cloud, and the voltage signal was sampled and recorded using a Yokogawa oscilloscope. The recorded data underwent secondary processing. Using a fixed number of samples as a time window, the standard deviation of all voltage signals within the window was calculated sequentially according to the time series. The continuous variation of the standard deviation for the corresponding concentration was obtained and used to construct a labeled dataset. The specific calculation method is as follows: ； in, Indicates the standard deviation value. Represents a window sequence. It is the number of samples within the time window. It is the first The first in the window Voltage signal at sampling point It is the first The average voltage signal within the window; After sampling and processing dust at all concentrations, the dataset was divided; signal processing data corresponding to different dust concentrations were randomly selected, and the measurement sequences were... with standard deviation The vectors are combined to form a two-dimensional vector, which serves as the input to the neural network. The overall mean of the standard deviation corresponding to the process is used as the target value, and the data is divided into training sets according to a certain ratio. and verification set ; S2. Building the Model A network model for predicting the mean standard deviation is constructed, with its hidden layers consisting of LSTM units, excluding the input layer, as shown below: ； Where: each hidden layer The output is The weight is , bias is The initial weights and bias parameters are initialized randomly, i.e. Hidden layer nodes use the ReLU activation function and are initialized with hidden states, as shown in the following formula: ； When updating weights and bias parameters using the backpropagation algorithm, the L2 norm loss is used to measure the difference between the model's predictions and the target values. The formula is as follows: ； in: It is a loss function; It is the output of the model; It is the target value; This represents the L2 norm, also known as the Euclidean norm. In the backpropagation algorithm, the gradient of the loss function with respect to the weights and biases is calculated, and then the parameters are updated using the gradient descent algorithm; the one-step update rule of gradient descent is expressed as: ； in, Indicates the weight or bias parameter. It's the learning rate. It is the gradient of the loss function with respect to the parameters; For L2 norm loss, the gradient is calculated using the chain rule; for weights... The update, gradient calculation is as follows: ； in: It is a loss function pair The partial derivatives; S3. Model Training training set The data is fed into the network model and trained using a pre-defined pattern. The model parameters are continuously iterated and adjusted to enable it to effectively predict the mean standard deviation of the actual concentration based on changes in the overall standard deviation. The process is essentially as follows: ； in: This is the theoretical predicted value of the standard deviation. For network model prediction, For model weights, Input to the model, For model bias; The overall training data is divided into different batches and imported into the model. The loss function is monitored. When the loss function no longer decreases with each training round, the training process ends and a model weight parameter file is generated.

Images