Dust concentration inversion method based on multi-modal feature level fusion

Abstract

The invention relates to a dust concentration inversion method based on multi-modal feature level fusion, belongs to the technical field of dust monitoring, and aims to solve the problem that the traditional light scattering method is easily interfered by temperature, humidity and ambient light, so that the measurement precision is insufficient. According to the technical scheme, light scattering signals and environment parameters are synchronously collected; local features of the scattering signals are extracted through a convolutional neural network; a Transform self-attention mechanism is adopted to capture global features; fusing the optical features and the environment features; and inverting the dust concentration based on the full-connection network. According to the method, environmental interference is effectively inhibited, and the precision, robustness and environmental adaptability of dust concentration measurement are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of dust monitoring and relates to a dust concentration inversion method based on multi-modal feature-level fusion. Background Art

[0002] Dust pollution is a common safety hazard in industrial production processes, especially in closed environments such as underground coal mines. High dust concentrations not only endanger human health but can also cause explosions. Therefore, real-time and accurate monitoring of dust concentrations is crucial. Traditional detection methods such as light scattering, weighing, and beta-ray methods all have significant limitations in practical applications. Light scattering is widely used for online monitoring due to its fast response speed, but its measurement accuracy is easily affected by environmental factors: humidity changes can cause particles to absorb moisture and expand, changing the light scattering characteristics; temperature fluctuations affect sensor stability; ambient light noise interferes with signal acquisition; and uneven dust particle size distribution causes scattering model mismatch. These interferences lead to systematic deviations in measurement results, making it difficult to meet the needs of high-precision monitoring.

[0003] Existing dust concentration detection methods based on single optical sensors generally suffer from poor environmental adaptability, and their measurement accuracy fluctuates significantly with operating conditions. More critically, existing methods fail to effectively leverage the complementary nature of multi-source data. For example, environmental parameter sensors can capture information on interfering factors such as temperature and humidity, while optical sensors can capture particle scattering characteristics. However, these two data sources are processed independently or simply superimposed, lacking a deep feature fusion mechanism. This data fragmentation prevents the system from adaptively compensating for environmental interference, resulting in severe deficiencies in the robustness and accuracy of dust concentration inversion results.

[0004] Therefore, there is an urgent need for a method that can collaboratively integrate multimodal data features, mine the implicit correlation between optical signals and environmental parameters through deep learning models, fundamentally overcome the defects of traditional detection technology, and improve the accuracy and adaptability of dust concentration inversion in complex industrial scenarios. Summary of the Invention

[0005] In view of this, the object of the present invention is to provide a dust concentration inversion method based on multimodal feature-level fusion.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A dust concentration inversion method based on multimodal feature-level fusion includes the following steps:

[0008] S1: Data acquisition and preprocessing: synchronously collect dust light scattering signal data, environmental parameter data and corresponding real dust concentration data;

[0009] S2: Extraction of local spatial features of light scattering signals: The light scattering signals from S1 are input into a convolutional neural network (CNN), and the local spatial features of the scattered light signals are extracted through multi-layer convolution operations;

[0010] S3: Global spatial feature extraction of light scattering signals: The local spatial features output by S2 are input into the Transformer module, and the global spatial features of the scattered light signal are extracted through the multi-head self-attention mechanism;

[0011] S4: Multimodal feature-level fusion: The light scattering global spatial features output by S3 are concatenated with the environmental parameter features, and the concatenated feature vectors are subjected to linear transformation and activation function processing to obtain fused features.

[0012] S5: Establishment of concentration inversion model: The fusion features of S4 are input into the fully connected neural network to output the dust concentration prediction value. The model parameters are optimized through back propagation based on the error between the prediction value and the actual dust concentration.

[0013] S6: Concentration inversion: Input the real-time collected light scattering signal and environmental parameters into the trained model and output the current dust concentration value.

[0014] Furthermore, in said S1:

[0015] The light scattering signal data is the multi-angle scattered light intensity data matrix I (i) =[I(λ,θ1),I(λ,θ2),...,I(λ,θ M )], where λ represents the wavelength of the light source, θ j represents the scattering angle, j = 1, 2, ..., M, where M is the number of scattering angles;

[0016] The environmental parameter data includes at least one of temperature T(i), humidity H(i), wind speed V(i), and wind direction D(i).

[0017] Furthermore, in S2, the calculation process of CNN satisfies:

[0018] H (l) =ReLU(W (l) *H (l-1) +b (l) )

[0019] Among them H (l) is the output feature map of the lth layer, W (l) is the convolution kernel weight, b (l) is the bias term, representing the convolution operation, and ReLU(·) is the activation function.

[0020] Furthermore, in S3, the calculation of the multi-head self-attention mechanism satisfies:

[0021]

[0022] Z att =[head1,head2,...,head H ]W O ,

[0023] in is the output of the last layer of CNN, is the mapping matrix of the h-th attention head, and H is the number of attention heads.

[0024] Furthermore, in S4, the calculation of feature fusion satisfies:

[0025] F fusion =σ(W f [F scatt ; F env ]+b f )

[0026] Among them F scatt is the global spatial characteristic of light scattering, F env is the environmental parameter feature, [;] represents vector splicing, W f is the weight matrix, b f is the bias term, and σ(·) is the activation function.

[0027] Furthermore, in S5, the output of the fully connected neural network satisfies:

[0028]

[0029] where h (L-1) is the output of the last hidden layer, W (L) and b (L) are the output layer parameters.

[0030] A dust concentration inversion system based on multi-modal feature-level fusion, including:

[0031] Data acquisition module, used to synchronously obtain light scattering signals, environmental parameters and actual dust concentration;

[0032] A local feature extraction module, whose input end is connected to the light scattering signal output end of the data acquisition module, is composed of a convolutional neural network (CNN) and is used to extract local spatial features of the light scattering signal;

[0033] A global feature extraction module, whose input end is connected to the output end of the local feature extraction module, is composed of a Transformer and is used to extract global spatial features from local spatial features;

[0034] A fusion module, having a first input end connected to the output end of the global feature extraction module and a second input end connected to the environmental parameter output end of the data acquisition module, for concatenating and linearly transforming the global spatial features and the environmental parameter features to generate fusion features;

[0035] an inversion module, whose input end is connected to the output end of the fusion module, and is composed of a fully connected neural network, and is used to output a dust concentration prediction value based on the fusion features;

[0036] The model optimization module has a first input end connected to the output end of the inversion module and a second input end connected to the real end of the data acquisition module.

[0037] Furthermore, the global feature extraction module includes a multi-head self-attention unit, which is configured to calculate the correlation weight between features of different scattering angles, and the input end of the multi-head self-attention unit is connected to the output end of the CNN.

[0038] Furthermore, the fusion module includes:

[0039] The splicing unit, whose input end is connected to the output end of the global feature extraction module and the environmental parameter output end of the data acquisition module, is used to convert the light scattering global feature vector F scatt and the environmental feature vector F env Spliced ​​into [F scatt ; F env ];

[0040] The linear transformation unit, whose input end is connected to the output end of the splicing unit, is used to perform W f [F scatt ; F env ]+b f Linear transformation of

[0041] The activation unit has an input end connected to the output end of the linear transformation unit and is used to generate fusion features through the activation function σ(·).

[0042] Furthermore, the loss function calculation unit of the model optimization module satisfies:

[0043]

[0044] Where N is the number of samples, is the predicted concentration value output by the inversion module, C (i) is the actual concentration value obtained by the data acquisition module, α is the regularization coefficient, ‖W f ‖2 is the weight square norm of the linear transformation unit in the fusion module.

[0045] The beneficial effects of the present invention are:

[0046] (1) By integrating the multimodal characteristics of optical scattering signals and environmental parameters, the model autonomously establishes a correlation mapping between environmental interference and dust scattering characteristics, overcoming the systematic measurement deviation caused by factors such as humidity and temperature in traditional light scattering methods, and significantly enhancing the anti-interference ability under complex working conditions.

[0047] (2) The convolutional neural network captures the local spatial pattern of the scattered signal, and the Transformer self-attention mechanism analyzes the global angular distribution characteristics. The two work together to form a multi-scale feature extraction architecture to achieve in-depth characterization of the dust light field characteristics.

[0048] (3) The feature-level fusion mechanism embeds the environmental state information into the light field feature space, enabling the inversion model to dynamically compensate for the coupling effect of environmental interference on light scattering, significantly reducing the concentration prediction error.

[0049] (4) The data-driven training method supervises the model optimization process through standard concentration labels and combines regularization constraints to ensure the generalization stability of the model in complex scenarios such as sudden changes in dust particle size distribution and extreme temperature and humidity.

[0050] (5) While maintaining the real-time advantage of the light scattering method, the collaborative design of multimodal feature fusion and deep learning architecture solves the bottleneck of response hysteresis of traditional high-precision detection methods, meeting the dual needs of industrial sites for real-time and high-precision monitoring.

[0051] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0053] Figure 1 This is a schematic diagram of the principle of the present invention. DETAILED DESCRIPTION

[0054] The following describes the embodiments of the present invention by means of specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0055] Among them, the accompanying drawings are only for illustrative purposes and represent only schematic diagrams rather than actual pictures, and should not be understood as limiting the present invention. In order to better illustrate the embodiments of the present invention, some parts of the accompanying drawings may be omitted, enlarged or reduced, and do not represent the dimensions of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions may be omitted in the accompanying drawings.

[0056] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "back", etc. indicating directions or positional relationships, they are based on the directions or positional relationships shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operate in a specific direction. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0057] See also Figure 1 , the present invention comprises the following steps:

[0058] (1) Data collection and preprocessing

[0059] Optical sensors are used to obtain multi-angle scattered light signals from dust:

[0060] I( i ) The dust concentration is the scattered light signal data matrix obtained by the i-th measurement:

[0061] I (i) =[I(λ,θ1),I(λ,θ2),...,I(λ,θ M )]

[0062] Where λ represents the wavelength of the light source, θ j (j=1, 2, ..., M) represents different scattering angles, and M is the number of sampled light scattering angles.

[0063] Synchronously, the environmental parameter sensors are used to obtain the environmental status at the time of sampling, including temperature T(i), humidity H(i), wind speed V(i), wind direction D(i), etc., to construct the environmental feature vector

[0064] E (i) =[T (i) ,H (i) ,V (i) ,D (i) ]

[0065] In addition, a standard dust concentration meter (such as filter membrane weighing method or other calibrated instruments) must be used to measure the actual dust concentration C at the sampling time. (i) , as the label value of the sample, the unit is mg / m 3 .

[0066] Through the above data collection, a light scattering signal I (i) , environment vector E (i) The corresponding true concentration C (i) In the preprocessing stage, necessary normalization or standardization can be performed on features of different dimensions, such as taking the logarithm of light intensity signals to amplify weak signals and subtracting the mean of environmental parameters, to improve the fusion and numerical stability of different modal data.

[0067] (2) Local spatial feature extraction of light scattering signals

[0068] First, the local features of the scattered light signal are extracted through a one-dimensional convolutional neural network (CNN). The output feature map of the l-th convolution layer is denoted as H (l) , which is calculated as:

[0069] H (l) =ReLU(W (l) *H (l-1) +b (l) )

[0070] Where W (l) represents the convolution kernel weight matrix of the lth layer, b (l) is the bias term of this layer, the symbol * represents the convolution operation, and ReLU(×) is the activation function (rectified linear unit). (0) is the initial input scattered light intensity data, H (0) =X (i) After multiple layers of convolution stacking and nonlinear transformation, a deep local feature representation of the multi-angle scattering signal can be obtained. The convolution feature can capture the changing trend of the scattering angle signal in the local neighborhood, such as the pattern of sharp increase or decrease of light intensity at certain adjacent angles, reflecting the local spatial information of the particle scattering intensity. The feature map output by the last layer is

[0071] (3) Global spatial feature self-attention extraction of light scattering signals

[0072] In order to further obtain the global distribution characteristics of the scattered light signal in the entire angle range, the present invention introduces the Transformer self-attention mechanism after CNN. Specifically, the feature map H output by the last layer of the CNN is converted into (LC) Flattened into a series of feature vector sequences and fed into the Transformer module. The Transformer first obtains the query Q, key K, and value V matrices through linear transformation, and then calculates multi-head self-attention to capture the correlation of features between angles. The calculation of multi-head self-attention can be expressed as:

[0073]

[0074] Z att =[head1,head2,...,head H ]W O ,

[0075] in is the query, key, and value mapping matrix of the h-th attention head, W O is the multi-head attention output mapping matrix, and H is the number of attention heads. Through the above self-attention mechanism, the model can calculate the weighted representation of each position based on the correlation between the features of different positions (different scattering angles) within the scattering signal sequence, and realize the extraction of global features of the light field. The Transformer module finally outputs a global feature representation F that integrates the correlation between multiple angles. scatt , which can be viewed as a high-dimensional eigenvector of the optical scattering mode.

[0076] (4) Multimodal feature-level fusion

[0077] The light scattering global eigenvector F obtained above is scatt The corresponding environmental parameter vector E (i) In this invention, the feature level connection plus linear transformation is used to fuse different modal information. env is the characteristic representation of the environmental parameter vector (in simple cases, the original environmental vector E can be directly used (i) As feature representation, or if necessary, first pass through a fully connected network to extract environmental features), then the fusion feature representation F fusion It can be expressed as:

[0078] F fusion =σ(W f [F scatt ; Fenv ]+b f )

[0079] The symbol [;] represents the vector connection operation, W f is the weight matrix of the fusion layer, b f is the bias vector, and σ(×) represents the activation function (such as ReLU or LeakyReLU). After fusion, the model can simultaneously consider the impact of optical scattering patterns and environmental conditions on dust concentration in a unified feature space, thereby improving the model's environmental adaptability and prediction accuracy.

[0080] (5) Establishment of concentration inversion model

[0081] After completing the multimodal feature fusion, the fused feature vector F fusion Enter the fully connected neural network (multi-layer perceptron, MLP) for the final concentration regression mapping. The fully connected network consists of several hidden layers and an output layer, mapping the high-dimensional fusion features into scalar dust concentration values. Assuming that the fully connected network contains L layers (excluding the input feature layer), the output of the lth layer can be expressed as:

[0082] h (l) =φ(W (l) h (l-1) +b (l) ),l=1,2,...,L

[0083]

[0084] Among them, h (0) =F fusion is the input of the fusion feature, W (l) 、b (l) are the weight matrix and bias vector of the fully connected layer l, φ(×) is the hidden layer activation function (such as ReLU or Sigmoid), Indicates the dust concentration value predicted by the output layer (unit: mg / m 3 ).

[0085] In the model training stage, the optimal weight W of the concentration inversion model is determined by a supervised learning method. f and bias parameter b f Construct a loss function with dust concentration prediction error as the target, such as mean square error (MSE) loss:

[0086]

[0087] Where N is the total number of training samples, C (i) is the actual dust concentration of the i-th sample (reference value measured by a standard instrument), is the concentration value predicted by the model, α is the regularization coefficient, |W f | 2 Represents the squared norm of the fusion layer weights. By minimizing the aforementioned loss function, the model parameters are iteratively updated using an error backpropagation algorithm (such as stochastic gradient descent or the Adam optimizer). All model parameters (including the CNN convolution kernels, Transformer weights, and the weights of the fusion and subsequent fully connected layers) are updated with gradients to continuously reduce the error between the predicted and true concentrations. Training continues until the loss function converges or a preset stopping condition is reached. During this process, the model gradually learns the relationship between each modal feature and dust concentration.

[0088] (6) Concentration inversion

[0089] After the model training is completed, the new light scattering multi-angle signal and the corresponding environmental parameters are input into the model, and the output layer can be obtained through forward propagation. As a real-time estimate of the dust concentration in the current environment. Since the model parameters have been trained through a large number of samples with calibration data, It has good adaptability and accuracy to various environmental conditions and changes in dust characteristics.

[0090] Example 1: Real-time monitoring of dust concentration in coal mines

[0091] 1. Data Collection

[0092] Multi-angle optical sensors are deployed in coal mine tunnels, and the light source wavelength uses 660-nanometer infrared laser.

[0093] Synchronously collect signals at five scattering angles: 30°, 60°, 90°, 120°, and 150°.

[0094] Environmental sensors obtain temperature, humidity, and wind speed data in real time.

[0095] The true value of dust concentration is obtained every 10 minutes by filter membrane weighing method for calibration.

[0096] 2. Feature Extraction

[0097] Local feature extraction: The light scattering signal is input into a 3-layer CNN network with a convolution kernel size of 3, a stride of 1, and the number of output channels of each layer is 16, 32, and 64, respectively.

[0098] Global feature extraction: The CNN output features are fed into the Transformer module, which uses four attention heads to calculate the correlation weights of features at different scattering angles.

[0099] 3. Feature Fusion

[0100] The 128-dimensional light scattering global feature output by Transformer is concatenated with the 4-dimensional environment feature vector.

[0101] Perform linear transformation through the fully connected layer: the weight matrix dimension is 132×64, and the activation function uses LeakyReLU.

[0102] 4. Concentration inversion

[0103] The fused features are input into a 2-layer fully connected network: the first layer has 64 neurons and the second layer has 1 neuron that outputs the concentration value.

[0104] The Huber loss function is used during model training, and the Adam optimizer is used to update the parameters.

[0105] 5. Real-time monitoring

[0106] The system outputs the concentration prediction value every 2 seconds.

[0107] When the humidity exceeds 85%, the environmental feature weight coefficient is automatically enhanced to dynamically compensate for humidity interference.

[0108] Example 2: Dust explosion risk warning in industrial workshops

[0109] 1. Multimodal Data Collaboration

[0110] The optical sensor collects light intensity signals at a 90° scattering angle and a wavelength of 532 nanometers.

[0111] Add dust particle size distribution sensor as a supplementary modality.

[0112] Environmental monitoring includes temperature, humidity, and ventilation intensity.

[0113] 2. Adaptive feature extraction

[0114] Local feature optimization: The CNN convolution kernel dynamically adjusts its size according to the particle size distribution, and a 5×5 kernel is used to capture wide-band scattering for large-size particles.

[0115] Global feature enhancement: Transformer assigns low attention weight to signal fluctuations under high ventilation intensity conditions.

[0116] 3. Hierarchical integration mechanism

[0117] First-level fusion: light scattering features and particle size features are spliced ​​together, and dimensionality is reduced through autoencoders.

[0118] Second level fusion: The reduced dimension features are fused with the environment vector for the second time to output 32-dimensional features.

[0119] 4. Risk warning linkage

[0120] While the inversion network outputs the concentration value, a risk level classification layer is added:

[0121] Safety level: concentration <10mg / m 3

[0122] Warning level: 10-20 mg / m 3

[0123] Danger level: >20mg / m 3

[0124] Dangerous levels automatically trigger ventilation system pressurization and sound and light alarms.

[0125] 5. Model Self-Update

[0126] Fine-tune the fully connected layer parameters with new calibration data every week to maintain environmental adaptability.

[0127] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A dust concentration inversion method based on multimodal feature-level fusion, characterized by: The following steps are involved: S1: Data acquisition and preprocessing: synchronously collect dust light scattering signal data, environmental parameter data and corresponding real dust concentration data; S2: Extraction of local spatial features of light scattering signals: The light scattering signals from S1 are input into the convolutional neural network (CNN), and the local spatial features of the scattered light signals are extracted through multi-layer convolution operations. S3: Global spatial feature extraction of light scattering signals: The local spatial features output by S2 are input into the Transformer module, and the global spatial features of the scattered light signal are extracted through the multi-head self-attention mechanism; S4: Multimodal feature-level fusion: The light scattering global spatial features output by S3 are concatenated with the environmental parameter features, and the concatenated feature vectors are subjected to linear transformation and activation function processing to obtain fused features. S5: Establishment of concentration inversion model: The fusion features of S4 are input into the fully connected neural network to output the dust concentration prediction value. The model parameters are optimized through back propagation based on the error between the prediction value and the actual dust concentration. S6: Concentration inversion: Input the real-time collected light scattering signal and environmental parameters into the trained model and output the current dust concentration value.

2. The dust concentration inversion method based on multimodal feature-level fusion according to claim 1 is characterized in that: In S1: The light scattering signal data is the multi-angle scattered light intensity data matrix I (i) =[I(λ,θ1),I(λ,θ2),...,I(λ,θ M )], where λ represents the wavelength of the light source, θ j represents the scattering angle, j = 1, 2, ..., M, where M is the number of scattering angles; The environmental parameter data includes at least one of temperature T(i), humidity H(i), wind speed V(i), and wind direction D(i).

3. The dust concentration inversion method based on multimodal feature-level fusion according to claim 2 is characterized in that: In S2, the calculation process of CNN satisfies: A (l) =ReLU(W (l) *H (l-1) +b (l) ) Among them H (l) is the output feature map of the lth layer, W (l) is the convolution kernel weight, b (l) is the bias term, representing the convolution operation, and ReLU(·) is the activation function.

4. The dust concentration inversion method based on multimodal feature-level fusion according to claim 1 is characterized in that: In S3, the calculation of the multi-head self-attention mechanism satisfies: Z att =[head1,head2,...,head H ]W O , in is the output of the last layer of CNN, is the mapping matrix of the h-th attention head, and H is the number of attention heads.

5. The dust concentration inversion method based on multimodal feature-level fusion according to claim 1 is characterized in that: In S4, the calculation of feature fusion satisfies: F fusion =σ(W f [F scatt ;F env ]+b f ) Among them F scatt is the global spatial characteristic of light scattering, F env is the environmental parameter characteristic, [; ] represents vector concatenation, W f is the weight matrix, b f is the bias term, and σ(·) is the activation function.

6. The dust concentration inversion method based on multimodal feature-level fusion according to claim 1 is characterized in that: In S5, the output of the fully connected neural network satisfies: where h (L-1) is the output of the last hidden layer, W (L) and b (L) are the output layer parameters.

7. A dust concentration inversion system based on multimodal feature-level fusion, characterized by: include: Data acquisition module, used to synchronously obtain light scattering signals, environmental parameters and actual dust concentration; A local feature extraction module, whose input end is connected to the light scattering signal output end of the data acquisition module, is composed of a convolutional neural network (CNN) and is used to extract local spatial features of the light scattering signal; A global feature extraction module, whose input end is connected to the output end of the local feature extraction module, is composed of a Transformer and is used to extract global spatial features from local spatial features; A fusion module, having a first input end connected to the output end of the global feature extraction module and a second input end connected to the environmental parameter output end of the data acquisition module, for concatenating and linearly transforming the global spatial features and the environmental parameter features to generate fusion features; an inversion module, whose input end is connected to the output end of the fusion module, and is composed of a fully connected neural network, and is used to output a dust concentration prediction value based on the fusion features; The model optimization module has a first input end connected to the output end of the inversion module and a second input end connected to the real end of the data acquisition module.

8. The dust concentration inversion system based on multi-modal feature-level fusion according to claim 7 is characterized in that: The global feature extraction module includes a multi-head self-attention unit, which is configured to calculate the correlation weight between different scattering angle features, and the input end of the multi-head self-attention unit is connected to the output end of the CNN.

9. The dust concentration inversion system based on multi-modal feature-level fusion according to claim 7 is characterized in that: The fusion module includes: The splicing unit, whose input end is connected to the output end of the global feature extraction module and the environmental parameter output end of the data acquisition module, is used to convert the light scattering global feature vector F scatt and the environmental feature vector F env Spliced ​​into [F scatt ; F env ]; The linear transformation unit, whose input end is connected to the output end of the splicing unit, is used to perform W f [F scatt ; F env ]+b f Linear transformation of The activation unit has an input end connected to the output end of the linear transformation unit and is used to generate fusion features through the activation function σ(·).

10. The dust concentration inversion system based on multi-modal feature-level fusion according to claim 7 is characterized in that: The loss function calculation unit of the model optimization module satisfies: Where N is the number of samples, is the predicted concentration value output by the inversion module, C (i) is the actual concentration value obtained by the data acquisition module, α is the regularization coefficient, ‖W f ‖ 2 is the weight square norm of the linear transformation unit in the fusion module.

Images