Machine learning-based method for constructing a model for predicting harmful substance release from a coating

Abstract

The present application relates to the technical field of intelligent coating monitoring data analysis, and particularly relates to a harmful substance release prediction model construction method based on machine learning, which comprises collecting heat flux gradient data on the surface of a semiconductor power device in a constant voltage control branch, solving carrier mobility deviation through a semiconductor drift diffusion equation, and then generating a feedforward compensation gate voltage sequence to maintain a constant volatile matter sensor power supply voltage. The time sequence trajectory of voltage compensation is extracted and reconstructed with a harmful substance concentration waveform in phase space, and a high-fidelity chemical release feature vector is extracted by using a persistent homology to strip thermal fluctuation distortion characteristics. The feature vector is input into a deep neural network, an orthogonal subspace projection constraint is introduced to calculate an orthogonal penalty term and update the weight, and the network parameters are saved to establish a prediction model when the target function converges. The scheme effectively eliminates the interference of curing heat release on the sensor baseline, and improves the robustness and accuracy of the prediction model.

Description

Technical Field

[0001] This invention relates to the field of intelligent coating monitoring data analysis technology, and in particular to a method for constructing a predictive model for the release of harmful substances in coatings based on machine learning. Background Technology

[0002] In the field of intelligent coating production monitoring, dynamic sensing technology is widely used to detect the concentration of volatile harmful substances. For the sensing elements to operate normally, a high-precision regulated power supply is required to provide the driving force. The regulated power supply includes a power supply branch, within which semiconductor power devices are arranged on the circuit board and connected in series in the load circuit. The system uses a voltage adjustment module to maintain the voltage level. The underlying raw data is then converted into feature vectors by a sampling circuit, and the evolution of these feature vectors is analyzed by a machine learning model to predict the release amount of harmful substances.

[0003] The curing process of the coating releases a significant amount of heat, and the resulting environmental thermal stress is then conducted to the semiconductor regulator in the voltage stabilization system. As electronic devices, the semiconductor regulator is highly susceptible to thermal drift, causing sudden jumps in the output level of the voltage regulator circuit. This further induces a nonlinear shift in the output signal of the sensing front-end. Along with this nonlinear shift in the output signal, the sampled low-level feature data exhibits severe phase difference and amplitude distortion. Because the prediction model lacks sufficient robustness to distorted inputs, the feature analysis process reveals significant logical flaws. For example, the rapid film formation of the coating creates a high-temperature environment, causing the turn-on voltage of the transistors inside the voltage regulator circuit's package to fluctuate irregularly with changes in ambient temperature. This fluctuation in output voltage leads to pulse interference and induces breaks in the waveform signal acquired by the sensor. The algorithm logic receives these broken waveforms and determines the input data as invalid, ultimately causing the machine learning prediction model to crash. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for constructing a predictive model for the release of harmful substances in coatings based on machine learning. This invention solves the technical problem that sudden changes in sensor power supply voltage caused by the exothermic curing of coatings lead to distortion of underlying feature data and consequently, the collapse of the machine learning prediction model.

[0005] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows: The present invention provides a method for constructing a machine learning-based model for predicting the release of harmful substances in coatings, comprising: Collect heat flux gradient data transferred to the surface of semiconductor power devices in the voltage regulation control branch; Substitute the heat flux gradient data into the semiconductor drift-diffusion equation to calculate the carrier mobility deviation of the semiconductor power device. A feedforward compensated gate voltage sequence is generated based on the carrier mobility deviation. The feedforward compensation gate voltage sequence is input to the control terminal of the semiconductor power device to maintain a constant power supply voltage to the volatile sensor. Extract the timing voltage evolution trajectory corresponding to the feedforward compensated gate voltage sequence; The constant power supply voltage is used to drive the volatile substance sensor to obtain the harmful substance concentration sensing waveform output by the volatile substance sensor; A phase space reconstruction operation is performed on the time-series voltage evolution trajectory and the sensor waveform of the concentration of harmful substances to generate a phase space reconstruction matrix. The phase space reconstruction matrix is ​​continuously homologous to eliminate the distortion dimension features caused by residual thermal fluctuations and extract the chemical release feature vector. The chemical release feature vector is input into a deep neural network to perform forward propagation calculation, outputs a network mapping matrix, obtains the true label vector, calculates the prediction error between the network mapping matrix and the true label vector, and calculates the orthogonal penalty term according to the orthogonal subspace projection constraint condition. The prediction error and the orthogonal penalty term are combined to construct a global objective function. The prediction error is backpropagated along the gradient descent direction to update the network node mapping weights of the deep neural network. When the global objective function converges, the network structure parameters corresponding to the network node mapping weights are saved to establish a prediction model.

[0006] Furthermore, in the method for constructing a predictive model for the release of harmful substances in coatings based on machine learning described in this invention, the step of collecting heat flux gradient data transmitted to the surface of semiconductor power devices within the voltage regulation control branch includes: A thin-film thermocouple array located on the surface of the semiconductor power device is used to obtain multi-point temperature difference potential signals; The multi-point temperature difference potential signal is differentially amplified to output the heat flux gradient data.

[0007] Furthermore, the method for constructing a predictive model for the release of harmful substances in coatings based on machine learning, as described in this invention, includes substituting the heat flux gradient data into the semiconductor drift-diffusion equation to calculate the carrier mobility deviation of the semiconductor power device, comprising: Substitute the heat flux gradient data into the semiconductor drift-diffusion equation to calculate the change in carrier scattering cross section caused by lattice thermal vibration. The carrier mobility deviation of the semiconductor power device is calculated based on the change in the carrier scattering cross section.

[0008] Furthermore, the method for constructing a predictive model for the release of harmful substances in coatings based on machine learning, as described in this invention, includes generating a feedforward compensated gate voltage sequence based on the carrier mobility deviation, comprising: The carrier mobility deviation is subjected to signal microprocessing operations to generate reverse voltage envelope data; The reverse voltage envelope data is analyzed by digital-to-analog conversion, and the feedforward compensated gate voltage sequence is output.

[0009] Furthermore, the method for constructing a predictive model for the release of harmful substances in coatings based on machine learning, as described in this invention, includes performing a phase space reconstruction operation on the time-series voltage evolution trajectory and the sensor waveform of the harmful substance concentration to generate a phase space reconstruction matrix, comprising: The concentration sensing waveform of the harmful substance and the time-series voltage evolution trajectory are converted into a discrete time series. Calculate the mutual information parameters of the discrete time series and output the optimal time delay value; The discrete time series is processed using the spurious nearest neighbor algorithm, and the optimal embedding dimension is output. Based on the optimal time delay value and the optimal embedding dimension, the discrete time series is subjected to delay coordinate embedding operation to generate the phase space reconstruction matrix.

[0010] Furthermore, the method for constructing a predictive model for the release of harmful substances in coatings based on machine learning, as described in this invention, includes the following steps: performing continuous homology calculations on the phase space reconstruction matrix, eliminating distortion dimensional features caused by residual thermal fluctuations, and extracting chemical release feature vectors. Construct a Vittoris Lipps simplex complex on the phase space reconstruction matrix; The connected radii of the Vitoris-Lipps simplex are traversed to perform homology calculations, and a continuously homology barcode is output. Extract short-lifetime homology classes from the continuous homology barcodes whose lifetime is below a set threshold.

[0011] Furthermore, the method for constructing a predictive model for the release of harmful substances in coatings based on machine learning, as described in this invention, after extracting short-lifetime homogeneous classes with lifetimes below a set threshold from the continuous homogeneous barcode, further includes: The short-lifecycle homogeneity class is determined to be a distortion dimension feature caused by the residual thermal fluctuation and is then removed. Extract the topological feature data corresponding to long-lifetime homology classes whose lifetime exceeds the set threshold from the continuous homology barcode; The topological feature data is reorganized into the chemical release feature vector.

[0012] Furthermore, the method for constructing a prediction model for the release of harmful substances in coatings based on machine learning according to the present invention includes the following steps: inputting the chemical release feature vector into a deep neural network to perform forward propagation calculation, outputting a network mapping matrix, obtaining the true label vector, calculating the prediction error between the network mapping matrix and the true label vector, and calculating an orthogonal penalty term based on the orthogonal subspace projection constraint condition, including: The chemical release feature vectors are divided into a training sample set and a validation sample set; The training sample set is input into the deep neural network to perform the forward propagation calculation, and the network mapping matrix is ​​output. Calculate the prediction error between the network mapping matrix and the real label vector; Extract the hidden feature data output from the hidden layer of the deep neural network, and divide the hidden feature data into a chemical feature matrix and a thermal fluctuation residual matrix; Calculate the orthogonal penalty term between the chemical characteristic matrix and the thermal fluctuation residual matrix.

[0013] Furthermore, the method for constructing a prediction model for the release of harmful substances in coatings based on machine learning according to the present invention, wherein the step of summarizing the prediction error and the orthogonal penalty term to construct a global objective function, and backpropagating the prediction error along the gradient descent direction to update the network node mapping weights of the deep neural network, includes: The orthogonal penalty term is defined as an independence constraint parameter; By summing the prediction error and the independence constraint parameters, the global objective function is constructed. Calculate the loss gradient matrix of the global objective function; The prediction error and the independence constraint parameter are backpropagated according to the gradient descent direction of the loss gradient matrix, and the network node mapping weights of the deep neural network are updated.

[0014] Furthermore, the method for constructing a prediction model for the release of harmful substances in coatings based on machine learning, as described in this invention, includes the step of saving the network structure parameters corresponding to the network node mapping weights to establish the prediction model when the global objective function converges, comprising: The verification sample set is input into the deep neural network with updated weights for verification calculation. When the rate of change of the global objective function is lower than the convergence threshold within a preset iteration period, the global objective function is determined to have converged. Save the network structure parameters at the convergence threshold and output the prediction model.

[0015] Beneficial effects of this invention: This invention provides a machine learning-based method for constructing a predictive model for the release of harmful substances in coatings. This method integrates physical state feedforward control, topological coherence and coupling calculations, and network space orthogonal constraint mechanisms to form a robust closed-loop data processing layer. The underlying hardware utilizes a thin-film thermocouple array to collect heat flux gradient data transferred to the surface of the semiconductor power device within the voltage regulation control branch. The computation unit substitutes this heat flux gradient data into the semiconductor drift diffusion equation to calculate the carrier mobility deviation within the semiconductor power device's packaging structure. Based on the calculated carrier mobility deviation, the control module generates a feedforward compensation gate voltage sequence to maintain a constant supply voltage to the volatile substance sensor. This hardware level compensation logic directly overcomes the technical deficiency of conventional passive heat dissipation methods in existing technologies, which cannot eliminate thermal drift in the deep structure of semiconductor devices.

[0016] After the hardware compensation is completed, the data processing node receives the time-series voltage evolution trajectory and the sensor waveform of hazardous substance concentration, and performs phase space reconstruction calculation on the input waveform data, outputting the corresponding phase space reconstruction matrix. The processing program then performs continuous cohomology calculation on the phase space reconstruction matrix to remove the distortion dimensionality features caused by residual thermal fluctuations; after dimensionality reduction and feature extraction operations, the system finally outputs a chemical release feature vector with topological invariance.

[0017] The algorithm at the model prediction level receives chemical release feature vectors and inputs them into a deep neural network to perform forward propagation calculations. Based on orthogonal subspace projection constraints, the computation logic calculates the orthogonal penalty term between the chemical feature matrix and the thermal fluctuation residual matrix. Simultaneously, the algorithm aggregates the prediction error and the orthogonal penalty term to construct a global objective function, and then backpropagates the error along the gradient descent direction to accurately update the mapping weights of the network nodes.

[0018] The system achieves deep synergy through three layers of logic: feedforward voltage compensation at the hardware level, topology feature purification at the data analysis layer, and orthogonal feature isolation at the model prediction layer. This collaborative action first mitigates voltage disturbances caused by thermal stress at the hardware control layer, then cleans residual thermal noise from the waveform at the data analysis layer, and finally forcibly isolates underlying hardware thermal stress interference and chemical release characteristics at the algorithm mapping layer. This invention blocks the hardware state and data mapping transmission path from environmental thermal disturbances to algorithm feature mapping, constructs a machine learning model with anti-interference capabilities against physical thermal stress disturbances, and ultimately achieves accurate prediction of the dynamic patterns of harmful substance release from coatings. Attached Figure Description

[0019] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the method for constructing a predictive model for the release of harmful substances in coatings based on machine learning, as described in this invention. Detailed Implementation

[0021] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 The present invention provides a method for constructing a machine learning-based prediction model for the release of harmful substances in coatings, comprising: Step 1: Collect heat flux gradient data transferred to the surface of semiconductor power devices in the voltage regulation control branch; Step 2: Substitute the heat flux gradient data into the semiconductor drift-diffusion equation to calculate the carrier mobility deviation of the semiconductor power device; Step 3: Generate a feedforward compensated gate voltage sequence based on the carrier mobility deviation; Step 4: Input the feedforward compensation gate voltage sequence into the control terminal of the semiconductor power device to maintain a constant power supply voltage to the volatile sensor; Step 5: Extract the timing voltage evolution trajectory corresponding to the feedforward compensated gate voltage sequence; Step 6: Use the constant power supply voltage to drive the volatile substance sensor and obtain the harmful substance concentration sensing waveform output by the volatile substance sensor; Step 7: Perform phase space reconstruction operation on the time-series voltage evolution trajectory and the sensor waveform of the harmful substance concentration to generate a phase space reconstruction matrix; Step 8: Perform continuous cohomology calculation on the phase space reconstruction matrix, remove the distortion dimension features caused by residual thermal fluctuations, and extract the chemical release feature vector; Step 9: Input the chemical release feature vector into the deep neural network to perform forward propagation calculation, output the network mapping matrix, obtain the real label vector, calculate the prediction error between the network mapping matrix and the real label vector, and calculate the orthogonal penalty term according to the orthogonal subspace projection constraint condition. Step 10: Summarize the prediction error and the orthogonal penalty term to construct a global objective function, backpropagate the prediction error along the gradient descent direction to update the network node mapping weights of the deep neural network, and save the network structure parameters corresponding to the network node mapping weights when the global objective function converges to establish a prediction model.

[0023] In the process of acquiring heat flux gradient data transmitted to the surface of the semiconductor power device in the voltage regulation control branch, the underlying hardware utilizes a thin-film thermocouple array pre-mounted on the surface of the semiconductor power device package to synchronously detect changes in the surface temperature field, thereby acquiring microvolt-level multi-point thermoelectric potential signals generated at multiple spatial locations. Subsequently, a differential amplifier circuit receives the multi-point thermoelectric potential signals and performs common-mode interference suppression and signal gain amplification. The differential amplifier circuit is constructed from operational amplifiers, and the processed signals undergo high-frequency discrete sampling operations via an analog-to-digital converter, ultimately outputting discrete digital format heat flux gradient data.

[0024] When calculating the carrier mobility deviation of semiconductor power devices, the system substitutes heat flux gradient data as a temperature field variable into a preset semiconductor drift-diffusion equation to analyze the phonon collision frequency change caused by non-uniform thermal stress on the crystal structure. By analyzing the phonon collision frequency change, the computational logic calculates the change in carrier scattering cross-section caused by lattice thermal vibration. Based on the change in carrier scattering cross-section and the standard cross-sectional area at nominal ambient temperature, the computational logic performs algebraic operations to calculate the carrier mobility deviation caused by the Joule thermal accumulation effect inside the semiconductor power device structure package, thus completing the dimensionality reduction mapping of heat source state data from macroscopic heat flux to microscopic electronic characteristics.

[0025] The control loop that generates the feedforward compensated gate voltage sequence based on carrier mobility deviation relies on specific signal conditioning logic. The control module first performs digital microprocessing calculations on the carrier mobility deviation, and then combines these with amplitude inversion and phase alignment operations to generate reverse voltage envelope data with inverse cancellation characteristics. The digital-to-analog converter chip receives the reverse voltage envelope data in real time and performs digital-to-analog conversion analysis, outputting a continuous analog level signal according to the set clock frequency. This continuous analog level signal is ultimately converted into a feedforward compensated gate voltage sequence applied to the control terminal of the semiconductor power device, suppressing abrupt changes in the supply voltage at the source.

[0026] In the data fusion step, which performs phase space reconstruction operations on the time-series voltage evolution trajectory and the hazardous substance concentration sensing waveform to generate a phase space reconstruction matrix, the data processing module discretizes the hazardous substance concentration sensing waveform and the time-series voltage evolution trajectory according to a fixed sampling period, converting them into discrete time series. The data processing module uses Shannon entropy theory to calculate the mutual information parameters of the discrete time series at different delay time steps, and determines the time delay node corresponding to the first local minimum value of the mutual information parameter as the optimal time delay value. Simultaneously, the data processing module executes a false nearest neighbor algorithm to process the discrete time series, gradually increasing the phase space dimension and calculating the geometric distance abrupt change ratio of adjacent trajectory points. When the geometric distance abrupt change ratio is lower than the spatial folding tolerance limit, the data processing module outputs the optimal embedding dimension. Based on the optimal time delay value and the optimal embedding dimension, the delay coordinate embedding operation processes the discrete time series, expanding and reconstructing the one-dimensional time series into a phase space reconstruction matrix with complete dynamic characteristics.

[0027] In the topological data analysis process of continuously cohomologically calculating the phase space reconstruction matrix and extracting short-lifetime cohomology classes, the analysis module constructs a Vittoris-Lipps simplex complex with matrix elements as vertex data points within the high-dimensional feature space defined by the phase space reconstruction matrix. The analysis module progressively increases the connectivity radius parameter at multiple dimensions and traverses the connectivity states of the Vittoris-Lipps simplex complex, recording the continuous evolution range of each topological hole from its generation to closure. After cohomology algebraic mapping operations, the analysis module outputs a continuously cohomological barcode representing multidimensional topological invariants. The analysis module scans the continuously cohomological barcode and calculates the lifetime length of each cohomology class, then extracts short-lifetime cohomology classes with lifetime values ​​below a set threshold from the continuously cohomological barcode, serving as a priori reference for hardware thermal disturbance noise separation.

[0028] The feature recombination process following the extraction of short-lifecycle homology classes includes rigorous data cleansing logic. The logic identifies short-lifecycle homology classes as distorted dimensional features introduced by residual thermal fluctuations in the sensor signal and performs a removal operation on the connected components corresponding to these distorted dimensional features within the phase space topology. Subsequently, the recombination logic extracts long-lifecycle homology classes from the continuously homology barcodes whose lifecycles exceed a set threshold. By mapping these long-lifecycle homology classes back to the original multidimensional space, the recombination logic outputs topological feature data that is topologically invariant to perturbations in the underlying hardware's power supply voltage. This topological feature data is combined along the temporal and spatial dimensions to reconstruct a chemical release feature vector representing the pure kinetic state of the coating's volatilization.

[0029] The input layer of a deep neural network receives a chemical release feature vector containing both temporal and spatial features. The number of nodes in the input layer is exactly the same as the number of feature dimensions in the chemical release feature vector, and the chemical release feature vector is converted into a multidimensional tensor of the network, which is then passed to the hidden layers consisting of multiple fully connected layers.

[0030] The hidden layers employ a dual-branch network topology. The first branch of the hidden layer is dedicated to performing chemical feature mapping, while the second branch is dedicated to performing physical thermal noise feature mapping. Furthermore, each fully connected layer is followed by a modified linear unit activation function.

[0031] After network branch mapping calculations, the first network branch outputs the chemical feature matrix, and the second network branch outputs the thermal fluctuation residual matrix. The hidden layer passes the chemical feature matrix to the output layer, and uses a linear regression mapping function to convert the chemical feature matrix into a network mapping matrix for predicting the release rate of hazardous substances.

[0032] In the data processing path, the training sample set enters the input layer according to the set batch size. The input layer passes the batch feature tensor to the hidden layer, which performs forward propagation matrix multiplication operations and then outputs the chemical feature matrix and the thermal fluctuation residual matrix simultaneously.

[0033] The computation module extracts the chemical feature matrix and the thermal fluctuation residual matrix, performs Pearson correlation coefficient calculation, and outputs an orthogonal penalty term characterizing the degree of linear independence between the chemical feature matrix and the thermal fluctuation residual matrix. The output layer receives the chemical feature matrix to generate the network mapping matrix, and the computation module applies the mean squared error formula to calculate the prediction error between the network mapping matrix and the true label vector.

[0034] The optimization module summarizes the prediction error and orthogonal penalty term to construct a global objective function. Using the backpropagation chain rule, the optimization module calculates the partial derivative matrix of the global objective function with respect to the weights mapped to each network node of the deep neural network, and performs iterative update operations on the network node mapping weights in conjunction with the adaptive moment estimation optimization algorithm until the set iteration cycle is completed.

[0035] In the gradient optimization phase, which aggregates prediction errors and orthogonal penalty terms to construct the global objective function and update the network node mapping weights, the algorithm logic defines the orthogonal penalty term as an independence constraint parameter to limit feature entanglement in the latent feature space. The algorithm logic introduces a constant as a weight adjustment coefficient, proportionally aggregating prediction errors and independence constraint parameters to construct a global objective function that balances prediction accuracy and feature decoupling. By calculating the partial derivatives of the global objective function with respect to the connection weights of each layer in the deep neural network, the optimization module calculates the loss gradient matrix of the global objective function. Based on the gradient descent direction indicated by the loss gradient matrix, the optimization module uses an adaptive moment estimation optimization algorithm to perform backpropagation, transforming the prediction error and independence constraint parameters into gradient descent step sizes, and synchronously updating the network node mapping weights of the deep neural network.

[0036] The final step of determining global objective function convergence and saving network structure parameters to build a prediction model is equipped with early stopping and convergence verification logic. The verification sample set is independently input into the deep neural network after one round of weight updates for verification calculations, and the system obtains the objective function loss value corresponding to the verification sample set. The monitoring program establishes a sliding time window to monitor the relative change rate of the objective function loss value. When the rate of change of the global objective function is lower than a preset convergence threshold parameter within a continuously set preset iteration period, the global objective function is determined to have entered a minimum convergence state. At the node that triggers the convergence state, the system stops parameter updates and saves the network structure parameters below the convergence threshold to non-volatile storage, completing the algorithm model parameter solidification and outputting the prediction model.

[0037] Formula for calculating carrier mobility deviation in semiconductor drift-diffusion equation:

[0038] In the formula, Indicates the deviation in carrier mobility. Represents the fundamental charge constant. Indicates the effective mass of charge carriers. This represents the phonon collision frequency at the nominal ambient temperature. This represents the numerical value of the phonon collision frequency change. This represents the reference carrier mobility at the nominal ambient temperature.

[0039] Formula for calculating the change in carrier scattering cross section induced by lattice thermal vibration:

[0040] In the formula, This represents the change in carrier scattering cross section. This represents the standard cross-sectional area at the nominal ambient temperature. This represents the numerical value of the phonon collision frequency change. This indicates the phonon collision frequency at the nominal ambient temperature.

[0041] Shannon entropy theory mutual information parameter calculation formula:

[0042] In the formula, Represents mutual information parameters. Indicates the delay time step. This represents the data points in the discrete-time series at the initial node. This represents the data points in a discrete-time series after a time delay step. Representing data points With data points The joint probability distribution function, Representing data points The marginal probability distribution function, Representing data points The marginal probability distribution function.

[0043] Formula for calculating the geometric distance abrupt change ratio in the spurious nearest neighbor algorithm:

[0044] In the formula, Indicates the geometric distance abrupt change ratio. This represents the original phase space dimension before the optimal embedding dimension is increased. This represents the phase space dimension after increasing the optimal embedding dimension. Represents the trajectory points of a discrete-time series in the original phase space dimension. Represents adjacent trajectory points of a discrete-time series in the original phase space dimension. This represents the trajectory points of the discrete-time series after the phase space dimension is increased. This represents the adjacent trajectory points of a discrete-time series after the dimension of the phase space has increased. The operator represents the Euclidean distance norm calculation operator between points on the trajectory.

[0045] Mean square error formula:

[0046] In the formula, Indicates the prediction error. This represents the total amount of sample data in the training sample set. This represents the true label value corresponding to the sample in the true label vector. This represents the predicted mapping value corresponding to the sample in the network mapping matrix.

[0047] The formula for calculating the orthogonal penalty term of the Pearson correlation coefficient is as follows:

[0048] In the formula, Indicates the orthogonal penalty term. Represents the characteristic elements in the chemical characteristic matrix. This represents the arithmetic mean of all characteristic elements in the chemical characteristic matrix. Represents the characteristic elements in the residual matrix of thermal fluctuations. This represents the arithmetic mean of all characteristic elements in the residual matrix of thermal fluctuations.

[0049] Formula for calculating the global objective function:

[0050] In the formula, Represents the global objective function. Indicates the prediction error. This represents the weighting adjustment coefficient. Indicates the independence constraint parameter. This represents the absolute value of the independence constraint parameter.

[0051] In an embodiment of this invention, in the application scenario of dynamic sensing and monitoring of volatile organic compounds in indoor building decoration environment, the coating undergoes a cross-linking and curing reaction on the wall surface, thereby releasing a large amount of reaction heat into the monitoring environment. In order to collect the heat flux gradient data transferred to the surface of the semiconductor power device in the voltage regulation control branch, the underlying hardware utilizes a thin-film thermocouple array pre-mounted on the surface of the semiconductor power device package to simultaneously detect the temperature field distribution changes on the surface of the semiconductor power device.

[0052] Through the above detection operation, the detection node acquires multi-point temperature difference potential signals, including the physical coordinates of the central heating area and the edge heat dissipation area. Subsequently, the differential amplifier circuit receives the multi-point temperature difference potential signals and performs common-mode interference suppression and signal gain amplification processing; after the analog-to-digital conversion module performs high-frequency discrete sampling operation at 1000 Hz, it finally outputs discrete digital format heat flux gradient data.

[0053] The calculation of carrier mobility deviation in semiconductor power devices depends on the semiconductor drift-diffusion equation. The computational logic substitutes heat flux gradient data as a temperature field variable into the semiconductor drift-diffusion equation to analyze the change in phonon collision frequency generated by the silicon lattice structure under non-uniform thermal stress, and then calculates the change in carrier scattering cross section caused by lattice thermal vibration.

[0054] Based on the change in carrier scattering cross-section and the standard cross-sectional area at nominal ambient temperature, algebraic operations are performed. The computational logic further calculates the carrier mobility deviation within the semiconductor power device packaging material due to the Joule heat accumulation effect. The calculated carrier mobility deviation is specifically manifested as a decrease in electron mobility, thus completing the dimensionality reduction mapping of heat source state data from macroscopic heat flux to microscopic electronic characteristics.

[0055] The process of generating the feedforward compensated gate voltage sequence is based on discrete signal modulation operations. The digital microprocessor module performs discrete signal conversion on the carrier mobility deviation, and combines amplitude inversion and phase alignment processing to generate reverse voltage envelope data with inverse cancellation characteristics.

[0056] The digital-to-analog converter circuit receives the inverted voltage envelope data in real time, performs digital-to-analog conversion analysis, and outputs a continuous analog level signal according to the set clock frequency. This ultimately generates a feedforward compensation gate voltage sequence applied to the control terminal of the semiconductor power device. This feedforward compensation gate voltage sequence directly drives a change in the depletion layer barrier width inside the semiconductor junction of the semiconductor power device, completing the underlying hardware offset compensation before a sudden change in the power supply level occurs. This cuts off the thermal disturbance conduction path from environmental thermal stress to sensor baseline drift.

[0057] The data fusion step of the phase space reconstruction matrix relies on high-dimensional mapping operations. The data processing module synchronously converts the hazardous substance concentration sensing waveform representing the concentration fluctuation per million percent, as well as the time-series voltage evolution trajectory recording the hardware's thermal disturbance resistance actions, into a discrete time series with a time resolution of 5 milliseconds.

[0058] Using Shannon entropy theory, the data processing module calculates the mutual information parameters of the discrete-time series at different delay time steps, and outputs the time delay node corresponding to the first local minimum of the mutual information parameter as the optimal time delay value. Simultaneously, the data processing module executes the spurious nearest neighbor algorithm to process the discrete-time series. By gradually increasing the phase space dimension and calculating the geometric distance abrupt change ratio between adjacent trajectory points, it determines the spatial folding state and outputs the optimal embedding dimension.

[0059] In the computational process of the false nearest neighbor algorithm, the data processing module pre-sets a spatial folding tolerance limit, constraining its specific value within the range of 0.10 to 0.15. Simultaneously, the data processing module calculates the geometric distance abrupt change ratio between adjacent trajectory points and compares this ratio with the spatial folding tolerance limit. When it is determined that the geometric distance abrupt change ratio is lower than the spatial folding tolerance limit, the data processing module outputs the conclusion that the spatial folding operation has been completed.

[0060] Based on the optimal time delay value and the optimal embedding dimension, the delay coordinate embedding operation performs a high-dimensional expansion on the discrete time series, generating a phase space reconstruction matrix that integrates chemical release kinetics and hardware thermal compensation dynamics.

[0061] The topological data analysis process for extracting short-lifetime homology classes is built upon the phase space reconstruction matrix. The analysis module constructs a Vittoria Lipps simplex complex within the high-dimensional feature space defined by the phase space reconstruction matrix, using matrix elements as vertex data points.

[0062] The analysis module incrementally increases the connectivity radius parameter at a multidimensional scale, traversing the connectivity states of the Vitoris-Lipps simplex complex and recording the continuous evolution range of topological holes from their generation to closure. After homology algebra mapping operations, the analysis module outputs a persistent homology barcode representing multidimensional topological invariants. The analysis module then scans the persistent homology barcodes and calculates the lifetime length of each homology class, finally extracting short-lifetime homology classes from the persistent homology barcodes whose lifetime values ​​are below a set threshold. The set threshold is pre-constrained to a range of 0.15 to 0.25; in this embodiment, the specific value of the set threshold is 0.2.

[0063] The feature recombination step is based on short-lifetime homology classes. The logic module determines that short-lifetime homology classes are distorted dimensional features introduced into the sensing signal by residual thermal fluctuations, and performs forced removal operations on the connected components corresponding to the distorted dimensional features in the phase space topology.

[0064] After the elimination operation is completed, the recombination logic extracts long-lifetime coherence classes from the persistent coherence barcodes whose lifetime exceeds a set threshold, and maps these long-lifetime coherence classes back to the original multidimensional space, outputting topological feature data that is topologically invariant to disturbances in the underlying hardware power supply voltage. The topological feature data is combined along the time and spatial dimensions and recombined into a chemical release feature vector containing 64 feature dimensions.

[0065] In the model computation stage, chemical release feature vectors are input into a deep neural network for forward propagation calculations. The data processing flow divides the chemical release feature vectors, containing 20,000 data points, into a training sample set for model fitting and a validation sample set for generalization evaluation, using an 8:2 ratio. The training sample set is then input batch by batch into the input layer of the deep neural network for forward propagation calculations. After passing through multiple layers of nonlinear activation functions, the deep neural network outputs a network mapping matrix representing the predicted release rate of hazardous substances at its output layer.

[0066] The computation module acquires the true label vectors calibrated by the gas chromatograph, quantifies the deviation distance between the network mapping matrix and the true label vectors using the mean squared error formula, and calculates the prediction error. Simultaneously, the computation module extracts the hidden feature data output from the hidden layer neurons of the deep neural network, and divides the hidden feature data into a chemical feature matrix characterizing the gas release pattern and a thermal fluctuation residual matrix characterizing hardware thermal stress drift, according to the feature dimension attribute. Using Pearson correlation coefficient calculation logic, the computation module calculates the inner product between the chemical feature matrix and the thermal fluctuation residual matrix, outputting an orthogonal penalty term to measure the degree of linear independence between the chemical feature matrix and the thermal fluctuation residual matrix within the feature mapping space.

[0067] The algorithm constructs a global objective function and updates the network node mapping weights based on gradient optimization. The algorithm logic defines the orthogonal penalty term as an independence constraint parameter to limit feature entanglement in the latent feature space, and introduces a constant of 0.05 as a weight adjustment coefficient. It then proportionally aggregates the prediction error and the independence constraint parameter to construct a global objective function that balances prediction accuracy and feature decoupling.

[0068] By calculating the partial derivatives of the global objective function with respect to the connection weights of each layer in the deep neural network, the optimization module calculates the loss gradient matrix of the global objective function. Based on the gradient descent direction indicated by the loss gradient matrix, the optimization module uses the adaptive moment estimation optimization algorithm to perform backpropagation, transforming the prediction error and independence constraint parameters into gradient descent step sizes, and synchronously updating the mapping weights of the network nodes in the deep neural network.

[0069] The validation logic for the prediction model is based on the convergence state of the global objective function. The validation sample set is independently input into the deep neural network after one round of weight updates for validation computation, and the system obtains the objective function loss value corresponding to the validation sample set.

[0070] The monitoring program establishes a sliding time window to monitor the relative change rate of the objective function loss value. When it is determined that the rate of change of the global objective function is lower than the convergence threshold for 50 consecutive preset iteration cycles, the monitoring program determines that the global objective function has entered the minimum convergence state. The convergence threshold is preset to a range of 0.0001 to 0.0005; in this embodiment, the specific value of the convergence threshold is selected as 0.0001. At the node that triggers the convergence state, the system stops the parameter update action and saves the network structure parameters below the convergence threshold to a non-volatile storage medium, completing the solidification of the algorithm model parameters and finally outputting the prediction model.

Claims

1. A method for constructing a predictive model for the release of harmful substances in coatings based on machine learning, characterized in that, include: Collect heat flux gradient data transferred to the surface of semiconductor power devices in the voltage regulation control branch; Substitute the heat flux gradient data into the semiconductor drift-diffusion equation to calculate the carrier mobility deviation of the semiconductor power device. A feedforward compensated gate voltage sequence is generated based on the carrier mobility deviation. The feedforward compensation gate voltage sequence is input to the control terminal of the semiconductor power device to maintain a constant power supply voltage to the volatile sensor. Extract the timing voltage evolution trajectory corresponding to the feedforward compensated gate voltage sequence; The constant power supply voltage is used to drive the volatile substance sensor to obtain the harmful substance concentration sensing waveform output by the volatile substance sensor; A phase space reconstruction operation is performed on the time-series voltage evolution trajectory and the sensor waveform of the concentration of harmful substances to generate a phase space reconstruction matrix. The phase space reconstruction matrix is ​​continuously homologous to eliminate the distortion dimension features caused by residual thermal fluctuations and extract the chemical release feature vector. The chemical release feature vector is input into a deep neural network to perform forward propagation calculation, outputs a network mapping matrix, obtains the true label vector, calculates the prediction error between the network mapping matrix and the true label vector, and calculates the orthogonal penalty term according to the orthogonal subspace projection constraint condition. The prediction error and the orthogonal penalty term are combined to construct a global objective function. The prediction error is backpropagated along the gradient descent direction to update the network node mapping weights of the deep neural network. When the global objective function converges, the network structure parameters corresponding to the network node mapping weights are saved to establish a prediction model.

2. The method for constructing a machine learning-based prediction model for the release of harmful substances in coatings according to claim 1, characterized in that, The heat flux gradient data acquired and transmitted to the surface of the semiconductor power devices within the voltage regulation control branch includes: A thin-film thermocouple array located on the surface of the semiconductor power device is used to obtain multi-point temperature difference potential signals; The multi-point temperature difference potential signal is differentially amplified to output the heat flux gradient data.

3. The method for constructing a machine learning-based prediction model for the release of harmful substances in coatings according to claim 2, characterized in that, The step of substituting the heat flux gradient data into the semiconductor drift-diffusion equation to solve for the carrier mobility deviation of the semiconductor power device includes: Substitute the heat flux gradient data into the semiconductor drift-diffusion equation to calculate the change in carrier scattering cross section caused by lattice thermal vibration. The carrier mobility deviation of the semiconductor power device is calculated based on the change in the carrier scattering cross section.

4. The method for constructing a machine learning-based prediction model for the release of harmful substances in coatings according to claim 3, characterized in that, The step of generating a feedforward compensated gate voltage sequence based on the carrier mobility deviation includes: The carrier mobility deviation is subjected to signal microprocessing operations to generate reverse voltage envelope data; The reverse voltage envelope data is analyzed by digital-to-analog conversion, and the feedforward compensated gate voltage sequence is output.

5. The method for constructing a machine learning-based prediction model for the release of harmful substances in coatings according to claim 4, characterized in that, The step of performing phase space reconstruction calculations on the time-series voltage evolution trajectory and the sensor waveform of the hazardous substance concentration to generate a phase space reconstruction matrix includes: The concentration sensing waveform of the harmful substance and the time-series voltage evolution trajectory are converted into a discrete time series. Calculate the mutual information parameters of the discrete time series and output the optimal time delay value; The discrete time series is processed using the spurious nearest neighbor algorithm, and the optimal embedding dimension is output. Based on the optimal time delay value and the optimal embedding dimension, the discrete time series is subjected to delay coordinate embedding operation to generate the phase space reconstruction matrix.

6. The method for constructing a machine learning-based prediction model for the release of harmful substances in coatings according to claim 5, characterized in that, The continuous cohomological calculation of the phase space reconstruction matrix, eliminating the distortion dimension features caused by residual thermal fluctuations, and extracting the chemical release feature vector includes: Construct a Vittoris Lipps simplex complex on the phase space reconstruction matrix; The connected radii of the Vitoris-Lipps simplex are traversed to perform homology calculations, and a continuously homology barcode is output. Extract short-lifetime homology classes from the continuous homology barcodes whose lifetime is below a set threshold.

7. The method for constructing a machine learning-based prediction model for the release of harmful substances in coatings according to claim 6, characterized in that, After extracting short-lifetime homogeneity classes with lifetimes below a set threshold from the continuous homogeneity barcodes, the method further includes: The short-lifecycle homogeneity class is determined to be a distortion dimension feature caused by the residual thermal fluctuation and is then removed. Extract the topological feature data corresponding to long-lifetime homology classes whose lifetime exceeds the set threshold from the continuous homology barcode; The topological feature data is reorganized into the chemical release feature vector.

8. The method for constructing a machine learning-based prediction model for the release of harmful substances in coatings according to claim 7, characterized in that, The process of inputting the chemical release feature vector into a deep neural network to perform forward propagation calculation, outputting a network mapping matrix, obtaining the true label vector, calculating the prediction error between the network mapping matrix and the true label vector, and calculating an orthogonal penalty term based on orthogonal subspace projection constraints includes: The chemical release feature vectors are divided into a training sample set and a validation sample set; The training sample set is input into the deep neural network to perform the forward propagation calculation, and the network mapping matrix is ​​output. Calculate the prediction error between the network mapping matrix and the real label vector; Extract the hidden feature data output from the hidden layer of the deep neural network, and divide the hidden feature data into a chemical feature matrix and a thermal fluctuation residual matrix; Calculate the orthogonal penalty term between the chemical characteristic matrix and the thermal fluctuation residual matrix.

9. The method for constructing a machine learning-based prediction model for the release of harmful substances in coatings according to claim 8, characterized in that, The process of summing the prediction error and the orthogonal penalty term to construct a global objective function, and backpropagating the prediction error along the gradient descent direction to update the network node mapping weights of the deep neural network, includes: The orthogonal penalty term is defined as an independence constraint parameter; By summing the prediction error and the independence constraint parameters, the global objective function is constructed. Calculate the loss gradient matrix of the global objective function; The prediction error and the independence constraint parameter are backpropagated according to the gradient descent direction of the loss gradient matrix, and the network node mapping weights of the deep neural network are updated.

10. The method for constructing a machine learning-based prediction model for the release of harmful substances in coatings according to claim 9, characterized in that, The step of saving the network structure parameters corresponding to the network node mapping weights to establish a prediction model when the global objective function converges includes: The verification sample set is input into the deep neural network with updated weights for verification calculation. When the rate of change of the global objective function is lower than the convergence threshold within a preset iteration period, the global objective function is determined to have converged. Save the network structure parameters at the convergence threshold and output the prediction model.

Images