A method for measuring spatial dust mass concentration based on natural sedimentation process

Abstract

The application provides a kind of space dust mass concentration measurement method based on natural sedimentation process, belong to dust concentration measurement technical field, the application is by adding experimental powder to fluidized bed dust generator and starting mixing fan, make the dust mass concentration in black box rise to target initial concentration after closing fan, make dust enter natural sedimentation state, filter membrane clamp is placed on analytical balance is connected between sampling port and air sampler, simultaneously start camera to record the balance figure continuously and start high-speed camera to photograph at equal time interval, calculate the dust mass concentration time sequence of each sampling time interval after extracting balance figure frame by frame, utilize artificial intelligence vision estimation model to output full concentration range image and dust mass concentration complete corresponding data set, solve the technical problem that dust mass concentration annotation data and image frame cannot be reliably aligned in time sequence, leading to the distortion of full concentration range image and dust mass concentration corresponding relationship.

Description

Technical Field

[0001] This invention belongs to the field of dust concentration measurement technology, and more specifically, relates to a method for measuring the mass concentration of spatial dust based on natural settling processes. Background Technology

[0002] Spatial measurement of dust mass concentration is a core requirement for safety monitoring in dust-hazardous scenarios such as mining, tunnel construction, and grain storage and transportation. Existing technologies typically use filter membrane weighing to obtain reference values ​​for dust mass concentration and combine this with an optical image sensor to construct a visual estimation model for non-contact continuous measurement. In this approach, the labeled data is provided by the filter membrane weighing results during discrete sampling periods, while the image sequence is obtained by continuous shooting with a high-speed camera. This inherently leads to a mismatch in sampling rates between the two over time.

[0003] In existing technologies, the temporal resolution of filter membrane weighing sampling is much lower than the frame rate of high-speed cameras, and the analytical balance suffers from thermal drift error, making it difficult to accurately establish the correspondence between sampling interval boundaries and image frames. Under the unsteady-state condition of natural dust settling, the dust concentration monotonically decreases over time. If the alignment method fails to utilize this physical constraint, image frames are incorrectly assigned to concentration intervals that do not conform to the settling pattern, leading to distorted labeled data and consequently causing systematic bias in the visual estimation model at low concentration stages. In other words, existing technologies suffer from the technical problem of unreliable temporal alignment between dust concentration labeled data and image frames, resulting in distorted correspondence between images and dust concentrations across the entire concentration range. Summary of the Invention

[0004] In view of this, the present invention provides a spatial dust mass concentration measurement method based on natural sedimentation process, which can solve the technical problem in the prior art that the dust mass concentration labeling data and image frames cannot be reliably aligned in time, resulting in the distortion of the correspondence between the full concentration range image and the dust mass concentration.

[0005] This invention is implemented as follows: This invention provides a method for measuring the mass concentration of spatial dust based on a natural settling process, comprising the following steps:

[0006] Add the experimental powder to the fluidized bed dust generator of the dust natural settling experimental equipment, start the fluidized bed dust generator and mixing fan to raise the dust mass concentration in the black box to the target initial concentration, and then turn off all the mixing fans to allow the dust to enter the natural settling state.

[0007] Place the filter membrane in the filter membrane clamp and weigh the entire filter membrane clamp to obtain the unloaded mass. Connect the filter membrane clamp in series between the black box sampling port and the air sampler inlet. Place the entire filter membrane clamp on the analytical balance. After the analytical balance reading stabilizes, start the air sampler.

[0008] Simultaneously start the camera to continuously record the reading of the analytical balance, and simultaneously start the high-speed camera to take pictures of the spatial distribution of dust inside the black box at equal time intervals.

[0009] The analysis balance readings at each moment in the camera recording are extracted frame by frame. Combined with the sampling flow rate and cumulative sampling duration, the dust mass concentration in each sampling time interval is calculated according to the mass concentration calculation formula to obtain the dust mass concentration time series.

[0010] The dust mass concentration time series and the high-speed camera image series are input into the optimal alignment algorithm for time-series concentration images. A directed acyclic graph is constructed with the physical monotonic decay constraint of natural dust settling as a penalty term. The optimal time-series registration path is solved by dynamic programming to obtain the dust mass concentration label value and alignment confidence score corresponding to each frame image.

[0011] The low signal-to-noise ratio image frames are processed by time-series cumulative averaging and phase-sensitive demodulation. After scattering feature extraction, the data are input into the dust concentration visual estimation model. The corrected dust mass concentration value is then combined with the zero-point thermal drift compensation of the balance to output a dataset that corresponds completely to the image and dust mass concentration across the entire concentration range.

[0012] Specifically, the target initial concentration is determined by multiplying the upper limit of the target dust mass concentration by an over-generation coefficient based on the natural settling rate of dust and the range of dust mass concentration to be covered in the experiment. The over-generation coefficient is determined by measuring the settling loss of dust from the shutdown of the self-fluidized bed dust generator to the shutdown of the mixing fan in the preliminary experiment, and then taking the average value after repeated experiments.

[0013] The formula for calculating the mass concentration is as follows: ,in To determine the dust mass concentration, For reference dust mass concentration, Unloaded mass To determine the overall mass of the filter membrane clamp after sampling. For reference quality, For sampling flow rate, To accumulate sampling time, For reference volume.

[0014] Specifically, the zero-point thermal drift compensation of the balance involves arranging a multi-point platinum resistance temperature sensor array in the weighing chamber of the analytical balance, establishing a thermal compensation model of the zero point and temperature field of the analytical balance, fitting the drift curve through periodic no-load zero-point calibration interpolation, and subtracting the thermal drift component from the original weighing data in real time to obtain the corrected dust mass concentration value.

[0015] The parameters of the thermal compensation model are obtained by performing multi-point temperature calibration experiments on the analytical balance within a preset temperature range, and the mean value is taken after repeated calibration at each temperature point for fitting.

[0016] Specifically, the time-series frame cumulative averaging and phase-sensitive demodulation involves applying a preset modulation frequency to the modulated light source, averaging the continuous multi-frame images acquired by the high-speed camera at the preset modulation frequency according to the modulation phase, suppressing non-co-frequency background stray light using the lock-in amplification principle, and extracting the dust scattering signal.

[0017] The preset modulation frequency is determined by measuring the frequency distribution of background stray light in a pre-experiment and selecting a value within the frequency range with the lowest background stray light energy. The number of frames accumulated over multiple frames is determined by measuring the change curve of image signal-to-noise ratio with the number of accumulated frames under the lowest target dust mass concentration condition in the experiment, and the minimum number of accumulated frames corresponding to an image signal-to-noise ratio not lower than the signal-to-noise ratio threshold is used.

[0018] Specifically, the time-series concentration image optimal alignment algorithm constructs a two-dimensional directed acyclic graph with sampling time as the row and captured frame as the column. The weight of the node in the graph is defined as the weighted Euclidean distance between the dust mass concentration value calculated in the sampling interval and the dust mass concentration estimated by the image scattering feature vector. The penalty coefficient is increased for edges that violate the physical monotonic decay constraint of natural dust settling. Dynamic programming is used to solve the shortest path of the directed acyclic graph to obtain the optimal dust mass concentration interval interpolation value and alignment confidence score corresponding to each frame image.

[0019] The penalty coefficient is determined by conducting pre-experiments under no less than 10 different initial dust mass concentration conditions for different targets, using the weight expansion ratio corresponding to the maximum error as the lower bound of the penalty coefficient, and the weight expansion ratio that causes the path search to jump beyond the jump threshold number of sampling intervals as the upper bound of the penalty coefficient.

[0020] The weight vector of the weighted Euclidean distance is determined by calculating the Pearson correlation coefficient between each scattering feature component and the dust mass concentration on the labeled dataset, normalizing the absolute value of the correlation coefficient, and taking the average value after cross-validation.

[0021] The dust concentration visual estimation model consists of two parts: a physically forward rendering module and a normalized flow inverse estimation module, which are coupled through a physical verification jump layer. The physically forward rendering module takes the dust mass concentration scalar and particle size distribution vector as input, embeds a Mie scattering coefficient lookup table and a Bell-Lambert extinction integral operator, and outputs the theoretically predicted image grayscale field. The normalized flow inverse estimation module adopts an improved Glow architecture, calculates the posterior probability density from the observed image through layer-by-layer reversible transformation, and samples the posterior distribution of dust mass concentration.

[0022] The physical verification jump layer is inserted between every two reversible affine coupling blocks to calculate the theoretically predicted image grayscale field and the observed image grayscale field corresponding to the latent variable after the current stream transformation. If the distance exceeds a preset verification threshold, the process jumps back to the previous flow layer and re-transforms. The upper limit of internal iterations is the threshold number of iterations. The preset verification threshold is taken as the distance between correctly aligned samples on the labeled dataset. The corresponding value of the quantile threshold of the distance distribution.

[0023] The training dataset for the dust concentration visual estimation model consists of both labeled and unlabeled samples. The labeled samples are trained under supervision using the sum of normalized flow negative log-likelihood loss and dust mass concentration posterior mean regression loss. The unlabeled samples are trained using a physically-oriented rendering module to compare the grayscale field of the theoretically predicted image with the observed image for dust mass concentration estimation. The loss function is trained under self-supervised conditions, and the two loss functions are summed with adjustable weights.

[0024] During training, a memory allocation strategy function calculates the estimated memory requirement based on the current batch image frame count, the number of reversible affine coupling block layers, and the size of the Mie scattering coefficient lookup table. Based on the ratio of the estimated memory requirement to the total available memory, the function dynamically selects a memory allocation scheme to allocate the physically forward rendering module and the normalized flow inverse estimation module to the same CUDA stream, independent CUDA streams, or allocate them to multiple CUDA streams in layers.

[0025] Among them, the over-sponge coefficient ranges from 1.2 to 1.4; the preset modulation frequency ranges from 50 to 500 Hz; the number of frames accumulated over multiple frames ranges from 8 to 64 frames; the signal-to-noise ratio threshold is 10 dB; the penalty coefficient ranges from 5 to 20; the jump threshold is 2 sampling intervals; the number of layers in the reversible affine coupling block ranges from 12 to 24; the upper limit of internal iteration is 3 times; the quantile threshold is the 95th quantile; and the ratio of labeled samples to unlabeled samples is not less than 1:3.

[0026] This invention constructs an optimal alignment algorithm for time-series concentration images, introducing the physical monotonic decay constraint of natural dust settling into the edge penalty term of a directed acyclic graph. Dynamic programming is used to force the alignment path to satisfy the physical laws of settling, ensuring physical consistency between image frames and dust mass concentration labels on the time axis. Furthermore, this invention employs zero-point thermal drift compensation to correct weighing errors and uses time-series frame cumulative averaging and phase-sensitive demodulation to improve the signal-to-noise ratio of low-concentration images, further guaranteeing the accuracy of the labeled data and reliably establishing the correspondence between images and dust mass concentrations across the entire concentration range. In summary, this invention solves the technical problem mentioned in the background art where dust mass concentration labeling data and image frames cannot be reliably aligned in time, leading to distortion of the correspondence between images and dust mass concentrations across the entire concentration range. Attached Figure Description

[0027] Figure 1 This is a flowchart of the method of the present invention.

[0028] Figure 2 This is a schematic diagram of the components of the dust natural settling experimental equipment.

[0029] Figure 3 The distribution of alignment confidence scores for the optimal alignment algorithm for time-series concentration images at different concentration stages is shown in the figure.

[0030] Figure 4 This is a graph showing the error distribution of the posterior mean estimation for the full concentration range of the visual dust concentration estimation model. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0032] like Figure 1 The diagram shown is a flowchart of a spatial dust mass concentration measurement method based on natural sedimentation process provided by the present invention. This method includes the following steps:

[0033] S01. Add the experimental powder to the fluidized bed dust generator of the dust natural settling test equipment, start the fluidized bed dust generator and the mixing fan, so that the dust mass concentration in the black box rises to the target initial concentration, and then turn off all the mixing fans to allow the dust to enter the natural settling state.

[0034] S02. Place the filter membrane in the filter membrane clamp and weigh the entire filter membrane clamp to obtain the unloaded mass. Connect the filter membrane clamp in series between the black box sampling port and the air sampler inlet. Place the entire filter membrane clamp on the analytical balance. Start the air sampler after the analytical balance reading stabilizes.

[0035] S03. Simultaneously start the camera to continuously record the reading of the analytical balance, and simultaneously start the high-speed camera to take pictures of the spatial distribution of dust in the black box at equal time intervals.

[0036] S04. Extract the analysis balance readings at each moment in the camera recording frame by frame, combine the sampling flow rate and cumulative sampling duration, and calculate the dust mass concentration in each sampling time interval according to the mass concentration calculation formula to obtain the dust mass concentration time series.

[0037] S05. Input the dust mass concentration time series and the high-speed camera image sequence into the optimal alignment algorithm of the time series concentration image. Construct a directed acyclic graph with the physical monotonic decay constraint of natural dust settling as the penalty term. Solve the optimal time series registration path through dynamic programming to obtain the dust mass concentration label value and alignment confidence score corresponding to each frame image.

[0038] S06. Perform time-series frame cumulative averaging and phase-sensitive demodulation on low signal-to-noise ratio image frames. After scattering feature extraction, input the data into the dust concentration visual estimation model. Combine the corrected dust mass concentration value after zero-point thermal drift compensation of the balance, and output the image and dust mass concentration complete corresponding dataset for the entire concentration range.

[0039] The target initial concentration is determined as follows: based on the natural settling rate of dust and the range of dust mass concentration to be covered in the experiment, the upper limit of the target dust mass concentration is multiplied by an over-generation coefficient; the value of the over-generation coefficient is in the range of 1.2 to 1.4, and is determined by measuring the settling loss of dust from the shutdown of the self-fluidized bed dust generator to the shutdown of the mixing fan in the preliminary experiment, and taking the average value after no less than 5 repeated experiments.

[0040] The unloaded mass is the total mass of the filter membrane and the filter membrane clamp, denoted as . The unit is The sampling flow rate of the air sampler is denoted as . The unit is The cumulative sampling duration is recorded as follows: The unit is The formula for calculating the mass concentration is as follows: ;in The desired dust mass concentration is expressed in units of... ; For reference dust mass concentration, the unit is... ; The mass of the filter membrane clamp after sampling is expressed in units of... ; For reference quality, the unit is... ; For reference volume, the unit is... .

[0041] The specific method for zero-point thermal drift compensation of the balance is as follows: a multi-point platinum resistance temperature sensor array is arranged in the weighing chamber of the analytical balance to establish a thermal compensation model of the zero point and temperature field of the analytical balance; the parameters of the thermal compensation model are obtained by performing multi-point temperature calibration experiments on the analytical balance in the range of 0 to 50℃ with a step size of 1℃, and the average value is taken after each temperature point is calibrated at least 5 times; the drift curve is fitted by interpolation through periodic no-load zero-point calibration, and the thermal drift component is subtracted from the original weighing data in real time to obtain the corrected dust mass concentration value; the corrected dust mass concentration value is used to input the dust concentration visual estimation model in step S06.

[0042] The specific implementation of the time-series frame cumulative averaging and phase-sensitive demodulation is as follows: A periodic intensity change at a preset modulation frequency is applied to the modulated light source. The preset modulation frequency ranges from 50 to 500 Hz. The frequency distribution of background stray light is measured through pre-experiments, and the value within the frequency range with the lowest background stray light energy is selected as the preset modulation frequency. Multiple consecutive frames of images acquired by the high-speed camera at the preset modulation frequency are grouped by modulation phase and averaged. The non-co-frequency background stray light is suppressed using the lock-in amplification principle, and the dust scattering signal is extracted. The number of frames accumulated is in the range of 8 to 64 frames. The actual value is determined by measuring the change curve of the image signal-to-noise ratio with the number of accumulated frames under the lowest target dust mass concentration condition in the experiment, with the minimum accumulated frame number corresponding to an image signal-to-noise ratio of not less than 10 dB. The output of the time-series frame cumulative averaging and phase-sensitive demodulation is an enhanced image of the low signal-to-noise ratio image frame, used for scattering feature extraction.

[0043] The principle and specific implementation of the Time-Series Concentration Image Optimal Alignment Algorithm (Graph-DTW-CAM) are as follows: A two-dimensional directed acyclic graph is constructed with sampling times as rows and captured frames as columns. The nodes in the graph... The weight is defined as the first Dust mass concentration value calculated from the sampling interval With the The weighted Euclidean distance between dust mass concentrations is estimated using the scattering feature vectors of frame images; for edges where the rate of change of dust mass concentration within the time span of adjacent nodes violates the physical monotonic decay constraint of natural dust settling, the penalty coefficient is increased. The The value range is 5 to 20. Preliminary experiments were conducted under at least 10 different initial dust mass concentrations. The actual alignment error distribution corresponding to nodes violating the monotonic decay constraint in each experiment was statistically analyzed. The weight amplification ratio corresponding to the maximum error was used as the... The lower bound is used as the weight amplification ratio that causes the path search to jump more than 2 sampling intervals. The upper bound is determined; dynamic programming is used to solve the shortest path in the directed acyclic graph, and an improved dynamic time warping algorithm is applied to temporal alignment to obtain the optimal dust mass concentration interval interpolation value and alignment confidence score for each frame image; the alignment confidence score is used to filter out low-quality corresponding points in step S06; the algorithm uses... Achieving GPU parallel acceleration through matrix dynamic programming. The number of sampling intervals. This represents the number of image frames.

[0044] Graph-DTW-CAM transforms the temporal alignment problem into a shortest path search problem with physical constraints. By introducing a monotonic decay penalty due to natural dust settling into the dynamic programming path, it forces the aligned path to satisfy the physical law of decreasing dust concentration over time, ensuring physical consistency between dust concentration labels and image frames on the time axis. Alignment confidence scores are used to subsequently filter out low-quality corresponding points, thus guaranteeing the temporal reliability of the labeled data across the entire concentration range. The algorithm complexity is O(log n). A matrix dynamic programming implementation suitable for GPU parallel acceleration, supporting real-time streaming processing.

[0045] The weight vector of the weighted Euclidean distance is determined as follows: the Pearson correlation coefficient between each scattering feature component and the dust mass concentration is calculated on the labeled dataset, and the absolute value of the correlation coefficient is normalized as the weight corresponding to each feature component. After 5-fold cross-validation, the mean value is taken to determine the final weight vector.

[0046] The specific structure of the dust concentration visual estimation model (Dust-CNF-PhySim model) is as follows: the model consists of two parts: a physically-oriented forward rendering module and a normalized flow inverse estimation module, which are coupled through a physical verification jump layer. The physically-oriented forward rendering module takes the dust mass concentration scalar and particle size distribution vector as input, and embeds a Mie scattering coefficient lookup table and a Beer-Lambert extinction integral operator, both of which are implemented as differentiable operators to support gradient backpropagation, and outputs the theoretically predicted image grayscale field; the Mie scattering coefficient lookup table covers a particle size range of 1 to 100 μm, discretized with a step size of 0.5 μm, and the Mie scattering coefficients corresponding to each particle size are stored after pre-calculation for lookup during training and inference. The normalized flow inverse estimation module employs an improved Glow architecture, consisting of 12–24 stacked reversible affine coupling blocks. Each reversible affine coupling block contains three layers of convolutional feature extraction subnetworks with 64–256 feature channels. Starting from the observed image, it accurately calculates the posterior probability density through layer-by-layer reversible transformation, sampling to obtain the posterior distribution of dust mass concentration. The number of layers in the reversible affine coupling block is determined by training 12, 16, 20, and 24 layers on the validation set, using the minimum posterior mean error of dust mass concentration as the criterion. A physical verification jump layer is inserted between every two reversible affine coupling blocks to calculate the difference between the theoretically predicted image grayscale field corresponding to the latent variable after the current flow transformation and the observed image. If the distance exceeds a preset verification threshold, the process jumps back to the previous flow layer and re-transforms, with an internal iteration limit of 3 times. The preset verification threshold is determined by statistically analyzing the correctly aligned samples on the labeled dataset. The distance distribution is used, and its 95th percentile is taken as the preset verification threshold.

[0047] The VRAM-Alloc function allocates memory based on the number of frames in the current batch of images. Number of reversible affine coupling block layers in the normalized flow inverse estimation module and the size of the Mie scattering coefficient lookup table Calculate the estimated video memory requirement. The estimated video memory requirement The calculation formula is: ;in , , As weighting coefficients, through different , , The actual memory usage under the combined configuration was obtained by performing least squares regression fitting. , , , For the corresponding reference value; when When, the physically-based forward rendering module and the normalized flow inverse estimation module are assigned to the same CUDA flow; when When the physical forward rendering module and the normalized flow inverse estimation module are assigned to independent CUDA streams and gradient checkpointing is enabled to reuse intermediate active memory; when At that time, the normalized flow inverse estimation module is allocated to multiple CUDA streams in layer-by-layer segments, and the Mie scattering coefficient lookup table is migrated to memory for asynchronous prefetching by the physical forward rendering module; This represents the total available video memory of the current GPU.

[0048] The specific steps for establishing the training dataset of the Dust-CNF-PhySim model include: conducting experiments using a dust natural settling experimental device under 5-10 different initial dust mass concentration conditions, collecting high-speed camera image sequences and corresponding dust mass concentration labels to form labeled image-dust mass concentration pairs; for image frames that cannot be fully labeled, only the image sequences are saved as unlabeled samples; after performing time-series frame cumulative averaging and phase-sensitive demodulation on all image frames, scattering feature vectors are extracted to form scattering feature image pairs; the final dataset consists of labeled and unlabeled samples, with a ratio of labeled to unlabeled samples of no less than 1:3.

[0049] The specific training steps of the Dust-CNF-PhySim model include: training with labeled samples using a supervised loss function, wherein the supervised loss function is the sum of the normalized flow negative log-likelihood loss and the posterior mean regression loss of dust mass concentration; and training with unlabeled samples using a self-supervised Mie rendering consistency loss function, wherein the self-supervised Mie rendering consistency loss function is the difference between the theoretically predicted image grayscale field of the dust mass concentration estimated by the model and the observed image obtained by the physically forward rendering module. Loss; two types of loss functions are summed with adjustable weights. In the early stages of training, the supervised loss function is dominant, and the weights of the self-supervised Mie rendering consistency loss function are gradually increased with each training epoch. The weight adjustment strategy is determined by the convergence curve of the dust concentration estimation error in the validation set. During training, the VRAM-Alloc function is used to dynamically adjust the memory allocation scheme based on the number of frames in the current batch of images. Number of reversible affine coupling block layers and the size of the Mie scattering coefficient lookup table Allocate CUDA streams and memory resources in real time.

[0050] The Dust-CNF-PhySim model treats the reconstruction of the dust mass concentration field as an inverse Bayesian problem. Through a physical forward rendering module, it explicitly couples dust mass concentration with the physical laws of Mie scattering, making the model's forward predictions physically interpretable. A normalized flow inverse estimation module calculates the posterior distribution of dust mass concentration using precise posterior probability density calculations, quantifying the uncertainty in the low-concentration stage. A physical verification jump layer applies physical consistency constraints in real-time during the flow transformation process, preventing inverse estimation from deviating from the physically feasible region. A semi-supervised training strategy fully utilizes a large number of unlabeled samples, alleviating the model's insufficient generalization caused by the scarcity of labeled samples in the low dust mass concentration stage, making the mapping relationship estimation more accurate and reliable across the entire concentration range.

[0051] The composition and connection of the dust natural settling experimental equipment are as follows: it includes one black box, one fluidized bed dust generator, eight mixing fans, one high-speed camera, one modulated light source, one air sampler, one filter membrane clamp, one filter membrane, one camera, and one analytical balance. A background plate is installed on the inner wall of the black box. The top of the black box is connected to the dust outlet pipe of the fluidized bed dust generator. The eight mixing fans are evenly arranged around the inner wall of the black box. A sampling port is provided on the black box. The sampling port is connected in series with the filter membrane clamp and the air sampler through a pipe. The filter membrane is placed inside the filter membrane clamp, and the entire filter membrane clamp is placed on the weighing platform of the analytical balance. The modulated light source is installed inside the black box. The high-speed camera is installed on the outer wall of the black box and aimed at the internal viewing window. The camera is positioned directly in front of the analytical balance display. The power control lines of the fluidized bed dust generator and the mixing fans are both connected to an external controller.

[0052] Optionally, the present invention also provides a method for implementing a spatial dust mass concentration measurement system based on a natural settling process using a computer. The computer is equipped with a readable storage medium that stores program instructions, which can execute the above-described method when the computer is run.

[0053] The specific implementation of step S01 is as follows: Technicians add the experimental powder to the fluidized bed dust generator, and after starting the fluidized bed dust generator, simultaneously start the fans on the inner wall of the black box (or multiple mixing fans evenly arranged around the inner wall of the black box, preferably 8 fans), so that the dust in the black box can be rapidly dispersed and reach a uniform distribution. The determination of the target initial concentration depends on the over-generation coefficient, which is determined by averaging no less than 5 pre-experiments, with a value range of 1.2 to 1.4. Its physical meaning is to compensate for the concentration loss caused by gravity settling during the period from when the dust generator stops to when the mixing fans are completely shut down, ensuring that the dust mass concentration in the black box is not lower than the target upper limit when the mixing fans are shut down. When the analytical balance reading confirms that the concentration in the black box has reached the target initial concentration, all mixing fans are simultaneously shut down, and the dust enters a state of pure natural settling, providing a prerequisite guarantee for the physical rationality of the subsequent time-series monotonic decay constraint.

[0054] The specific implementation of step S02 is as follows: After the filter membrane is placed in the filter membrane holder, the entire assembly is precisely weighed, and the unloaded mass is recorded. The unit is mg. Connect the filter membrane clip in series between the sampling port of the black box and the air inlet of the air sampler via a conduit, ensuring that all the sampling airflow passes through the filter membrane. Simultaneously, place the entire filter membrane clip stably on the weighing platform of the analytical balance, allowing the real-time mass of the filter membrane clip to be continuously read by the analytical balance during air sampling. The air sampler can only be started after the analytical balance reading stabilizes, and the sampling flow rate is set accordingly. The unit is L / min, to ensure the stability of the baseline for subsequent mass concentration calculations.

[0055] The specific implementation of step S03 is as follows: A camera is mounted directly in front of the analytical balance display to continuously record the balance readings, ensuring that the weighing data at each moment is recorded for subsequent frame-by-frame extraction. A high-speed camera is installed on the outer wall of the black box and aimed at the internal viewing window, capturing images of the spatial distribution of dust inside the black box at equal time intervals to form an image frame sequence. A modulated light source is installed inside the black box to provide controlled illumination for the high-speed camera. Its modulation frequency is selected from the lowest energy frequency range after pre-experimental measurement of the background stray light frequency distribution, with a value range of 50–500 Hz, to ensure optimal suppression of subsequent phase-sensitive demodulation.

[0056] The specific implementation of step S04 is as follows: digital recognition is performed on each frame of the image recorded by the camera, the balance reading is extracted and analyzed frame by frame, and the overall mass of the sampled filter membrane clip is obtained. The sequence of changes over time. Based on the mass concentration calculation formula. With no-load mass Sampling flow rate and cumulative sampling time Using the input as input, calculate the dust mass concentration corresponding to each sampling time interval. The unit is This forms a time series of dust mass concentration. The physical meaning of this formula is to normalize the ratio of filter membrane weight gain to sampling volume to a standard reference quantity, ensuring dimensional consistency and calculation traceability.

[0057] The specific implementation of step S05 is as follows: The optimal alignment algorithm for time-series concentration images constructs a two-dimensional directed acyclic graph with sampling times as rows and captured frames as columns, and nodes... The weight is the first Dust mass concentration value in sampling interval With the The weighted Euclidean distance between dust mass concentrations is estimated using the scattering feature vectors of frame images. The weight vectors are determined by normalizing the absolute values ​​of the Pearson correlation coefficients between each scattering feature component and the dust mass concentration, and then averaging the results through 5-fold cross-validation. For edges where the rate of concentration change between adjacent nodes violates the monotonic decay constraint of natural dust settling, the penalty coefficient is increased. The value ranges from 5 to 20, and the upper and lower bounds are determined through statistical analysis of no fewer than 10 sets of preliminary experiments. Dynamic programming is used in... Solving for the shortest path on the matrix, an improved dynamic time warping algorithm is applied to temporal alignment to obtain the optimal dust mass concentration interval interpolation value and alignment confidence score for each frame. The algorithm supports GPU parallel acceleration and has a complexity of O(n log n). , The number of sampling intervals. This represents the number of image frames.

[0058] The specific implementation of step S06 is as follows: For low signal-to-noise ratio image frames, time-series frame cumulative averaging processing is first performed. Multiple consecutive frames of images acquired by the high-speed camera at a preset modulation frequency are grouped by modulation phase and then averaged. The lock-in amplification principle is used to suppress non-co-frequency background stray light. The cumulative frame count ranges from 8 to 64 frames, with the minimum cumulative frame count corresponding to an image signal-to-noise ratio of not less than 10dB being the actual value. After the above processing, the dust scattering feature vector is extracted and input into the dust concentration visual estimation model along with the corrected dust mass concentration value obtained from the zero-point thermal drift compensation of the balance. Thermal drift compensation is achieved by establishing a thermal compensation model using a multi-point platinum resistance temperature sensor array. Multi-point temperature calibration is performed in the range of 0–50℃ with a step size of 1℃. Each point is repeated at least 5 times, and the average value is used for fitting. After periodic no-load zero-point calibration, the drift curve is interpolated and fitted, and the thermal drift component is subtracted from the original weighing data in real time. The dust concentration visual estimation model outputs the theoretically predicted image grayscale field using a physically-based forward rendering module, and outputs the posterior distribution of dust mass concentration using a normalized flow inverse estimation module. A physical validation jump layer performs real-time validation between every two reversible affine coupling blocks. If the distance exceeds a preset verification threshold, the algorithm will jump back and recalculate. The maximum number of iterations is 3. The preset verification threshold is taken from the correctly aligned samples on the labeled dataset. The 95th percentile of the distance distribution is used to finally output an image of the full concentration range that corresponds completely to the dust mass concentration dataset.

[0059] It should be noted that the key technologies of this invention include: the optimal alignment algorithm for time-series concentration images embeds the physical monotonic decay constraint of natural dust settling into the edge penalty term of the directed acyclic graph, enabling dynamic programming to automatically avoid alignments that violate the settling law during global path search, thereby eliminating the labeling chaos caused by the traditional dynamic time warping algorithm's failure to distinguish the temporal physical directionality; the balance zero-point thermal drift compensation combines a temperature sensor array with a thermal compensation model to deduct the thermal drift component in real time, preventing the weighing benchmark error from accumulating over experimental time and ensuring the accuracy of the dust mass concentration reference value; the time-series frame cumulative averaging processing and phase-sensitive demodulation utilize the lock-in amplification principle to separate background stray light from the dust scattering signal, ensuring the distinguishability of the scattering characteristics of the image in the low-concentration stage. With the synergistic effect of these three technologies, the temporal reliability of the labeled data, the accuracy of the reference value, and the quality of image features are simultaneously guaranteed, enabling the dust concentration visual estimation model to establish a physically consistent mapping relationship between the image and the dust mass concentration across the entire concentration range. Compared to existing methods that rely solely on a single weighing or image processing method, this approach offers stronger systemic assurance capabilities.

[0060] It should be noted that in experimental conditions where the natural dust settling process covers a wide concentration range and the low-concentration phase lasts for a long time, the number of labeled samples in the low-concentration phase is far less than that in the high-concentration phase. This results in a significant deficiency in the generalization ability of the dust concentration visual estimation model in the low-concentration range, and the model's estimation error for low-concentration images is systematically large. The reason for this technical problem is that the natural dust settling rate is faster in the high-concentration phase, corresponding to a relatively large number of sampling intervals. However, in the low-concentration phase, the settling rate tends to be slower, the concentration change per unit time is small, the appearance differences between images are subtle, and the discriminative power of the scattering feature vectors is reduced. If the labeled samples are scarce, the model struggles to learn the feature distribution patterns of the low-concentration phase from a limited number of samples. The usual solution to this technical problem is to increase the number of experiments to expand the labeled samples in the low-concentration phase. However, since the images in the low-concentration phase themselves have a low signal-to-noise ratio in each experiment, effective feature extraction is difficult. Simply increasing the number of labeled samples cannot fundamentally solve the problem of insufficient generalization of the model in the low-concentration phase. In addition, some methods use data augmentation to expand low-concentration images, but data augmentation cannot introduce new physical constraints, and the generated samples deviate from the physical distribution of real low-concentration scattering images, which may even exacerbate the estimation bias of the model. This invention effectively solves this technical problem by introducing a physically forward rendering module into the dust concentration visual estimation model. Through a Mie scattering coefficient lookup table and the Bell-Lambert extinction integral operator, the dust mass concentration is explicitly coupled with the physical laws of optical scattering, so that the model's estimation in the low-concentration stage is constrained by physical laws and does not completely depend on the number of labeled samples. Simultaneously, a semi-supervised training strategy is adopted, using a large number of unlabeled low-concentration images for training through a self-supervised Mie rendering consistency loss function. This allows the model to learn the scattering feature distribution in the low-concentration stage from the physical rendering error signals of unlabeled samples, alleviating the problem of insufficient generalization caused by the scarcity of labeled samples, and making the mapping relationship estimation across the entire concentration range more accurate and reliable.

[0061] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the above-mentioned technical problems is that the optimal alignment algorithm for time-series concentration images transforms the alignment problem into a shortest path search problem with physical constraints, thereby eliminating the possibility of alignment that violates the dust settling law from the algorithm structure level.

[0062] Specifically, after the mixing fan is turned off, dust enters a natural settling state, and its mass concentration decreases monotonically over time—a deterministic physical law. Traditional dynamic time warping algorithms do not distinguish the physical directionality of time sequences during time alignment, allowing jumps that violate monotonicity on the time axis. Therefore, for scenarios with strong physical constraints, such as dust settling, traditional algorithms may incorrectly map high-frame-rate images to sampling intervals with higher concentrations, causing annotation confusion. This invention applies a penalty coefficient to edges that violate the monotonic decay constraint in the edge weights of the directed acyclic graph, enabling dynamic programming to automatically avoid physically unreasonable alignments during the global optimal path search, fundamentally ensuring the physical consistency of the aligned path.

[0063] The introduction of zero-point thermal drift compensation for the balance solves the problem of weighing reference drift. Analysis shows that zero-point drift occurs during long-term operation of the balance due to changes in ambient temperature. Without compensation, the calculated weight gain of the filter membrane will introduce systematic errors, leading to deviations in the calculated dust mass concentration across different sampling intervals. This invention establishes a thermal compensation model using a multi-point platinum resistance temperature sensor array, subtracting the thermal drift component from the original weighing data in real time, thus ensuring the accuracy of the dust mass concentration reference value.

[0064] The introduction of temporal frame cumulative averaging and phase-sensitive demodulation solves the problem of insufficient signal-to-noise ratio in low-concentration images. Low-concentration dust scatters light very weakly, and background stray light has a significant drowning effect on the scattered signal. This invention applies periodic intensity modulation to a modulated light source, uses lock-in amplification to suppress non-co-frequency background stray light, and simultaneously performs cumulative averaging of multiple frames according to their modulation phase, enabling the scattered signal to be effectively extracted from the noise. This processing gives the low-concentration images sufficient feature discriminability, providing reliable input for scattering feature extraction.

[0065] With the synergistic effect of these three factors, this invention simultaneously ensures the reliability of the labeled data across the entire concentration range from three dimensions: reference value accuracy, image quality, and temporal alignment. This enables the dust concentration visual estimation model to establish an accurate mapping relationship between images and dust mass concentration across the entire concentration range.

[0066] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0067] The specific implementation of step S01 is as follows: Add the experimental powder to the fluidized bed dust generator, start the fluidized bed dust generator and mixing fans, and after the dust concentration in the black box reaches the target initial concentration, turn off all mixing fans to allow the dust to settle naturally. Target initial concentration The formula for determining it is expressed as follows:

[0068] ;

[0069] In the formula, The initial concentration of the target, in units of ; This represents the upper limit of the target dust concentration to be covered in the experiment, in units of... ; For reference dust mass concentration, the unit is... ; The over-generation coefficient is dimensionless and ranges from 1.2 to 1.4. It is determined by taking the average value of the settling loss during the period from the shutdown of the self-fluidized bed dust generator to the shutdown of the mixing fan through no less than 5 preliminary experiments.

[0070] The specific implementation of step S02 is as follows: place the filter membrane in the filter membrane holder and weigh the whole thing to obtain the empty weight. The unit is This value represents the overall mass of the filter membrane and filter membrane clamp. The filter membrane clamp is connected in series between the sampling port of the black box and the air inlet of the air sampler. The entire filter membrane clamp is placed on the analytical balance. After the reading stabilizes, the air sampler is started, and the sampling flow rate is recorded as [value missing]. The unit is The overall mass of the filter membrane clamp after sampling is recorded as follows: The unit is The cumulative sampling time is recorded as The unit is The formula for calculating dust mass concentration is as follows:

[0071] ;

[0072] In the formula, The desired dust mass concentration is expressed in units of... ; Unloaded mass, unit: ; The mass of the filter membrane clamp after sampling is expressed in units of... ; For reference quality, the unit is... ; For reference volume, the unit is... ; The dimensions are ,and Dimensions are consistent; The dimensions are ,and The dimensions are consistent, and both sides of the equal sign are dimensionless ratios.

[0073] The specific implementation of step S03 is as follows: the camera is started simultaneously to continuously record the reading of the analytical balance, and the high-speed camera is started simultaneously to take pictures of the spatial distribution of dust in the black box at equal time intervals. The two data are kept synchronized on the time axis to provide raw data input for subsequent steps.

[0074] The specific implementation of step S04 is as follows: extract the balance readings frame by frame from the camera recording, and combine them with the sampling flow rate. With cumulative sampling time The dust mass concentration at each sampling time interval is calculated based on the above formula, thus obtaining the dust mass concentration time series. ,in , This represents the total number of sampling intervals.

[0075] The specific implementation of step S05 is as follows: The dust mass concentration time series... With high-speed camera image sequences ( , The algorithm for optimal alignment of temporal density images (total number of image frames) is a graph-constrained dynamic temporal warping density alignment matching algorithm. This algorithm constructs a data structure with sampling times as rows and captured frames as columns. Directed acyclic graph, nodes The weight is defined as the first Dust mass concentration in sampling interval With the Frame image scattering feature vector Estimate the weighted Euclidean distance between dust mass concentrations. For the first The scattering feature vector obtained after scattering feature extraction of the frame image has a dimension of The node weights are obtained by calculating the output image from the time-series frame cumulative averaging and phase-sensitive demodulation through a feature extraction subnetwork. The node weight formula is expressed as follows:

[0076] ;

[0077] In the formula, For nodes The weighted Euclidean distance weights are dimensionless; Let be the dimension of the scattering eigenvectors; For the first The weights of each feature component are dimensionless. For the first Frame Image The estimated dust mass concentration corresponding to each scattering characteristic component, in units of ; and All terms are dimensionless ratios, and the summation and square root are then performed. The formula for determining it is expressed as follows:

[0078] ;

[0079] In the formula, For the first The Pearson correlation coefficients between the scattering characteristic components and the dust mass concentration are dimensionless and their mean values ​​are used to determine the final weight vector after 5-fold cross-validation. Penalty coefficients are applied to edges that violate the physical monotonic decay constraint of natural dust settling. The value ranges from 5 to 20, and is determined through no fewer than 10 sets of preliminary experiments. The formula for the edge weight with penalty is expressed as follows:

[0080] ;

[0081] In the formula, This indicates a directed acyclic graph consisting of nodes. Point to adjacent nodes The directed edges; and These are the row and column indices of the successor node, respectively. This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. It is dimensionless. Dimensionless; Dimensionless; and All quantities are dimensionless; this penalty term forces the alignment path to satisfy the physical law that dust concentration decreases with time. The dynamic programming optimal path formula is expressed as follows:

[0082] ;

[0083] In the formula, From the starting node to the node The optimal cumulative cost is dimensionless; For nodes The set of predecessor nodes. The algorithm uses... A matrix dynamic programming approach is used to achieve parallel acceleration of the graphics processing unit (GPU), with an algorithm complexity of O(n log n). Output the optimal dust mass concentration interval interpolation value and alignment confidence score for each frame of the image. . The calculation formula is expressed as follows:

[0084] ;

[0085] In the formula, Termination node The optimal cumulative cost at the location is dimensionless. and These are the minimum and maximum values ​​of the optimal cumulative cost among all terminating nodes, respectively, and are dimensionless. The larger the value, the higher the confidence level of the alignment between the image frame and the dust mass concentration label; frames below the confidence threshold are filtered out in step S06.

[0086] The specific implementation of step S06 is as follows: First, perform time-series frame cumulative averaging and phase-sensitive demodulation on the low signal-to-noise ratio image frames. Then, apply a preset modulation frequency to the modulated light source. The periodic intensity variation (range 50–500 Hz) is used to average the modulated phase groups of multiple consecutive frames acquired by a high-speed camera. Lock-in amplification is then used to suppress non-co-frequency background stray light and extract the dust scattering signal. (Cumulative frame count) The determination method is as follows: the image signal-to-noise ratio is measured under the condition of the lowest target dust mass concentration in the experiment. Follow The change curve, with The minimum corresponding to not less than 10dB The value is taken as the actual value. The value range is 8 to 64 frames. Regarding zero-point thermal drift compensation for the balance, a zero-point thermal compensation model is established by arranging a multi-point platinum resistance temperature sensor array within the weighing cavity of the balance. The formula is as follows:

[0087] ;

[0088] In the formula, To be at temperature The zero-point thermal drift mass is given in units of ; The current temperature is expressed in °C. This is a reference temperature, in °C. For the first The polynomial fitting coefficients are dimensionless and are obtained by fitting the average value of the polynomials within the range of 0 to 50℃ with a step size of 1℃ and repeated calibration at each temperature point at least 5 times. The order of the polynomial is 3, which is an empirical value. It can be adjusted appropriately based on whether the root mean square value of the fitted residuals meets the accuracy requirements of the balance. The residual term is the fitted value, in units of 1. ; left side of the equals sign Dimensionless, terms on the right side Dimensionless Dimensionless. Corrected dust mass concentration. The calculation formula is expressed as follows:

[0089] ;

[0090] In the formula, The corrected dust mass concentration is expressed in units of... ; This represents the zero-point thermal drift mass at the corresponding sampling time, in units of... The meanings of the remaining variables are the same as before; Dimensions are , divided by Dimensionless; Dimensions are , divided by The result is dimensionless; both sides of the equation are dimensionless ratios. The enhanced image, after scattering feature extraction, is compared with the corrected dust mass concentration value. The common input is the visual estimation model of dust concentration, namely the physically simulated normalized flow dust concentration estimation model, which outputs an image dataset that completely corresponds to the dust mass concentration across the entire concentration range. The physically-based forward rendering module uses dust mass concentration as the input. With particle size distribution vector For input, The particle size distribution vector represents the particle number concentration within a particle size range of 1–100 μm, discretized in 0.5 μm increments. This concentration was obtained from particle size analysis of the experimental powder prior to the experiment. The grayscale field of the image was predicted using a Mie scattering coefficient lookup table and the output theory of the Beer-Lambert extinction integral operator. The formula is expressed as follows:

[0091] ;

[0092] In the formula, For pixels The theoretically predicted grayscale value at the location is dimensionless. The reference grayscale value is dimensionless. Dimensionless; The dimensionless Mie extinction efficiency factor is derived from the particle size distribution vector. The result is obtained by weighted average after lookup table calculation; it is dimensionless. Optical path length, in units of ; For reference length, the unit is... ; Dimensionless; all factors in the exponential term are dimensionless, and both sides of the equals sign represent dimensionless ratios. The normalized flow inverse estimation module consists of 12–24 stacked reversible affine coupling blocks. The number of layers is determined by comparing on the validation set using the minimum posterior mean error of dust mass concentration as the criterion. The forward transformation formula for a reversible affine coupled block is expressed as follows:

[0093] ;

[0094] In the formula, For the first Latent variable vector, dimensionless; for The first part of the component is dimensionless; and The first The block's scale subnetwork and translation subnetwork each contain 3 layers of convolutional feature extraction subnetworks, with the number of feature channels ranging from 64 to 256; This is element-wise multiplication; For the first The latent variable vector is dimensionless; its inverse transformation formula is as follows:

[0095] ;

[0096] The physical verification jump layer is inserted between every two reversible affine coupling blocks to calculate the relationship between the theoretically predicted image grayscale field and the observed image grayscale field corresponding to the latent variables after the current stream transform. Distance, expressed by the formula as follows:

[0097] ;

[0098] In the formula, For observing the grayscale field of the image, it is dimensionless; Dimensionless for verification error; Dimensions and The same applies when both sides of the equal sign are dimensionless; when Exceeding the preset verification threshold If the process jumps back to the previous flow layer and transforms again, the internal iteration limit is 3 times. Take the correctly aligned samples from the labeled dataset The 95th quantile of the distribution, dimensionless. Estimated memory requirements. The calculation formula is expressed as follows:

[0099] ;

[0100] In the formula, This is an estimate of video memory requirements, in units of... ; Total available video memory for the current graphics processor, in units of ; This represents the number of image frames in the current batch. The reference batch image frame number; The number of reversible affine coupling block layers; For reference, the number of layers in the reversible affine coupling block; Size of the lookup table for Mie scattering coefficients; For reference lookup table size; , , The weighting coefficients are dimensionless and are determined by different... , , The actual memory usage under the combined configuration was obtained by least squares regression fitting; the left side of the equals sign Dimensionless, of the three terms on the right , , All are dimensionless. A supervised loss function is used during training. The formula is expressed as follows:

[0101] ;

[0102] In the formula, For the first The log-likelihood of the latent variables is dimensionless. This represents the number of labeled samples. For the first Corrected dust mass concentration for each labeled sample, in units of ; For the model to the first The posterior mean estimate of each sample output, in units of ; Dimensionless; divide both by The two sides of the last equal sign are dimensionless and uniform; the sum of the two terms constitutes the complete supervised loss. Self-supervised loss function. The formula is expressed as follows:

[0103] ;

[0104] In the formula, This represents the number of unlabeled samples. For the first The theoretical predicted image grayscale field of an unlabeled sample, dimensionless; The corresponding grayscale field of the observed image is dimensionless; Dimensionless; both sides of the equation are dimensionless. Total loss function The weighted summation formula is expressed as follows:

[0105] ;

[0106] In the formula, Total loss, dimensionless; The reference loss value is dimensionless. For the first The self-supervised weights corresponding to each training epoch are dimensionless and vary with the training epoch. The increase is gradual, and the specific adjustment strategy is determined by verifying the convergence curve of the dust mass concentration estimation error. and Both are dimensionless; both sides of the equal sign are dimensionless ratios.

[0107] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0108] To illustrate the specific application of the present invention, technicians built a dust natural settling experimental device and conducted a systematic experiment on a certain mining dust (median particle size of about 15 μm, particle size range of 1 to 80 μm) to verify the feasibility of the entire process of the method of the present invention, and constructed a dataset corresponding to the full concentration range of images and dust mass concentrations.

[0109] The composition of the dust natural settling test equipment is as follows: Figure 2 As shown, the internal volume of the black box is 1. The inner wall of the black box is lined with a black background panel, and the top is connected to the dust outlet pipe of the fluidized bed dust generator. Eight mixing fans are evenly distributed around the inner wall of the black box to rapidly disperse the dust into a uniform distribution. A modulated light source is installed inside the black box. The modulation frequency was determined by pre-experimental measurement of the background stray light frequency distribution, and the lowest background stray light energy was found at 120Hz. Therefore, the preset modulation frequency is set to 120Hz. A high-speed camera is installed at the viewing window on the outer wall of the black box, with a frame rate set to 500 frames / s. A sampling port is set on the black box, and the sampling port is connected in series with a filter membrane clamp and an air sampler through a pipeline. The air sampler samples the flow rate. The flow rate was set to 20 L / min. The analytical balance had an accuracy of 0.01 mg. A six-point platinum resistance temperature sensor array was arranged in the weighing chamber. The temperature calibration experiment was conducted in 1-℃ increments within the range of 0–50℃. Each temperature point was calibrated five times, and the average value was used to fit a thermal compensation model. A camera was mounted directly in front of the analytical balance display to continuously record the balance readings throughout the process.

[0110] Technicians conducted experiments at seven different initial dust concentrations, covering a target initial concentration range of 50. By 2000 The over-generation coefficient was determined to be 1.28 after averaging five preliminary experiments. The number of sampling intervals for the dust mass concentration time series in each experiment... Approximately 60, the total number of high-speed camera image frames. Approximately 18,000 frames. Penalty coefficient for the optimal alignment algorithm of temporal density images. The weights were determined to be 12 through 10 preliminary experiments. The weight vectors for the weighted Euclidean distance were determined by averaging after 5-fold cross-validation. The cumulative frame count for time-series cumulative averaging was determined by 50. The curve of the signal-to-noise ratio of the measured image under the target dust mass concentration condition was determined. The results showed that the image signal-to-noise ratio reached 12dB when the cumulative frame number was 32 frames, which met the requirement of not less than 10dB. Therefore, the actual cumulative frame number was taken as 32 frames.

[0111] After processing by the time-series concentration image optimal alignment algorithm, the corresponding image frames yielded dust mass concentration labels and alignment confidence scores for each frame. Frames with low confidence scores (below the lower quartile of the distribution) were filtered out. The resulting number of labeled samples was approximately 4200, and the number of unlabeled samples was approximately 17500, with a ratio of approximately 1:4.2, meeting the requirement of at least 1:3. The dust mass concentration label distribution in each experiment is shown in Table 1.

[0112] Table 1. Distribution of dust mass concentration in labeled samples for each target initial concentration experiment.

[0113]

[0114] The normalized flow inverse estimation module of the dust concentration visual estimation model adopts an improved Glow architecture. The number of reversible affine coupling block layers was compared by training 12, 16, 20, and 24 layers on the validation set, respectively, to determine the minimum error in the posterior mean of dust concentration. Therefore, 20 layers were chosen, and the number of feature channels was set to 128. The preset validation threshold for the physical validation jump layer was set to the correctly aligned samples on the labeled dataset. The distance distribution's 95th quantile is used, with an internal iteration limit of 3. The memory allocation strategy function during training is based on the number of image frames in the current batch. Number of reversible affine coupling block layers and the size of the Mie scattering coefficient lookup table Calculate the estimated video memory requirements ,in , , The data was obtained through least-squares regression fitting after measuring the video memory usage under different combinations. The experiment used a GPU with a total video memory of 24GB per card, and all batches were able to achieve the desired results. and The ratio relationship automatically selects the appropriate CUDA stream allocation scheme, and no memory overflow occurs.

[0115] like Figure 3 As shown, the alignment confidence score distribution of the optimal alignment algorithm for time-series concentration images at different concentration stages indicates that, after introducing a monotonically decaying penalty, the alignment confidence scores for both high and low concentration stages are significantly higher than the baseline method without the penalty, especially in the low concentration stage (below 100). The improvement in alignment confidence is more significant, indicating that the effect of physical constraints on improving the quality of temporal alignment is more prominent in the low-concentration stage with a low signal-to-noise ratio.

[0116] like Figure 4As shown, the posterior mean estimation error distribution of the visual dust concentration estimation model across the entire concentration range indicates that, after adopting a semi-supervised training strategy, the accuracy of the estimation is lower in the low concentration stage (below 100%). The estimation error of the model is significantly reduced compared to the comparative method using only supervised training, indicating that the self-supervised Mie rendering consistency loss function effectively utilizes a large number of unlabeled low-concentration samples, alleviating the problem of insufficient generalization caused by the scarcity of labeled samples in the low-concentration stage. The estimation error distribution of the model across the entire concentration range is shown in Table 2.

[0117] Table 2. Error Distribution of the Visual Estimation Model for Dust Concentration Across the Entire Concentration Range

[0118]

[0119] The advancements of this invention compared to traditional methods are reflected in the following aspects: Traditional filter weighing methods can only provide the average dust concentration during the sampling period, failing to establish a frame-by-frame correspondence with image frames. This invention, however, embeds physical constraints into the alignment process through a temporal concentration image optimal alignment algorithm, ensuring that each image frame receives a physically consistent dust concentration label, fundamentally overcoming the inherent defect of missing temporal correspondence in traditional methods. Traditional optical image methods fail in the low-concentration stage due to background stray light overwhelming the scattering signal. This invention, through phase-sensitive demodulation and lock-in amplification, suppresses background stray light from a physical mechanism, effectively extracting scattering features in the low-concentration stage, rather than relying on passive suppression through post-processing filtering. Traditional visual estimation models exhibit systematic biases in the low-concentration stage due to a lack of physical constraints. This invention explicitly couples the physical laws of Mie scattering into the model's forward rendering path, subjecting the model's estimation in the low-concentration stage to active constraints from physical laws, rather than relying entirely on implicit fitting through statistical learning.

[0120] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0121] Table 3. Variable Explanation Table (Part 1)

[0122]

[0123] Table 4. Variable Explanation Table (Part Two)

[0124]

[0125] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for measuring the mass concentration of spatial dust based on natural settling processes, characterized in that, Includes the following steps: Add the experimental powder to the fluidized bed dust generator of the dust natural settling experimental equipment, start the fluidized bed dust generator and mixing fan to raise the dust mass concentration in the black box to the target initial concentration, and then turn off all the mixing fans to allow the dust to enter the natural settling state. Place the filter membrane in the filter membrane clamp and weigh the entire filter membrane clamp to obtain the unloaded mass. Connect the filter membrane clamp in series between the black box sampling port and the air sampler inlet. Place the entire filter membrane clamp on the analytical balance. After the analytical balance reading stabilizes, start the air sampler. Simultaneously start the camera to continuously record the reading of the analytical balance, and simultaneously start the high-speed camera to take pictures of the spatial distribution of dust inside the black box at equal time intervals. The analysis balance readings at each moment in the camera recording are extracted frame by frame. Combined with the sampling flow rate and cumulative sampling duration, the dust mass concentration in each sampling time interval is calculated according to the mass concentration calculation formula to obtain the dust mass concentration time series. The dust mass concentration time series and the high-speed camera image series are input into the optimal alignment algorithm for time-series concentration images. A directed acyclic graph is constructed with the physical monotonic decay constraint of natural dust settling as a penalty term. The optimal time-series registration path is solved by dynamic programming to obtain the dust mass concentration label value and alignment confidence score corresponding to each frame image. The low signal-to-noise ratio image frames are processed by time-series cumulative averaging and phase-sensitive demodulation. After scattering feature extraction, the data are input into the dust concentration visual estimation model. The corrected dust mass concentration value is then combined with the zero-point thermal drift compensation of the balance to output a dataset that corresponds completely to the image and dust mass concentration across the entire concentration range.

2. The method for measuring spatial dust mass concentration based on natural settling process according to claim 1, characterized in that, The target initial concentration is specifically determined by multiplying the upper limit of the target dust mass concentration by an over-generation coefficient based on the natural dust settling rate and the dust mass concentration range to be covered in the experiment. The over-generation coefficient is determined by measuring the settling loss of the dust from the shutdown of the self-fluidized bed dust generator to the shutdown of the mixing fan in the preliminary experiment, and then taking the average value after repeated experiments.

3. The method for measuring spatial dust mass concentration based on natural settling process according to claim 2, characterized in that, The dust concentration visual estimation model consists of two parts: a physical forward rendering module and a normalized flow inverse estimation module. The two are coupled through a physical verification jump layer. The physical forward rendering module takes the dust mass concentration scalar and particle size distribution vector as input, embeds a Mie scattering coefficient lookup table and a Bell-Lamber extinction integral operator, and outputs the theoretically predicted image gray field. The normalized flow inverse estimation module adopts an improved Glow architecture. Starting from the observed image, it calculates the posterior probability density through layer-by-layer reversible transformation and samples to obtain the posterior distribution of dust mass concentration.

4. The method for measuring spatial dust mass concentration based on natural settling process according to claim 3, characterized in that, The zero-point thermal drift compensation of the balance is specifically achieved by arranging a multi-point platinum resistance temperature sensor array in the weighing chamber of the analytical balance, establishing a thermal compensation model of the zero point and temperature field of the analytical balance, fitting the drift curve through periodic no-load zero-point calibration interpolation, and subtracting the thermal drift component from the original weighing data in real time to obtain the corrected dust mass concentration value.

5. The method for measuring spatial dust mass concentration based on natural settling process according to claim 4, characterized in that, The parameters of the thermal compensation model are obtained by performing multi-point temperature calibration experiments on the analytical balance within a preset temperature range. The average value is then used to fit the model after repeated calibration at each temperature point.

6. The method for measuring spatial dust mass concentration based on natural settling process according to claim 5, characterized in that, The time-series frame cumulative averaging and phase-sensitive demodulation specifically involves applying a preset modulation frequency to the modulated light source with periodic intensity changes, averaging multiple consecutive frames of images acquired by the high-speed camera at the preset modulation frequency according to the modulation phase, using the lock-in amplification principle to suppress non-co-frequency background stray light, and extracting dust scattering signals.

7. The method for measuring spatial dust mass concentration based on natural settling process according to claim 6, characterized in that, The preset modulation frequency is determined by measuring the frequency distribution of background stray light in a pre-experiment and selecting a value within the frequency range with the lowest background stray light energy. The number of frames accumulated over multiple frames is determined by measuring the change curve of image signal-to-noise ratio with the number of accumulated frames under the lowest target dust mass concentration condition in the experiment, and the minimum number of accumulated frames corresponding to an image signal-to-noise ratio not lower than the signal-to-noise ratio threshold is used.

8. The method for measuring spatial dust mass concentration based on natural settling process according to claim 7, characterized in that, The optimal alignment algorithm for time-series concentration images specifically constructs a two-dimensional directed acyclic graph with sampling times as rows and captured frames as columns. The weight of each node in the graph is defined as the weighted Euclidean distance between the dust mass concentration value calculated in the sampling interval and the dust mass concentration estimated by the image scattering feature vector. The penalty coefficient is increased for edges that violate the physical monotonic decay constraint of natural dust settling. Dynamic programming is used to solve for the shortest path in the directed acyclic graph to obtain the optimal dust mass concentration interval interpolation value and alignment confidence score for each frame image.

9. The method for measuring spatial dust mass concentration based on natural settling process according to claim 8, characterized in that, The value of the penalty coefficient is determined by conducting pre-experiments under no less than 10 different initial dust mass concentration conditions for different targets. The lower bound of the penalty coefficient is the weight expansion ratio corresponding to the maximum error, and the upper bound of the penalty coefficient is the weight expansion ratio that causes the path search to jump more than the jump threshold number of sampling intervals.

10. The method for measuring spatial dust mass concentration based on natural settling process according to claim 9, characterized in that, The weight vector of the weighted Euclidean distance is calculated by taking the Pearson correlation coefficient between each scattering feature component and the dust mass concentration on the labeled dataset, normalizing the absolute value of the correlation coefficient, and taking the average value after cross-validation.

Images