Depth interpretable industrial sensor network detection method and system based on statistical constraint and variance adjustment

Abstract

The invention relates to the field of industrial intelligent detection, and particularly provides a depth interpretable industrial sensor network detection method and system based on statistical constraint and variance adjustment. The method aims to solve the contradiction between insufficient statistical model expression ability and unstable deep network black box characteristics and training. Comprising the following steps: acquiring industrial sensing data and key index labels, and constructing and dividing a data set; a multi-layer deep network is constructed, partial least squares (PLS) is utilized to provide a weight initialization direction for an intermediate layer, and feature and output association is ensured; common least square (OLS) is adopted to provide initialization for a linear output layer, and interpretability is enhanced; and furthermore, a Xavier method is combined to carry out variance scaling adjustment on the statistical weight, so that gradient disappearance and activation function saturation are avoided while statistical priori is reserved, and the training stability and generalization performance are improved. The model obtained by training can be called Deep-PLSaoNET. According to the invention, a deep learning solution with high precision, high reliability and interpretability is provided for an industrial small sample scene.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent inspection, specifically a method and system for deep interpretable industrial sensor network inspection based on statistical constraints and variance adjustment. Background Technology

[0002] In process industries (such as mineral processing, metallurgy, and chemical engineering), real-time, accurate, and stable sensing of key parameters during production using online sensing devices is the core foundation for achieving process optimization and intelligent control. However, process industry production processes are typically accompanied by a series of continuous and complex physicochemical changes, making it difficult to establish accurate mechanistic models between the inputs and outputs of online sensing devices. Measurement data (such as spectra, images, and vibration signals) often exhibit high-dimensional, nonlinear, and strongly coupled mapping relationships with target parameters. Traditional linear statistical methods, such as ordinary least squares (OLS) and partial least squares (PLS) regression models, have limited expressive power and cannot accurately represent such complex relationships, thus failing to meet the accuracy requirements of actual industrial production.

[0003] In recent years, with the rapid development of deep learning technology, neural networks have demonstrated powerful nonlinear fitting capabilities, and quantitative modeling methods based on neural networks have gradually been introduced into the field of industrial sensing and detection. However, deep neural networks have a large parameter scale and a complex optimization space. Their traditional training methods usually adopt a random initialization strategy, that is, randomly generating initial weights from a uniform or normal distribution, relying on a large number of samples to seek the optimal solution through multiple iterations. Industrial field data is not only high-dimensional, but also often accompanied by challenges such as measurement noise and multicollinearity. At the same time, the cost of acquiring labeled data is high, resulting in a limited sample size, making deep networks prone to overfitting and declining generalization performance, which restricts their application in real industrial scenarios.

[0004] Further analysis from the perspective of optimization mechanisms reveals that neural network training relies on the backpropagation of output errors, and its loss function only directly constrains the consistency between the network output and the label, with indirect constraints on the hidden layers. In contrast, PLS considers both the expressive power of latent variables on the input and their correlation with the output, using covariance maximization as the optimization objective. Its latent variable space and weight matrix possess clear statistical significance and interpretability. Although linear models such as PLS and OLS have limited predictive accuracy, their model parameters often possess physical interpretability, meeting the reliability requirements of industrial production.

[0005] It is worth noting that the OLS and PLS models can be structurally viewed as single-layer and two-layer linear networks, respectively. Compared to completely random initialization, the initial weights generated based on the linear statistical models PLS and OLS can serve as a useful reference for neural network initialization. Providing a more stable and physically meaningful starting point for network training helps reduce optimization difficulty, decrease dependence on large-scale samples, thereby suppressing overfitting and improving the model's practicality in industrial small-sample environments.

[0006] However, if the weights obtained from the statistical model are directly used as the initial parameters of the network, some weight values ​​may be too large, easily causing the activation function (such as Tanh) to enter the saturation region, leading to the vanishing gradient problem. The gradient decays layer by layer during backpropagation, making it difficult to update the parameters of the previous layers and limiting the model's training effectiveness. In contrast, commonly used neural network initialization methods (such as truncated normal distribution initialization, Xavier initialization, etc.) achieve a balance between forward propagation and backward gradient flow by constraining the range of weight values, effectively promoting training stability. Summary of the Invention

[0007] This invention provides a method and system for deep, interpretable industrial sensor network detection based on statistical constraints and variance adjustment. The method assigns initially identifiable, physically interpretable directions to the neural network weights using partial least squares (PLS) and ordinary least squares (OLS) regression. Then, it incorporates the Xavier algorithm to adaptively scale and adjust the generated weight magnitudes, thereby ensuring the stability of activation values ​​and gradient backpropagation at each layer of the network while maintaining the interpretability of the statistical model. This invention effectively integrates the advantages of statistical priors and deep learning, constructing a deep learning model that balances interpretability and performance in industrial scenarios with small sample sizes and high noise levels. It provides a new approach for intelligent detection in process industries and has significant application value in promoting the development of intelligent industries.

[0008] To achieve the above objectives, the present invention proposes the following technical solution:

[0009] A deep interpretable industrial sensor network detection method based on statistical constraints and variance adjustment is used for online detection of key indicators in industrial production processes, including the following steps:

[0010] S1: Acquire measurement data collected by industrial sensing devices and simultaneously acquire key indicator labels to generate a paired dataset {measurement data, labels};

[0011] S2: Divide the paired dataset into a training set, a validation set, and a test set;

[0012] S3: Construct a deep neural network consisting of an input layer, K non-linear hidden layers (K≥2), and a linear output layer;

[0013] S4: Using the training set data, based on the Partial Least Squares (PLS) model and the Ordinary Least Squares (OLS) model, statistical constraints are applied to the initialization process of the weights and biases of each layer of the deep neural network, specifically including:

[0014] (S4a) For the k-th hidden layer (k=1, 2, ..., K), establish a PLS model between its input and the prediction target. The number of PLS ​​latent variables is set to the number of neurons in the hidden layer. The weight matrix of the input to the latent variables calculated by the PLS model is used as the initialization direction of the weight matrix of the hidden layer. The bias is initialized to the negative mean of the linear output of the layer.

[0015] (S4b) For the output layer, establish an OLS model between the output of the Kth hidden layer (the input of this layer) and the prediction target, and use the obtained regression weights and biases as the initial values ​​of the weight initialization direction and bias of the output layer.

[0016] S5: Use the Xavier initialization method to obtain the theoretical variance of the weight initialization of each layer, and generate a scaling factor accordingly to adjust the distribution amplitude of the baseline value of the statistical constraint weight matrix obtained in step S4.

[0017] S6: Using the network initialized with weights and biases in steps S4 and S5, iteratively train it based on the backpropagation algorithm using the training set data, and finely adjust the network hyperparameters according to the performance of the validation set to obtain the optimal statistically constrained deep interpretable network model Deep-PLSaoNET.

[0018] S7: In actual industrial sensing scenarios, the measurement data or test set data collected by the sensors are input into the trained Deep-PLSaoNET model, which outputs the predicted values ​​of key indicators in real time.

[0019] When the measurement data collected by the industrial sensing device is high-dimensional spectrum, the key indicators in the industrial production process are the key components of materials throughout the entire production process, including but not limited to raw materials, intermediate products and final products.

[0020] The steps based on statistical constraints include:

[0021] (1) Define the nonlinear hidden layer L k Input-output pairs {X} (k=1, 2, ..., K, K≥2) (k) Y (k)}:X (1) For measurement data X (spec) X (k) (k≥2) represents network L k-1 Layer output O (k-1) Y (k)Both are prediction targets y; define the input-output pair {X} of the output layer. (k) Y (k)} for network L K Layer output O (K) With the predicted target y;

[0022] (2) For the nonlinear hidden layer L k The number of neurons h in the k-th layer of the network k To establish the optimal number of latent variables, a PLS model is built between their inputs and outputs, and the transformation weight matrix W from the corresponding inputs to the latent variables is obtained. PLS (k) , take it as L k The layer's weight matrix initializes the direction information, and the bias b is set. PLS (k) Initialize to N is the number of modeling samples; this ensures that the stepwise feature extraction process is correlated with the target grade and compensates for the systematic bias of the neuron input, making its mean close to 0.

[0023] (3) For the output layer, establish its input-output OLS model to obtain the weights W. OLS and bias b OLS Determine the linear interpretable initialization baseline for the output layer.

[0024] The variance-based adjustment steps include:

[0025] (1) Calculate the theoretical variance of the weight distribution of each layer of the network based on the Xavier initialization method: ,in , These represent the input and output dimensions of the network layer, respectively.

[0026] (2) Calculate the scaling factor:

[0027] (3) Nonlinear hidden layer L k The weight matrix is ​​initialized as follows: ;

[0028] (4) The weight matrix of the output layer is initialized as follows: .

[0029] The activation function used in the nonlinear hidden layer is the hyperbolic tangent Tanh function; during the network training process, the Adam optimizer is selected as the parameter optimization method, the root mean square error (RMSE) is selected as the loss function, and a maximum number of iterations and an early termination strategy are set to prevent overfitting.

[0030] The deep neural network employs, but is not limited to, fully connected networks, convolutional neural networks, LSTM networks, and Transformer networks. For different network topologies, the corresponding strategy only needs to perform statistical constraint initialization and variance readjustment on the weights and biases of each layer in the network based on the training set data, combining the partial least squares (PLS) model and the ordinary least squares (OLS) model and the Xavier initialization method.

[0031] A deep interpretable industrial sensor network detection system based on statistical constraints and variance adjustment is proposed. The system includes: industrial sensing devices, a processor, and a memory. The sensing devices collect measurement data in the industrial field for monitoring key indicators during industrial production. The memory stores the following program modules: the processor reads and executes the program to implement the method steps described above, thereby achieving prediction of key indicators of interest in the current industrial production process. The system's program modules include:

[0032] The offline modeling module is used to establish a paired dataset {measurement data acquired by sensors, corresponding key indicator labels}, and divide it into training, validation, and test sets; based on the training set data, it establishes a PLS model between the input and output of each layer of the network's nonlinear hidden layer to determine the initialization direction of its weights; based on the training set data, it establishes an OLS model between the input and output of the network's output layer to determine the initialization direction and bias value of its weights; based on Xavier initialization, it calculates the theoretical variance of the weight distribution of each layer of the network to determine the initial magnitude of the weights, and adjusts the distribution magnitude of the PLS and OLS statistical constraint weight initialization baseline values ​​accordingly; it retrains the initialized network using the training set data; and based on the performance indicators of the validation set, it determines the number of parameters of the Deep-PLSaoNET model to obtain the optimal Deep-PLSaoNET model.

[0033] Online monitoring module: This module is used to input measurement data or test set data collected by sensors in actual industrial sites into the trained Deep-PLSaoNET model to obtain predicted values ​​of key indicators, thereby realizing real-time monitoring and prediction of industrial production processes.

[0034] The industrial sensing equipment includes one or more of the following: parameter detection sensors, material control valves, and production equipment controllers.

[0035] The deep interpretable industrial sensor network detection system based on statistical constraints and variance adjustment also includes a production control module, which is used to adjust the production control quantities in the industrial field in reverse according to the predicted values ​​of key indicators output by the Deep-PLSaoNET model, thereby forming a closed-loop production control.

[0036] The production control quantity includes one or more of the following: the opening degree of the material control valve of the production equipment and the position feed amount of the motion actuator.

[0037] The beneficial effects and advantages of this invention are as follows:

[0038] 1. This invention proposes a deep, interpretable industrial sensor network detection method based on statistical constraints and variance adjustment. It innovatively combines the physically meaningful weight directions extracted by partial least squares (PLS) and ordinary least squares (OLS) regression with the Xavier algorithm for variance adjustment, achieving synergistic optimization of the initialization direction and magnitude of deep neural network weights. This method significantly enhances the model's stability, interpretability, and convergence efficiency while preserving the powerful nonlinear expressive capabilities of deep networks, effectively solving the problems of training difficulties and overfitting in industrial scenarios with small sample sizes.

[0039] 2. The Deep-PLSaoNET model constructed in this invention can more accurately and reliably detect key parameters such as material composition, equipment status and product quality in real time in complex process industrial sensing environments (such as mineral processing, metallurgy, chemical industry, etc.), providing stable and reliable data support for industrial process optimization and intelligent control.

[0040] 3. Starting from the network initialization mechanism, this invention provides a complete initialization framework for deep neural networks that integrates statistical priors and weight distribution theory. It has both the physical interpretability of traditional linear models and the high accuracy and strong fitting ability of deep networks, which significantly improves the overall performance of the model in terms of accuracy, robustness and engineering applicability. It is of great significance for promoting the development of intelligent detection and closed-loop control in industry. Attached Figure Description

[0041] Figure 1 This is a structural diagram of the method of the present invention.

[0042] Figure 2 This is a flowchart of the method of the present invention.

[0043] Figure 3 This is a schematic diagram illustrating the online detection of iron grade in iron concentrate slurry during the LIBS-based flotation process.

[0044] Figure 4 The original laser-induced breakdown spectrum of iron concentrate slurry.

[0045] Figure 5 This is a schematic diagram of the network structure used in an embodiment of the present invention.

[0046] Figure 6 The input distribution of neurons in the hidden layer of the Deep-PLSaoNET model is obtained by using PLS to obtain the weight matrix and then adjusting the variance using the Xavier algorithm.

[0047] Figure 7 The learning rate optimization process for the Deep-PLSaoNET model.

[0048] Figure 8 The modeling and prediction results are based on random initialization for deep networks with the same structure as those in this invention.

[0049] Figure 9 The modeling and prediction results of the Deep-PLSaoNET model proposed in this invention are shown. Detailed Implementation

[0050] To make the above-mentioned objectives, features and advantages of the present invention more apparent and understandable, the technical solution of the present invention will be further explained below with reference to an example of LIBS grade analysis of iron concentrate slurry.

[0051] Numerous specific details are set forth in the following description to provide a full understanding of the invention. However, the invention can be practiced in many other ways than those described herein, and similar modifications can be made by those skilled in the art without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0052] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0053] In mineral flotation, real-time detection of key components throughout the entire process (including raw materials, intermediate products, and final products) is a crucial foundation for process optimization and intelligent control. Laser-induced breakdown spectroscopy (LIBS), as an online sensing method based on high-dimensional spectral analysis, can establish a quantitative relationship model between spectra and components, enabling rapid, in-situ detection of key components. This provides critical data support for real-time monitoring and optimized control of the flotation process.

[0054] like Figure 1As shown, this invention first establishes a PLS model between the input and output of the hidden layer to obtain the transformation weight matrix from the corresponding input to the latent variable, and an OLS model between the input and output of the output layer to obtain the weight matrix and bias. Then, based on the Xavier initialization method, the theoretical variance of the weight distribution of each layer of the network is calculated, and a scaling factor is calculated accordingly to adjust the variance of the weight matrices obtained from the linear statistical model for the nonlinear hidden layer and the linear output layer, achieving coordinated optimization of the initialization direction and magnitude of the deep neural network weights. To further improve model performance, this invention further trains the network (Deep-PLSaoNET) after linear statistical model-assisted initialization based on the training set data. During training, this invention further optimizes the hyperparameters of the Deep-PLSaoNET model during training, ensuring the training effect of the model.

[0055] This embodiment provides a process for establishing an online monitoring model for iron concentrate slurry grade based on laser-induced breakdown spectroscopy (LIBS) technology. This process combines the method for constructing a deeply interpretable industrial sensor network model based on statistical constraints and variance adjustment proposed in this invention. The overall process is as follows: Figure 2 As shown, the specific steps include:

[0056] (1) Acquire spectral data:

[0057] Using LIBS technology, spectral data of slurry samples with iron grade labels were collected at the flotation site. Figure 3 The diagram shows a schematic of the iron ore flotation process, the LIBSlurry analyzer used in this invention, and the optical measurement devices inside the analyzer. The LIBSlurry analyzer, installed on the concentrate pipeline, collected 300 sets of iron concentrate pulp spectra during the iron ore beneficiation plant flotation process and simultaneously analyzed the corresponding Fe grade data.

[0058] The original laser-induced breakdown spectrum of the obtained iron concentrate slurry is as follows: Figure 4 As shown, the spectral dimension is 6116 and the wavelength range is 227-784 nm.

[0059] (2) Split the dataset:

[0060] Three hundred iron concentrate slurry samples were randomly divided into training, validation, and test sets in a 3:1:1 ratio. 180 samples were used as the training set to establish a spectral intensity-iron grade regression model; 60 samples were used as the validation set to optimize the model's hyperparameters; and the remaining 60 samples were used as the test set to compare and evaluate the model's generalization ability.

[0061] (3) Construct a deep fully connected neural network:

[0062] This invention uses a three-layer fully connected architecture as a paradigm to systematically explain the collaborative initialization mechanism of Deep-PLSaoNET, which integrates statistical constraints and variance adjustment. The network consists of one input layer, three nonlinear hidden layers, and one linear output layer, with the number of neurons in the hidden layers being 15, 10, and 5, respectively. The overall network architecture is as follows: Figure 5 As shown. Nonlinear hidden layer L 1 , L 2 , L 3 The weight matrices are respectively The bias vectors are respectively ,in For input dimensions, They are respectively L 1 , L 2 , L 3 Layer output dimension. The weight matrix of a linear output layer is... bias The mathematical expression for the current network fitting function is:

[0063] (1)

[0064] In quantitative regression tasks, the Tanh activation function is typically used to introduce a nonlinear transformation, i.e. The mathematical form is Specifically, L 1 -L 3 The progressive feature extraction process of each layer is as follows:

[0065] (2)

[0066] in, For network input.

[0067] For the output layer: .

[0068] in, For the prediction target.

[0069] (4) Solve for the weights of the PLS model and the OLS model:

[0070] (4.1) Define input-output pairs:

[0071] For nonlinear hidden layers L k Input-output pairs {X} of (k=1, 2, 3) (k) Y(k)}:X (1) For measurement data X (spec) X (k) (k=2, 3) represents network L k-1 Layer output O (k-1) Y (k) Both are prediction targets y. The input-output pairs of the output layer {X} (k) Y (k)} for network L K Layer output O (K) The target y for predicting iron grade concentration.

[0072] (4.2) For the nonlinear hidden layer L k : The number of neurons h in the k-th layer of the network k To establish the optimal number of latent variables, a PLS model is built between their inputs and outputs, and the transformation weight matrix W from the corresponding inputs to the latent variables is obtained. PLS (k) , take it as L k The layer's weight matrix is ​​initialized with orientation information, and the bias is initialized to [value]. N=180. This ensures the stepwise feature extraction process maintains its correlation with the target grade and compensates for systematic bias in the neuron inputs, making their mean close to 0.

[0073] (4.3) For the output layer: Establish its input-output OLS model and obtain the weights W. OLS and bias b OLS Determine the linear interpretable initialization baseline for the output layer.

[0074] (5) Variance adjustment using the Xavier algorithm:

[0075] (1) Calculate the theoretical variance of the weight distribution of each layer of the network based on the Xavier initialization method: ,in , These correspond to the input and output dimensions of each layer in the network, respectively. They are 2 / (285+15)=0.007, 2 / (15+10)=0.08, and 2 / (10+5)=0.13.

[0076] (2) Calculate the scaling factor for each layer of the network: ,

[0077] (3) Nonlinear hidden layer L k The weight matrix is ​​initialized as follows:

[0078] (4) The weight matrix of the output layer is initialized as follows: .

[0079] (6) Network training and parameter optimization:

[0080] In this invention, the root mean square error (RMSE) and the coefficient of determination (R²) are used. 2 ) are used as performance evaluation metrics for the model. Their definitions are shown in formulas (3) and (4):

[0081] (3)

[0082] (4)

[0083] in, It is the reference value for the nth sample. It is the predicted value of the nth sample. This is the average reference value for all samples. N represents the number of samples. In the subsequent model performance evaluation, we mainly observe the following four specific metrics: RMSEC (RMSE on the training set), RMSEV (RMSE on the validation set), RMSEP (RMSE on the test set), and R2P (R2 on the test set).

[0084] For nonlinear hidden layers containing activation functions, to ensure training stability and avoid gradient vanishing, the weight matrix obtained based on the PLS model needs to be variance-adjusted before being used as network initialization weights, as described in step (5). The standard deviation of the weights in the first layer of the Xavier algorithm network should be 0.007, the standard deviation of the weights in the second layer should be 0.08, and the standard deviation of the weights in the third layer should be 0.13. The input distribution of the neurons in the first three layers of the neural network after compression by the standard deviation of the weights is as follows: Figure 6 As shown, the neuron inputs are distributed within the unsaturated region of the Tanh function, which allows for effective learning and training.

[0085] After constraining the upper and lower bounds of the weight parameter distribution during network weight initialization, it is also necessary to optimize the learning rate parameter, which affects the training process, network convergence speed, and results. An excessively large learning rate will cause severe oscillations in the loss curve during training, making convergence difficult; while an excessively small learning rate not only slows down training but also easily leads to getting trapped in local minima. Therefore, we observe the decrease in RMSE and convergence of the training and validation sets with learning rates of 0.00001, 0.0001, and 0.001, respectively. Figure 7 Shown. According to Figure 7 As a result, the learning rate was ultimately set to 0.0001.

[0086] Furthermore, when a deep neural network structure includes not only fully connected layers but also convolutional layers (such as deep convolutional neural networks), recurrent layers (such as LSTM networks), and self-attention layers (such as Transformer networks), then the training data involved in the PLS statistical constraint initialization process needs to be adjusted as the network topology changes.

[0087] Taking the initialization of a convolutional layer as an example, when a deep neural network structure contains convolutional layers, its PLS statistical constraint initialization process needs to adjust the input data according to the characteristics of the convolution operation. The specific steps are as follows:

[0088] Assume the network training sample size is 100, the input image size is 20×20, and the convolutional layer parameters are: kernel size 5×5, output channel number 16, stride 2.

[0089] Constructing the initial training set: The initial dataset for this layer is constructed using the sliding property of the convolutional kernel on the input image. After sliding extraction of each input image, ((20 - 5) / 2) + 1 = 8, or 8 × 8 = 64 local receptive field image patches, are obtained. The image patches generated from all 100 training samples are summarized, resulting in a total of 64 × 100 = 6400 training samples.

[0090] Define the model input and output: Flatten each 5×5 local receptive field image patch to obtain a 25-dimensional input vector. The input X_conv of this convolutional layer to the dataset is 6400 25-dimensional vectors, and the output y_conv is the predicted target value y corresponding to the original 100 samples. Each y value is reused among its corresponding 64 image patches.

[0091] Establishing the PLS model and reshaping the weights: Using the number of output channels (16) of this convolutional layer as the principal components, a PLS regression model is established between X_conv and y_conv, resulting in a 25×16 weight matrix W_PLS. Each column of this matrix (a 25-dimensional weight vector) is reshaped into a 5×5 two-dimensional convolutional kernel, thereby generating initialized convolutional kernel weights for each of the 16 output channels.

[0092] Variance Adjustment: Finally, the initial weights obtained based on PLS statistical constraints are combined with the Xavier variance adjustment method to perform final scaling and shaping, so as to ensure the numerical stability of the network during training.

[0093] (7) Model Application:

[0094] Real-time acquisition of slurry spectra from industrial sites is used to input the trained Deep-PLSaoNET model for online measurement in actual industrial sensing environments, outputting predicted iron grade values ​​to further verify and optimize the model's performance.

[0095] To fully verify the effectiveness of the Deep-PLSaoNET method proposed in this invention in practical applications, we constructed a comparative model based on random initialization using the same dataset and a deep network with the same structure as in this invention. After network training and parameter optimization, its modeling and prediction results are as follows: Figure 8 As shown.

[0096] Figure 9 The results are from the Deep-PLSaoNET model proposed in this invention.

[0097] Observation and comparison Figure 8 and Figure 9 It can be seen that the Deep-PLSaoNET model provides higher accuracy and a smaller error range in predicting iron concentrate grade. This fully demonstrates the effectiveness and superiority of the Deep-PLSaoNET method proposed in this invention for iron concentrate grade prediction.

[0098] This invention also provides a deep interpretable industrial sensor network detection system based on statistical constraints and variance adjustment. The system includes: an industrial sensing device, a processor, and a memory; the sensing device collects measurement data in the industrial field for monitoring key indicators during industrial production; the memory stores the following program modules: the processor reads and executes the program to implement the method steps described above, thereby achieving the prediction of key indicators of interest in the current industrial production process. The system's program modules include:

[0099] The offline modeling module is used to establish a paired dataset {measurement data acquired by sensors, corresponding key indicator labels}, and divide it into training, validation, and test sets; based on the training set data, it establishes a PLS model between the input and output of each layer of the network's nonlinear hidden layer to determine the initialization direction of its weights; based on the training set data, it establishes an OLS model between the input and output of the network's output layer to determine the initialization direction and bias value of its weights; based on Xavier initialization, it calculates the theoretical variance of the weight distribution of each layer of the network to determine the initial magnitude of the weights, and adjusts the distribution magnitude of the PLS and OLS statistical constraint weight initialization baseline values ​​accordingly; it retrains the initialized network using the training set data; and based on the performance indicators of the validation set, it determines the number of parameters of the Deep-PLSaoNET model to obtain the optimal Deep-PLSaoNET model.

[0100] Online monitoring module: This module is used to input measurement data or test set data collected by sensors in actual industrial sites into the trained Deep-PLSaoNET model to obtain predicted values ​​of key indicators, thereby realizing real-time monitoring and prediction of industrial production processes.

[0101] The production control module is used to adjust the production control quantities in the industrial field in reverse according to the predicted values ​​of key indicators output by the Deep-PLSaoNET model, forming a closed-loop production control. The production control quantities include one or more of the following: the opening degree of the material control valve of the production equipment, and the position feed amount of the motion actuator.

[0102] Another embodiment of the present invention takes a chemical production process as an example. The key indicator labels related to industrial production are the content values ​​of specified substances in the target product. Measurement data includes reaction temperature, pressure, material flow rate, etc., which are collected in real time by sensors and transmitted to the system. Using the neural network detection method and system of the present invention, the measurement data of the target product serves as the input to the model for training and optimizing the Deep-PLSaoNET model. The key indicator labels are the output of the model, representing the predicted content values ​​of specified substances in the target product. Initialization based on statistical constraints and variance adjustment ensures the rationality of network initialization; continuous learning and adjustment ensure that the model can achieve higher prediction accuracy. The Deep-PLSaoNET model predicts the content values ​​of specified substances in the target product in real time, and production control quantities are adjusted based on the prediction results, such as adjusting the opening degree of material control valves and changing the position feed rate of motion actuators, thereby achieving precise control of the production process.

[0103] The embodiments described above will help those skilled in the art to further understand the present invention, but do not limit the present invention in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

Claims

1. A detection method for deep interpretable industrial sensor networks based on statistical constraints and variance adjustment, characterized in that, This method is used for online monitoring of key indicators in industrial production processes, and includes the following steps: S1: Acquire measurement data collected by industrial sensing devices and simultaneously acquire key indicator labels to generate a paired dataset {measurement data, labels}; S2: Divide the paired dataset into a training set, a validation set, and a test set; S3: Construct a deep neural network consisting of an input layer, K non-linear hidden layers (K≥2), and a linear output layer; S4: Using the training set data, based on the Partial Least Squares (PLS) model and the Ordinary Least Squares (OLS) model, statistical constraints are applied to the initialization process of the weights and biases of each layer of the deep neural network, specifically including: (S4a) For the k-th hidden layer (k=1, 2, ..., K), establish a PLS model between its input and the prediction target. The number of PLS ​​latent variables is set to the number of neurons in the hidden layer. The weight matrix of the input to the latent variables calculated by the PLS model is used as the initialization direction of the weight matrix of the hidden layer. The bias is initialized to the negative mean of the linear output of the layer. (S4b) For the output layer, establish an OLS model between the output of the Kth hidden layer (the input of this layer) and the prediction target, and use the obtained regression weights and biases as the initial values ​​of the weight initialization direction and bias of the output layer. S5: Use the Xavier initialization method to obtain the theoretical variance of the weight initialization of each layer, and generate a scaling factor accordingly to adjust the distribution amplitude of the baseline value of the statistical constraint weight matrix obtained in step S4. S6: Using the network initialized with weights and biases in steps S4 and S5, iteratively train it based on the backpropagation algorithm using the training set data, and finely adjust the network hyperparameters according to the performance of the validation set to obtain the optimal statistically constrained deep interpretable network model Deep-PLSaoNET. S7: In actual industrial sensing scenarios, the measurement data or test set data collected by the sensors are input into the trained Deep-PLSaoNET model, which outputs the predicted values ​​of key indicators in real time.

2. The method for constructing a deep interpretable industrial sensor network model based on statistical constraints and variance adjustment according to claim 1, characterized in that, When the measurement data collected by the industrial sensing device is high-dimensional spectrum, the key indicators in the industrial production process are the key components of materials throughout the entire production process, including but not limited to raw materials, intermediate products and final products.

3. The method for constructing a deep interpretable industrial sensor network model based on statistical constraints and variance adjustment according to claim 1, characterized in that, The steps based on statistical constraints include: (1) Define the nonlinear hidden layer L k Input-output pairs {X} (k=1, 2, ..., K, K≥2) (k) Y (k) }:X (1) For measurement data X (spec) X (k) (k≥2) represents network L k-1 Layer output O (k-1) Y (k) Both are prediction targets y; define the input-output pair {X} of the output layer. (k) Y (k) } for network L K Layer output O (K) With the predicted target y; (2) For the nonlinear hidden layer L k The number of neurons h in the k-th layer of the network k To establish the optimal number of latent variables, a PLS model is built between their inputs and outputs, and the transformation weight matrix W from the corresponding inputs to the latent variables is obtained. PLS (k) , take it as L k The layer's weight matrix initializes the direction information, and the bias b is set. PLS (k) Initialize to N is the number of modeling samples; this ensures that the stepwise feature extraction process is correlated with the target grade and compensates for the systematic bias of the neuron input, making its mean close to 0. (3) For the output layer, establish its input-output OLS model to obtain the weights W. OLS and bias b OLS Determine the linear interpretable initialization baseline for the output layer.

4. The method for constructing a deep interpretable industrial sensor network model based on statistical constraints and variance adjustment according to claim 1, characterized in that, The variance-based adjustment steps include: (1) Calculate the theoretical variance of the weight distribution of each layer of the network based on the Xavier initialization method: ,in , These represent the input and output dimensions of the network layer, respectively. (2) Calculate the scaling factor: ; (3) Nonlinear hidden layer L k The weight matrix is ​​initialized as follows: ; (4) The weight matrix of the output layer is initialized as follows: .

5. The method for constructing a deep interpretable industrial sensor network model based on statistical constraints and variance adjustment according to claim 1, characterized in that, The activation function used in the nonlinear hidden layer is the hyperbolic tangent Tanh function; during the network training process, the Adam optimizer is selected as the parameter optimization method, the root mean square error (RMSE) is selected as the loss function, and a maximum number of iterations and an early termination strategy are set to prevent overfitting.

6. The method for constructing a deep interpretable industrial sensor network model based on statistical constraints and variance adjustment according to any one of claims 1-5, characterized in that, The deep neural network employs, but is not limited to, fully connected networks, convolutional neural networks, LSTM networks, and Transformer networks. For different network topologies, the corresponding strategy only needs to perform statistical constraint initialization and variance readjustment on the weights and biases of each layer in the network based on the training set data, combining the partial least squares (PLS) model and the ordinary least squares (OLS) model and the Xavier initialization method.

7. A deep interpretable industrial sensor network detection system based on statistical constraints and variance adjustment, characterized in that, The system includes: an industrial sensing device, a processor, and a memory; the sensing device collects measurement data in the industrial field for monitoring key indicators during industrial production; the memory stores the following program modules: the processor reads and executes the program to implement the steps of the method as described in any one of claims 1 to 6, thereby achieving the prediction of key indicators of concern in the current industrial production process; the program modules of the system include: The offline modeling module is used to establish a paired dataset {measurement data acquired by sensors, corresponding key indicator labels}, and divide it into training, validation, and test sets; based on the training set data, it establishes a PLS model between the input and output of each layer of the network's nonlinear hidden layer to determine the initialization direction of its weights; based on the training set data, it establishes an OLS model between the input and output of the network's output layer to determine the initialization direction and bias value of its weights; based on Xavier initialization, it calculates the theoretical variance of the weight distribution of each layer of the network to determine the initial magnitude of the weights, and adjusts the distribution magnitude of the PLS and OLS statistical constraint weight initialization baseline values ​​accordingly; it retrains the initialized network using the training set data; and based on the performance indicators of the validation set, it determines the number of parameters of the Deep-PLSaoNET model to obtain the optimal Deep-PLSaoNET model. Online monitoring module: This module is used to input measurement data or test set data collected by sensors in actual industrial sites into the trained Deep-PLSaoNET model to obtain predicted values ​​of key indicators, thereby realizing real-time monitoring and prediction of industrial production processes.

8. The deep interpretable industrial sensor network detection system based on statistical constraints and variance adjustment according to claim 7, characterized in that, The industrial sensing equipment includes one or more of the following: parameter detection sensors, material control valves, and production equipment controllers.

9. The deep interpretable industrial sensor network detection system based on statistical constraints and variance adjustment according to claim 7 or 8, characterized in that, It also includes a production control module, which is used to adjust the production control quantities in the industrial field in reverse according to the predicted values ​​of key indicators output by the Deep-PLSaoNET model, so as to form a closed-loop production control.

10. The deep interpretable industrial sensor network detection system based on statistical constraints and variance adjustment according to claim 9, characterized in that, The production control quantity includes one or more of the following: the opening degree of the material control valve of the production equipment and the position feed amount of the motion actuator.

Images