Clean room particle concentration prediction method based on edge computing

Abstract

The present application relates to the technical field of edge computing, and particularly relates to a clean room particle concentration prediction method based on edge computing. A multi-branch time series prediction network containing different network depth prediction branches is deployed on an edge computing node; a computing power state vector is generated by obtaining the central processing unit occupancy rate and the memory remaining space of the edge computing node; a particle concentration difference value at an adjacent time is calculated to generate an environment change rate vector; both are input into a decision tree, the decision tree determines a selectable branch set according to the computing power state vector, and selects a target prediction branch according to the environment change rate vector; when the environment change rate vector exceeds a set threshold, the deepest first branch is forced to be selected for prediction. The present application dynamically schedules network branches of different depths, shortens the prediction response time under the particle concentration mutation working condition on the premise of meeting the edge node computing power constraint, and suppresses the prediction lag phenomenon of a single shallow model under the mutation working condition.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, and more specifically to a method for predicting particulate concentration in cleanrooms based on edge computing. Background Technology

[0002] Existing edge computing-based methods for predicting particulate concentration in cleanrooms typically deploy time-series prediction networks with fixed structures at edge nodes. Due to the limited hardware resources of edge nodes, the common practice is to tailor the model structure, deploying lightweight models with few parameters and shallow network depth. During operation, the edge nodes continuously collect sensor data, input it into the fixed lightweight model, perform forward propagation according to a predetermined network hierarchy, and output the predicted particulate concentration results for subsequent time steps. To maintain stable operation of the edge nodes, the deployed network structure, number of parameters, and computational graph remain unchanged throughout the entire operation.

[0003] Cleanrooms experience equipment start-up and shutdown cycles, leading to drastic changes in local airflow and abrupt shifts in particulate concentration data. Fixed, lightweight models, limited by network depth, cannot extract the rapid temporal changes under these abrupt changes, resulting in prediction lag. Deploying deep, complex models at edge nodes exceeds the capacity of their CPUs and memory, causing computational blockage or system crashes. Existing prediction methods with fixed network architectures fail to strike a balance between the limited computing power of edge nodes and the predictive demands of abrupt particulate concentration changes. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting particulate concentration in cleanrooms based on edge computing, which can effectively solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A cleanroom particulate concentration prediction method based on edge computing, the method comprising: deploying a multi-branch temporal prediction network on an edge computing node, wherein the multi-branch temporal prediction network contains multiple prediction branches with different network depths and parameter quantities;

[0007] The CPU occupancy rate and remaining memory space of the edge computing node are obtained, a computing power status vector is generated, and the particle concentration difference between adjacent moments in the cleanroom is calculated to generate an environmental change rate vector.

[0008] The computing power state vector and the environmental change rate vector are input into a preset decision tree. The decision tree determines a set of optional branches based on the computing power state vector and selects a target prediction branch from the set of optional branches based on the environmental change rate vector.

[0009] The target prediction branch is used to extract features from the time-series data of particle concentration and output the prediction results.

[0010] When the environmental change rate vector exceeds a set threshold, the decision tree forcibly selects the deepest first branch for prediction, and weights and fuses the historical output residual of the first branch with the output of the shallow branches to generate the final particle concentration prediction value.

[0011] Preferably, the step of generating the computing power state vector includes: setting a sliding time window of fixed length at the edge computing node, and collecting the CPU occupancy rate sequence and memory remaining space sequence corresponding to multiple timestamps according to the step size of the sliding time window;

[0012] The CPU utilization sequence and the memory remaining space sequence are subjected to max-min normalization processing. The normalized CPU utilization sequence and the memory remaining space sequence are then concatenated along the time dimension to obtain the initial computing power feature matrix.

[0013] The initial computing power feature matrix is ​​subjected to dimensionality reduction mapping using the principal component analysis algorithm. The principal component components with the largest corresponding feature values ​​in the initial computing power feature matrix are extracted, and the fixed-dimensional vector after dimensionality reduction mapping is used as the computing power state vector.

[0014] Preferably, the step of generating the environmental change rate vector includes: obtaining the particle concentration time series data within a first preset time length before the current moment, performing first-order difference calculation on the particle concentration time series data, and obtaining a sequence of concentration difference values ​​corresponding to multiple moments;

[0015] Remove the difference values ​​in the concentration difference value sequence whose absolute value is less than a preset noise threshold, and retain the target difference values ​​whose absolute value is greater than or equal to the preset noise threshold;

[0016] The target difference values ​​are rearranged in chronological order, and zero-padding is performed at the beginning and end of the sequence to generate an initial rate of change sequence.

[0017] Perform a Fourier transform on the initial rate of change sequence, extract the low-frequency coefficients for reconstruction, and use the reconstructed sequence as the environmental rate of change vector.

[0018] Preferably, the step of determining the set of optional branches based on the computing power state vector includes: obtaining the parameter quantity and required running memory space corresponding to each prediction branch in the multi-branch time series prediction network;

[0019] The remaining memory space in the computing power state vector is compared one by one with the required running memory space corresponding to each prediction branch. Prediction branches whose required running memory space is greater than the remaining memory space are eliminated to obtain the intermediate branch set.

[0020] Extract the CPU utilization rate from the computing power state vector, calculate the difference between the CPU utilization rate and the preset baseline utilization rate, remove the prediction branches from the intermediate branch set whose network depth is greater than the maximum tolerance depth corresponding to the difference, and use the remaining prediction branches as the optional branch set.

[0021] Preferably, the step of selecting the target prediction branch from the set of optional branches based on the environmental change rate vector includes: calculating the L2 norm of the environmental change rate vector and using the L2 norm as the current environmental fluctuation assessment value;

[0022] Obtain the preset fluctuation adaptation range corresponding to each prediction branch in the set of optional branches, and determine the preset fluctuation adaptation range into which the current environmental fluctuation assessment value falls;

[0023] The prediction branch corresponding to the preset fluctuation adaptation interval that falls into is selected as the target prediction branch. If the current environmental fluctuation assessment value falls into the overlapping area of ​​multiple preset fluctuation adaptation intervals at the same time, the prediction branch with the shallowest network depth is selected from the prediction branches corresponding to the multiple preset fluctuation adaptation intervals as the target prediction branch.

[0024] Preferably, the step of obtaining the historical output residual of the first branch includes: obtaining the first predicted value sequence of the first branch output within a second preset time length before the current time and the actual particle concentration sequence at the corresponding time.

[0025] Calculate the difference between the first predicted value sequence and the corresponding time data in the real particle concentration sequence to obtain the initial residual sequence;

[0026] The initial residual sequence is subjected to moving average filtering to remove high-frequency fluctuation components.

[0027] The filtered sequence is subjected to variance testing to remove abnormal residual points whose variance exceeds a preset variance threshold. The processed residual sequence is then used as the historical output residual of the first branch.

[0028] Preferably, the step of performing dimensionality reduction mapping on the initial computing power feature matrix using principal component analysis includes: calculating the covariance matrix of the initial computing power feature matrix, performing eigenvalue decomposition on the covariance matrix, and obtaining multiple eigenvalues ​​and corresponding eigenvectors.

[0029] Arrange the multiple feature values ​​in descending order, and calculate the cumulative contribution rate of the first m feature values ​​after arrangement, where m is a positive integer;

[0030] When the cumulative contribution rate reaches a preset contribution rate threshold, the feature vectors corresponding to the first m feature values ​​are extracted to construct a projection matrix.

[0031] The initial computing power feature matrix is ​​multiplied using the projection matrix to output a fixed-dimensional vector after dimensionality reduction mapping.

[0032] Preferably, the step of performing a Fourier transform on the initial rate of change sequence and extracting low-frequency coefficients for reconstruction includes: performing a fast Fourier transform on the initial rate of change sequence to obtain a frequency domain complex sequence;

[0033] Calculate the amplitude corresponding to each frequency point in the frequency domain complex sequence, and statistically distribute the amplitude according to the frequency magnitude to generate an amplitude-frequency distribution curve.

[0034] In the amplitude-frequency distribution curve, a first cutoff frequency corresponding to the amplitude attenuation to a preset attenuation ratio is determined, low-frequency coefficients with frequencies less than the first cutoff frequency are retained, and high-frequency coefficients with frequencies greater than or equal to the first cutoff frequency are set to zero.

[0035] Perform an inverse fast Fourier transform on the retained low-frequency coefficients and the zeroed high-frequency coefficients to output the reconstructed sequence.

[0036] Preferably, the step of calculating the L2 norm of the environmental change rate vector includes: extracting each element in the environmental change rate vector and calculating the square value of each element respectively;

[0037] Sum the squares of all the elements to obtain the sum of squares.

[0038] The square root of the sum of squares is calculated, and the result of the square root calculation is used as the L2 norm of the environmental change rate vector;

[0039] Before calculating the L2 norm, it is checked whether there are missing values ​​in the environmental change rate vector. If there are missing values, the average value of the element before and after the missing value is calculated using linear interpolation. The average value is then filled into the missing value position, and the L2 norm is calculated on the filled environmental change rate vector.

[0040] Preferably, the step of performing moving average filtering on the initial residual sequence includes: setting a third preset time window on the initial residual sequence;

[0041] The filter window is moved according to a preset sliding step size, and the arithmetic mean of the multiple residual data covered by the filter window at each stop is calculated.

[0042] The calculated arithmetic mean is used as the filtered output value at the time corresponding to the center position of the filtering window;

[0043] Traverse the initial residual sequence to generate a filter residual sequence composed of each filter output value. When the filter window moves to the beginning or end of the initial residual sequence, use the mirror symmetry method to symmetrically extend the virtual residual data before or after the beginning or end of the initial residual sequence to complete the residual data required by the filter window.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] 1. This invention deploys a multi-branch network containing prediction branches of different depths at edge nodes and constructs a two-factor linkage decision-making mechanism by combining computing power state vectors and environmental change rate vectors. The range of selectable branches is defined according to the real-time hardware load of the edge nodes, and the target branch is selected according to the changes in particle concentration. When dealing with sudden changes in particle concentration and the rate of change exceeds a set threshold, the deepest branch is forcibly invoked and weighted and fused with the historical output residuals and the outputs of shallow branches. This solves the contradiction between the limited edge computing power and the lag in prediction of sudden changes under a fixed network architecture, enabling edge nodes to cope with data changes within the limits of computing power.

[0046] 2. By acquiring the computing power sequence through a sliding time window and performing dimensionality reduction mapping using principal component analysis, redundant hardware fluctuation data is filtered out, reducing the dimension of the computing power state vector and decreasing the input computation of the decision tree. First-order difference calculation is performed on the concentration time series data, and difference values ​​with small absolute values ​​are removed. Then, Fourier transform is used to extract low-frequency coefficients for reconstruction, eliminating the interference of high-frequency noise on the rate of change judgment. When determining the set of optional branches, the remaining memory space and CPU occupancy rate are compared with the branch parameters and depth respectively to remove branches, avoiding system crashes caused by allocating branches beyond the remaining hardware resources. When acquiring historical output residuals, moving average filtering and variance testing are used to remove abnormal residual points, suppressing high-frequency fluctuation components in the residual sequence and ensuring the data stability of the weighted fusion process. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the overall execution process of the cleanroom particulate concentration prediction based on edge computing according to the present invention.

[0048] Figure 2 This is a flowchart of the computing power state vector generation process of the present invention;

[0049] Figure 3This is a flowchart of the environmental change rate vector generation process of the present invention;

[0050] Figure 4 This is a flowchart of the optional prediction branch set selection process of the present invention;

[0051] Figure 5 Selecting a decision flowchart for the target prediction branch of this invention;

[0052] Figure 6 This is a flowchart of the historical residual processing and predicted value weighted fusion of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0054] Please refer to Figure 1 This embodiment provides a cleanroom particulate concentration prediction method based on edge computing, executed on edge computing nodes deployed in the cleanroom. The edge computing nodes establish communication connections with multiple particulate concentration sensors within the cleanroom, collecting particulate concentration data from each monitoring point in real time. A multi-branch temporal prediction network is deployed on the edge computing nodes, comprising multiple prediction branches with different network depths and parameter counts. Each prediction branch adopts a temporal convolutional network architecture, achieving different network depths by stacking different numbers of causal convolutional layers. The first branch contains 12 causal convolutional layers with 2.3M parameters; the second branch contains 8 causal convolutional layers with 1.5M parameters; the third branch contains 4 causal convolutional layers with 0.8M parameters; and the fourth branch contains 2 causal convolutional layers with 0.3M parameters. Each causal convolutional layer has a kernel size of 3, and the dilation coefficient increases exponentially with the number of layers to expand the receptive field. All prediction branches share an input layer and an output layer. The input layer receives particulate concentration temporal data of length T, and the output layer outputs predicted particulate concentration values ​​for the next k time steps.

[0055] The system acquires the CPU occupancy and remaining memory space of the edge computing node, generates a computing power status vector, and calculates the particle concentration difference between adjacent moments in the cleanroom to generate an environmental change rate vector. The system monitoring module of the edge computing node collects CPU occupancy and remaining memory space data at a frequency of 1Hz and stores the collected data in a local circular buffer. Simultaneously, a particle concentration sensor collects particle concentration data in the cleanroom at a frequency of 0.5Hz, transmits it to the edge computing node via the MQTT protocol, and stores it in a time-series database after data verification and format conversion. When generating the computing power status vector, the CPU occupancy and remaining memory space data for the most recent N timestamps are read from the circular buffer, forming two one-dimensional sequences of length N. When generating the environmental change rate vector, the particle concentration time-series data within a first preset time length before the current moment is read from the time-series database, and the concentration difference between adjacent moments is calculated to form a concentration difference value sequence.

[0056] The computing power state vector and the environmental change rate vector are input into a preset decision tree. The decision tree determines a set of optional branches based on the computing power state vector and selects a target prediction branch from the set of optional branches based on the environmental change rate vector. The decision tree is constructed using the CART algorithm, with a depth of 3, containing 4 internal nodes and 5 leaf nodes. The root node of the decision tree uses the remaining memory space in the computing power state vector as the splitting feature, the left subtree uses the CPU utilization rate as the splitting feature, and the right subtree uses the L2 norm of the environmental change rate vector as the splitting feature. Each leaf node of the decision tree corresponds to one or more prediction branches. During the decision-making process, firstly, based on the remaining memory space and CPU utilization rate in the computing power state vector, prediction branches that can operate normally under the current hardware resource conditions are selected to form a set of optional branches. Then, based on the L2 norm of the environmental change rate vector, the prediction branch most suitable for the current environmental fluctuation level is selected from the set of optional branches as the target prediction branch.

[0057] The target prediction branch is used to extract features from the time-series particle concentration data and output prediction results. Time-series particle concentration data of length T is input into the target prediction branch, and features are extracted through multiple causal convolutional layers to obtain a high-dimensional time-series feature vector. This high-dimensional feature vector is then input into a fully connected layer for dimensionality transformation, outputting predicted particle concentration values ​​for the next k time steps. During feature extraction, the output of each causal convolutional layer undergoes ReLU activation and batch normalization to improve the network's training stability and convergence speed. The mean and variance of the batch normalization layer are calculated using batch data during model training and remain fixed during inference.

[0058] When the environmental change rate vector exceeds a set threshold, the decision tree forcibly selects the deepest first branch for prediction, and weights and fuses the historical output residuals of the first branch with the outputs of the shallow branches to generate the final predicted particle concentration value. The set threshold is determined based on the cleanroom's process requirements and historical data statistics. When the L2 norm of the environmental change rate vector exceeds this threshold, it indicates a drastic change in the cleanroom's particle concentration, requiring the use of the deepest first branch for prediction to capture rapidly changing characteristics. Simultaneously, the historical output residuals of the first branch over a past period and the current outputs of the shallow branches (second, third, or fourth branches) are obtained and weighted and fused. The weighting coefficients are determined based on the statistical characteristics of the historical residuals, ensuring that the fused predicted value combines the first branch's ability to capture abrupt changes with the stability of the shallow branches.

[0059] Table 1. Parameter configuration of each branch of the multi-branch temporal prediction network

[0060] First Branch 12 2.3 128 85 [0.8,+∞) Second branch 8 1.5 96 70 [0.5,1.2] Third Branch 4 0.8 64 55 [0.2,0.7] Fourth branch 2 0.3 32 40 [0,0.3]

[0061] Table 1 shows the parameter configurations of each prediction branch in the multi-branch time series prediction network. The first branch has the deepest network structure and the most parameters, enabling the extraction of complex time series features, but it requires more RAM and higher CPU utilization, making it suitable for environments with drastic fluctuations. The fourth branch has the shallowest network structure and the fewest parameters, requiring less hardware resources and suitable for stable environments. The preset fluctuation adaptation ranges of each branch partially overlap. When the environmental fluctuation assessment value falls into the overlapping region, the prediction branch with the shallowest network depth is selected to reduce computational resource consumption.

[0062] In this embodiment, each branch of the multi-branch temporal prediction network is trained simultaneously during the model training phase. The training dataset contains time-series data on particle concentration in a cleanroom under different operating conditions, including normal operation, equipment startup, equipment shutdown, and personnel entry / exit. The training process uses the mean squared error loss function and the Adam optimizer for parameter updates. The loss function for each branch is calculated independently, and only the network parameters of the corresponding branch are updated during backpropagation. After training, the network parameters of each branch are saved as independent weight files and deployed to edge computing nodes. During the inference phase, based on the selection result of the decision tree, the weight file of the corresponding branch is loaded for forward propagation calculation.

[0063] In a preferred embodiment, reference Figure 2The steps for generating the computing power state vector include: setting a fixed-length sliding time window on the edge computing node, and collecting CPU utilization sequences and memory remaining space sequences corresponding to multiple timestamps according to the step size of the sliding time window. The length of the sliding time window is set to 60 seconds, and the step size is set to 1 second, that is, the window slides once per second, and each window contains CPU utilization and memory remaining space data for the most recent 60 seconds. The CPU utilization value ranges from 0 to 100, representing the percentage of CPU usage at the current moment. The memory remaining space value ranges from 0 to the total memory capacity of the edge computing node, representing the amount of unused memory space at the current moment.

[0064] The CPU utilization sequence and the remaining memory space sequence are subjected to max-min normalization. The normalized CPU utilization sequence and the remaining memory space sequence are then concatenated along the time dimension to obtain an initial computing power feature matrix. Max-min normalization maps the data to the [0,1] interval, eliminating the influence of different units on subsequent processing. The normalization formula for the CPU utilization sequence is:

[0065]

[0066] in, This represents the original value of the CPU utilization rate at time t. This represents the minimum value in the CPU utilization sequence. This represents the maximum value in the CPU utilization sequence. This represents the normalized value of the CPU utilization rate at time t.

[0067] The normalization formula for the sequence of remaining memory space is:

[0068]

[0069] in, This represents the original value of the remaining memory space at time t. This represents the minimum value in the sequence of remaining memory space. This represents the maximum value in the sequence of remaining memory space. This represents the normalized value of the remaining memory space at time t.

[0070] The normalized CPU utilization sequence and the remaining memory space sequence are concatenated along the time dimension to form an initial computing power feature matrix X of 2 rows and N columns, where N is the length of the sliding time window. The element in the i-th row and j-th column of the initial computing power feature matrix X represents the normalized value of the i-th computing power feature corresponding to the j-th timestamp.

[0071] Principal Component Analysis (PCA) is used to perform dimensionality reduction mapping on the initial computing power feature matrix. This involves extracting the top-ranked principal component components based on their eigenvalues, and using the resulting fixed-dimensional vector as the computing power state vector. PCA maps high-dimensional features to a low-dimensional space through linear transformation while preserving the most important information. During dimensionality reduction, the covariance matrix of the initial computing power feature matrix is ​​first calculated. Then, eigenvalue decomposition is performed on the covariance matrix to obtain multiple eigenvalues ​​and their corresponding eigenvectors. The eigenvalues ​​are arranged in descending order, and the eigenvectors corresponding to the first m eigenvalues ​​are selected to construct a projection matrix. Matrix multiplication is then performed on the initial computing power feature matrix using the projection matrix to obtain the dimensionality-reduced eigenvectors, i.e., the computing power state vector.

[0072] Table 2. Examples of Principal Component Analysis Dimensionality Reduction Results

[0073] 1 1.872 62.4 62.4 2 0.756 25.2 87.6 3 0.234 7.8 95.4 4 0.138 4.6 100.0

[0074] Table 2 shows an example of the dimensionality reduction results of principal component analysis on the initial computing power feature matrix. As can be seen from the table, the cumulative contribution rate of the first two principal components reaches 87.6%, and the cumulative contribution rate of the first three principal components reaches 95.4%. In this embodiment, the preset contribution rate threshold is set to 95%, therefore the first three principal components are selected to construct the projection matrix, reducing the initial 2×60-dimensional feature matrix to a 3-dimensional computing power state vector. The dimensionality-reduced computing power state vector retains 95.4% of the information in the original data, while reducing the feature dimension and decreasing the input computation of the subsequent decision tree.

[0075] In this embodiment, the step of using principal component analysis (PCA) to perform dimensionality reduction mapping on the initial computing power feature matrix includes: calculating the covariance matrix of the initial computing power feature matrix, performing eigenvalue decomposition on the covariance matrix to obtain multiple eigenvalues ​​and corresponding eigenvectors. The formula for calculating the covariance matrix is:

[0076]

[0077] Where C is the covariance matrix and X is the initial computing power feature matrix. Let X be the transpose of X, and N be the length of the sliding time window.

[0078] Eigenvalue decomposition of the covariance matrix C yields:

[0079]

[0080] Where V is the eigenvector matrix, and each column corresponds to an eigenvector; It is an eigenvalue diagonal matrix, where the elements on the diagonal are the corresponding eigenvalues.

[0081] Arrange the multiple feature values ​​in descending order, and calculate the cumulative contribution rate of the first m feature values ​​after arrangement, where m is a positive integer. The formula for calculating the contribution rate of the i-th feature value is:

[0082]

[0083] in, Let be the i-th eigenvalue, and d be the dimension of the covariance matrix, i.e., the number of rows in the initial computing power feature matrix.

[0084] The formula for calculating the cumulative contribution rate of the first m eigenvalues ​​is:

[0085]

[0086] When the cumulative contribution rate reaches a preset contribution rate threshold, the feature vectors corresponding to the first m feature values ​​are extracted to construct a projection matrix. The projection matrix P is... A dimensional matrix, where the i-th column is the eigenvector corresponding to the i-th eigenvalue.

[0087] The initial computing power feature matrix is ​​multiplied using the projection matrix to output a fixed-dimensional vector after dimensionality reduction. The formula for calculating the dimensionality-reduced computing power state vector s is as follows:

[0088]

[0089] in, The mean vector is the d-dimensional value obtained by averaging the columns of the initial computing power feature matrix X.

[0090] In this embodiment, the length and step size of the sliding time window can be adjusted according to the hardware performance of the edge computing node and the frequency of changes in the operating conditions of the cleanroom. When the hardware performance of the edge computing node is high, the length of the sliding time window can be appropriately increased to obtain more historical computing power data and improve the accuracy of the computing power state vector. When the operating conditions of the cleanroom change frequently, the step size of the sliding time window can be appropriately decreased to increase the update frequency of the computing power state vector and reflect the changes in hardware resources in a timely manner.

[0091] In a preferred embodiment, reference Figure 3 The step of generating the environmental change rate vector includes: acquiring time-series data of particle concentration within a first preset time period prior to the current moment; performing first-order difference calculation on the particle concentration time-series data to obtain a sequence of concentration difference values ​​corresponding to multiple moments. The first preset time period is set to 30 minutes, meaning all particle concentration data within 30 minutes prior to the current moment are acquired. Since the sampling frequency of the particle concentration sensor is 0.5Hz, there are 900 sampling points within 30 minutes. The formula for first-order difference calculation is:

[0092]

[0093] in, This represents the particle concentration value at time t. This represents the particle concentration value at time t-1. This represents the concentration difference at time t.

[0094] First-order difference calculations can convert raw particle concentration time-series data into concentration change rate data, highlighting the variation characteristics in the data. When the particle concentration remains stable, the concentration difference value is close to 0; when the particle concentration changes, the absolute value of the concentration difference value increases.

[0095] Discrimination values ​​in the concentration difference sequence whose absolute values ​​are less than a preset noise threshold are removed, retaining target difference values ​​whose absolute values ​​are greater than or equal to the preset noise threshold. The preset noise threshold is determined based on the sensor's accuracy and the background noise level of the cleanroom, and is set to 0.01 mg / m³. Due to sensor measurement noise, when the particle concentration change is small, the concentration difference value may be mainly caused by noise and does not reflect the true environmental change. By removing difference values ​​whose absolute values ​​are less than the preset noise threshold, the influence of noise can be filtered out, improving the accuracy of the environmental change rate vector.

[0096] The target difference values ​​are rearranged in chronological order, and zero-padding is performed at the beginning and end of the sequence to generate an initial rate of change sequence. Since some difference values ​​were removed, the length of the target difference value sequence is shorter than the length of the original concentration difference value sequence. To maintain sequence length consistency, zero-padding is performed at the beginning and end of the target difference value sequence, making the length of the initial rate of change sequence the same as the length of the original concentration difference value sequence. Zero-padding does not change the main characteristics of the sequence and ensures that the input length for subsequent Fourier transforms remains fixed.

[0097] A Fourier transform is performed on the initial rate of change sequence, and the low-frequency coefficients are truncated for reconstruction. The reconstructed sequence is then used as the environmental rate of change vector. The Fourier transform converts a time-domain signal into a frequency-domain signal, separating different frequency components. The actual changes in cleanroom particulate concentration typically manifest as low-frequency signals, while noise primarily manifests as high-frequency signals. By truncating the low-frequency coefficients for reconstruction, high-frequency noise can be further filtered out, preserving the true environmental change information.

[0098] Table 3. Parameters for Low-Frequency Truncation of Fourier Transform

[0099] 899 0.5 0.01 18 0.1

[0100] Table 3 shows the parameter settings for the low-frequency band truncation of the Fourier transform. The initial rate of change sequence length is 899 (because 899 difference values ​​are obtained after first-order differencing of 900 sampling points), and the sampling frequency is 0.5Hz. The preset attenuation ratio is set to 0.1, that is, when the amplitude in the amplitude-frequency distribution curve attenuates to 10% of the maximum amplitude, the corresponding frequency is the first cutoff frequency. In this embodiment, the first cutoff frequency is 0.01Hz, and the corresponding number of retained coefficients is 18. By retaining the low-frequency coefficients with frequencies less than 0.01Hz, the high-frequency coefficients with frequencies greater than or equal to 0.01Hz are set to zero, and then an inverse Fourier transform is performed to obtain the reconstructed environmental rate of change vector.

[0101] In this embodiment, the step of performing a Fourier transform on the initial rate of change sequence and reconstructing the low-frequency coefficients includes: performing a Fast Fourier Transform on the initial rate of change sequence to obtain a frequency domain complex sequence. The formula for the Fast Fourier Transform is:

[0102]

[0103] in, Let N be the initial rate of change sequence, and N be the sequence length. Let k be a complex sequence in the frequency domain, where k is the frequency index and j is the imaginary unit.

[0104] Calculate the amplitude corresponding to each frequency point in the frequency domain complex sequence, and statistically distribute the amplitudes according to frequency magnitude to generate an amplitude-frequency distribution curve. The frequency value corresponding to frequency point k is:

[0105]

[0106] in, The sampling frequency.

[0107] The amplitude corresponding to frequency point k is:

[0108]

[0109] in, Let be the real part of the complex sequence in the frequency domain. Let be the imaginary part of the complex sequence in the frequency domain.

[0110] In the amplitude-frequency distribution curve, a first cutoff frequency corresponding to the amplitude attenuation to a preset attenuation ratio is determined. Low-frequency coefficients with frequencies lower than the first cutoff frequency are retained, while high-frequency coefficients with frequencies greater than or equal to the first cutoff frequency are set to zero. Preset attenuation ratio. Setting it to 0.1 means finding the minimum frequency. This makes it possible for all ,have ,in This represents the maximum amplitude value in the amplitude-frequency distribution curve.

[0111] Perform an inverse fast Fourier transform (IFT) on the retained low-frequency coefficients and the zeroed-out high-frequency coefficients to output the reconstructed sequence. The formula for the IFT is:

[0112]

[0113] in, The processed frequency domain complex sequence, This is the reconstructed environmental change rate vector.

[0114] In this embodiment, the first preset time length can be adjusted according to the process characteristics of the cleanroom and the variation pattern of particulate concentration. For cleanrooms where particulate concentration changes slowly, the first preset time length can be appropriately increased to obtain more historical data and improve the stability of the environmental change rate vector. For cleanrooms where particulate concentration changes rapidly, the first preset time length can be appropriately decreased to improve the response speed of the environmental change rate vector and reflect environmental changes in a timely manner.

[0115] In a preferred embodiment, reference Figure 4 The step of determining the set of selectable branches based on the computing power state vector includes: obtaining the parameter count and required runtime memory space corresponding to each prediction branch in the multi-branch temporal prediction network. The parameter count and required runtime memory space of each prediction branch are statistically analyzed after model training and pre-stored in the configuration file of the edge computing node. The required runtime memory space includes the memory occupied by model weights, the memory occupied by intermediate feature maps, and the memory occupied by temporary variables during computation.

[0116] The remaining memory space in the computing power state vector is compared one by one with the required operating memory space corresponding to each prediction branch. Prediction branches whose required operating memory space is greater than the remaining memory space are eliminated to obtain the intermediate branch set. The first principal component in the computing power state vector mainly reflects the change in the remaining memory space. Through a pre-established mapping relationship, the first principal component can be converted into the actual remaining memory space value. The converted remaining memory space value is compared with the required operating memory space of each prediction branch. If the required operating memory space of a prediction branch is greater than the current remaining memory space, then the branch cannot run under the current hardware conditions and is eliminated from the candidate branches.

[0117] The CPU utilization rate is extracted from the computing power state vector. The difference between the CPU utilization rate and the preset baseline utilization rate is calculated. Predicted branches with a network depth greater than the maximum tolerable depth corresponding to the difference are removed from the intermediate branch set, and the remaining predicted branches are used as the optional branch set. The second principal component in the computing power state vector mainly reflects the change in CPU utilization rate. Through a pre-established mapping relationship, the second principal component can be converted into the actual CPU utilization rate value. The preset baseline utilization rate is set to 30%, representing the normal CPU utilization rate of the edge computing node in an idle state. The difference between the CPU utilization rate and the preset baseline utilization rate reflects the current CPU load. The larger the difference, the heavier the CPU load and the shallower the network depth that can be supported.

[0118] Table 4. Correspondence between CPU load and maximum tolerable network depth

[0119] <10 12 Branches 1, 2, 3, and 4 10-25 8 Second, third, and fourth branches 25-40 4 Third and fourth branches >40 2 Fourth branch

[0120] Table 4 illustrates the relationship between CPU utilization difference and maximum tolerable network depth. When the CPU utilization difference is less than 10%, the CPU load is light, supporting networks up to 12 layers deep, meaning all prediction branches can run. When the CPU utilization difference is between 10% and 25%, the CPU load is moderate, supporting networks up to 8 layers deep, meaning the second, third, and fourth branches can run. When the CPU utilization difference is between 25% and 40%, the CPU load is heavy, supporting networks up to 4 layers deep, meaning the third and fourth branches can run. When the CPU utilization difference is greater than 40%, the CPU load is very heavy, supporting networks up to 2 layers deep, meaning only the fourth branch can run.

[0121] refer to Figure 5 The step of selecting a target prediction branch from the set of optional branches based on the environmental change rate vector includes: calculating the L2 norm of the environmental change rate vector and using the L2 norm as the current environmental fluctuation assessment value. The L2 norm of the environmental change rate vector reflects the severity of environmental change; the larger the L2 norm, the more severe the environmental change.

[0122] Obtain the preset fluctuation adaptation range corresponding to each prediction branch in the optional branch set, and determine which preset fluctuation adaptation range the current environmental fluctuation assessment value falls into. The preset fluctuation adaptation range of each prediction branch is pre-stored in the configuration file of the edge computing node, as shown in Table 1. The preset fluctuation adaptation range of the first branch is [0.8, +∞), which is suitable for cases of severe environmental fluctuations; the preset fluctuation adaptation range of the second branch is [0.5, 1.2], which is suitable for cases of moderate to severe environmental fluctuations; the preset fluctuation adaptation range of the third branch is [0.2, 0.7], which is suitable for cases of moderate environmental fluctuations; and the preset fluctuation adaptation range of the fourth branch is [0, 0.3], which is suitable for cases of minor environmental fluctuations.

[0123] The prediction branch corresponding to the preset fluctuation adaptation interval that the current environmental fluctuation assessment value falls into is selected as the target prediction branch. If the current environmental fluctuation assessment value falls into the overlapping area of ​​multiple preset fluctuation adaptation intervals, the prediction branch with the shallowest network depth is selected from the prediction branches corresponding to the multiple preset fluctuation adaptation intervals as the target prediction branch. For example, when the current environmental fluctuation assessment value is 0.6, it falls into the overlapping area of ​​the preset fluctuation adaptation interval [0.5, 1.2] of the second branch and the preset fluctuation adaptation interval [0.2, 0.7] of the third branch. In this case, the third branch with the shallower network depth is selected as the target prediction branch to reduce the consumption of computing resources.

[0124] In this embodiment, the step of calculating the L2 norm of the environmental rate of change vector includes: extracting each element of the environmental rate of change vector and calculating the square value of each element; summing the square values ​​of all elements to obtain a sum of squares; and taking the square root of the sum of squares, using the result as the L2 norm of the environmental rate of change vector. The formula for calculating the L2 norm of the environmental rate of change vector v is:

[0125]

[0126] in, Let be the i-th element of the environmental rate of change vector, and n be the length of the environmental rate of change vector.

[0127] Before calculating the L2 norm, it is checked whether there are missing values ​​in the environmental change rate vector. If missing values ​​exist, the average of the element preceding and following the missing value is calculated using linear interpolation. This average is then used to fill the missing value position, and the L2 norm is calculated on the filled environmental change rate vector. The formula for linear interpolation is as follows:

[0128]

[0129] in, Filler values ​​for missing value locations. The value of the element preceding the missing value. The value of the element at the position following the missing value.

[0130] refer to Figure 6 The step of obtaining the historical output residual of the first branch includes: obtaining the first predicted value sequence of the first branch output within a second preset time length before the current time and the corresponding actual particle concentration sequence at that time. The second preset time length is set to 10 minutes, that is, obtaining the predicted value and the corresponding actual value of the first branch within 10 minutes before the current time.

[0131] Calculate the difference between the first predicted value sequence and the corresponding time data in the actual particle concentration sequence to obtain the initial residual sequence. The formula for calculating the initial residual sequence is:

[0132]

[0133] in, Let be the predicted value of the first branch at time t. Let be the actual particle concentration value at time t. Let be the residual value at time t.

[0134] The initial residual sequence is subjected to a moving average filter to remove high-frequency fluctuations. The moving average filter smooths out high-frequency fluctuations in the data by averaging the data within a window.

[0135] The filtered sequence undergoes a variance test to remove outlier residuals whose variance exceeds a preset variance threshold. The processed residual sequence is then used as the historical output residual of the first branch. The variance test determines the presence of outliers by calculating the variance of the data within a window. If the variance within a window exceeds the preset variance threshold, an outlier residual is considered to exist within that window and is removed from the sequence.

[0136] In this embodiment, the step of performing moving average filtering on the initial residual sequence includes: setting a filtering window of a third preset time length on the initial residual sequence; moving the filtering window according to a preset sliding step size, and calculating the arithmetic mean of multiple residual data covered by the filtering window at each pause; using the calculated arithmetic mean as the filtered output value at the time corresponding to the center position of the filtering window; traversing the initial residual sequence to generate a filtered residual sequence composed of each filtered output value, wherein when the filtering window moves to the beginning or end of the initial residual sequence, virtual residual data is symmetrically extended before or after the beginning or end of the initial residual sequence using a mirror symmetry method to complete the residual data required by the filtering window. The third preset time length is set to 1 minute, the size of the filtering window is 30 data points, and the sliding step size is set to 1 data point.

[0137] The specific operation of the mirror symmetry method is as follows: For the first k data points at the beginning of the sequence, extend k virtual data points before the beginning of the sequence. The value of each virtual data point is a mirror image of the corresponding real data point, i.e., the virtual data point... Where i = 1, 2, ..., k. For the last k data points at the end of the sequence, k virtual data points are extended after the end of the sequence. The value of each virtual data point is a mirror image of the corresponding real data point, i.e., a virtual data point. , where i=1,2,...,k, and n is the length of the initial residual sequence.

[0138] The weighted fusion of the historical output residuals of the first branch and the output of the shallow branch yields the final particle concentration prediction value, calculated using the following formula:

[0139]

[0140] in, This is the final predicted value for particle concentration. This is the current predicted value for the first branch. This represents the current predicted value for the shallow branch. The average of the historical output residuals of the first branch. , and For the weighting coefficients, satisfying .

[0141] The weighting coefficients are determined based on the statistical characteristics of historical data. In this embodiment, Set to 0.6, Set to 0.3, The value was set to 0.1. By using weighted fusion, the ability of the first branch to capture mutation features, the stability of the shallow branch, and the correction effect of historical residuals were combined, thereby improving the prediction accuracy under mutation conditions.

Claims

1. A cleanroom particulate concentration prediction method based on edge computing, characterized in that, The method includes: deploying a multi-branch temporal prediction network on an edge computing node, wherein the multi-branch temporal prediction network contains multiple prediction branches with different network depths and parameter numbers; The CPU occupancy rate and remaining memory space of the edge computing node are obtained, a computing power status vector is generated, and the particle concentration difference between adjacent moments in the cleanroom is calculated to generate an environmental change rate vector. The computing power state vector and the environmental change rate vector are input into a preset decision tree. The decision tree determines a set of optional branches based on the computing power state vector and selects a target prediction branch from the set of optional branches based on the environmental change rate vector. The target prediction branch is used to extract features from the time-series data of particle concentration and output the prediction results. When the environmental change rate vector exceeds a set threshold, the decision tree forcibly selects the deepest first branch for prediction, and weights and fuses the historical output residual of the first branch with the output of the shallow branches to generate the final particle concentration prediction value.

2. The cleanroom particulate concentration prediction method based on edge computing according to claim 1, characterized in that, The steps for generating the computing power state vector include: setting a fixed-length sliding time window at the edge computing node, and collecting the CPU occupancy rate sequence and memory remaining space sequence corresponding to multiple timestamps according to the step size of the sliding time window; The CPU utilization sequence and the memory remaining space sequence are subjected to max-min normalization processing. The normalized CPU utilization sequence and the memory remaining space sequence are then concatenated along the time dimension to obtain the initial computing power feature matrix. The initial computing power feature matrix is ​​subjected to dimensionality reduction mapping using the principal component analysis algorithm. The principal component components with the largest corresponding feature values ​​in the initial computing power feature matrix are extracted, and the fixed-dimensional vector after dimensionality reduction mapping is used as the computing power state vector.

3. The cleanroom particulate concentration prediction method based on edge computing according to claim 1, characterized in that, The steps for generating the environmental change rate vector include: acquiring time-series data of particle concentration within a first preset time length before the current moment, performing first-order difference calculation on the time-series data of particle concentration, and obtaining a sequence of concentration difference values ​​corresponding to multiple moments; Remove the difference values ​​in the concentration difference value sequence whose absolute value is less than a preset noise threshold, and retain the target difference values ​​whose absolute value is greater than or equal to the preset noise threshold; The target difference values ​​are rearranged in chronological order, and zero-padding is performed at the beginning and end of the sequence to generate an initial rate of change sequence. Perform a Fourier transform on the initial rate of change sequence, extract the low-frequency coefficients for reconstruction, and use the reconstructed sequence as the environmental rate of change vector.

4. The cleanroom particulate concentration prediction method based on edge computing according to claim 1, characterized in that, The step of determining the set of optional branches based on the computing power state vector includes: obtaining the parameter quantity and required running memory space corresponding to each prediction branch in the multi-branch temporal prediction network; The remaining memory space in the computing power state vector is compared one by one with the required running memory space corresponding to each prediction branch. Prediction branches whose required running memory space is greater than the remaining memory space are eliminated to obtain the intermediate branch set. Extract the CPU utilization rate from the computing power state vector, calculate the difference between the CPU utilization rate and the preset baseline utilization rate, remove the prediction branches from the intermediate branch set whose network depth is greater than the maximum tolerance depth corresponding to the difference, and use the remaining prediction branches as the optional branch set.

5. The cleanroom particulate concentration prediction method based on edge computing according to claim 1, characterized in that, The step of selecting a target prediction branch from the set of optional branches based on the environmental rate of change vector includes: calculating the L2 norm of the environmental rate of change vector and using the L2 norm as the current environmental fluctuation assessment value; Obtain the preset fluctuation adaptation range corresponding to each prediction branch in the set of optional branches, and determine the preset fluctuation adaptation range into which the current environmental fluctuation assessment value falls; The prediction branch corresponding to the preset fluctuation adaptation interval that falls into is selected as the target prediction branch. If the current environmental fluctuation assessment value falls into the overlapping area of ​​multiple preset fluctuation adaptation intervals at the same time, the prediction branch with the shallowest network depth is selected from the prediction branches corresponding to the multiple preset fluctuation adaptation intervals as the target prediction branch.

6. The cleanroom particulate concentration prediction method based on edge computing according to claim 1, characterized in that, The steps for obtaining the historical output residuals of the first branch include: obtaining the first predicted value sequence of the first branch output within a second preset time length before the current time and the corresponding real particle concentration sequence; Calculate the difference between the first predicted value sequence and the corresponding time data in the real particle concentration sequence to obtain the initial residual sequence; The initial residual sequence is subjected to moving average filtering to remove high-frequency fluctuation components. The filtered sequence is subjected to variance testing to remove abnormal residual points whose variance exceeds a preset variance threshold. The processed residual sequence is then used as the historical output residual of the first branch.

7. The cleanroom particulate concentration prediction method based on edge computing according to claim 2, characterized in that, The steps of using principal component analysis to perform dimensionality reduction mapping on the initial computing power feature matrix include: calculating the covariance matrix of the initial computing power feature matrix, performing eigenvalue decomposition on the covariance matrix to obtain multiple eigenvalues ​​and corresponding eigenvectors; Arrange the multiple feature values ​​in descending order, and calculate the cumulative contribution rate of the first m feature values ​​after arrangement, where m is a positive integer; When the cumulative contribution rate reaches a preset contribution rate threshold, the feature vectors corresponding to the first m feature values ​​are extracted to construct a projection matrix. The initial computing power feature matrix is ​​multiplied using the projection matrix to output a fixed-dimensional vector after dimensionality reduction mapping.

8. The cleanroom particulate concentration prediction method based on edge computing according to claim 3, characterized in that, The step of performing a Fourier transform on the initial rate of change sequence and reconstructing the low-frequency coefficients includes: performing a fast Fourier transform on the initial rate of change sequence to obtain a frequency domain complex sequence; Calculate the amplitude corresponding to each frequency point in the frequency domain complex sequence, and statistically distribute the amplitude according to the frequency magnitude to generate an amplitude-frequency distribution curve. In the amplitude-frequency distribution curve, a first cutoff frequency corresponding to the amplitude attenuation to a preset attenuation ratio is determined, low-frequency coefficients with frequencies less than the first cutoff frequency are retained, and high-frequency coefficients with frequencies greater than or equal to the first cutoff frequency are set to zero. Perform an inverse fast Fourier transform on the retained low-frequency coefficients and the zeroed high-frequency coefficients to output the reconstructed sequence.

9. The cleanroom particulate concentration prediction method based on edge computing according to claim 5, characterized in that, The steps for calculating the L2 norm of the environmental rate of change vector include: extracting each element from the environmental rate of change vector and calculating the square value of each element respectively; Sum the squares of all the elements to obtain the sum of squares. The square root of the sum of squares is calculated, and the result of the square root calculation is used as the L2 norm of the environmental change rate vector; Before calculating the L2 norm, it is checked whether there are missing values ​​in the environmental change rate vector. If there are missing values, the average value of the element before and after the missing value is calculated using linear interpolation. The average value is then filled into the missing value position, and the L2 norm is calculated on the filled environmental change rate vector.

10. The cleanroom particulate concentration prediction method based on edge computing according to claim 6, characterized in that, The step of performing moving average filtering on the initial residual sequence includes: setting a third preset time-length filtering window on the initial residual sequence; The filter window is moved according to a preset sliding step size, and the arithmetic mean of the multiple residual data covered by the filter window at each stop is calculated. The calculated arithmetic mean is used as the filtered output value at the time corresponding to the center position of the filtering window; Traverse the initial residual sequence to generate a filter residual sequence composed of each filter output value. When the filter window moves to the beginning or end of the initial residual sequence, use the mirror symmetry method to symmetrically extend the virtual residual data before or after the beginning or end of the initial residual sequence to complete the residual data required by the filter window.

Images