A method for synergistic optimization of energy consumption and environmental factors in cleanrooms based on big data mining

Abstract

This invention relates to the field of industrial intelligent control technology, specifically a method for collaborative optimization of energy consumption and environmental factors in cleanrooms based on big data mining. The method includes: S1: Constructing a physical feature knowledge base, obtaining industry similarity, and improving the obtained target industry feature vector based on the industry similarity; S2: Determining the improvement method of the target industry feature vector based on the industry similarity; S3: Pre-training the spatiotemporal graph convolutional network after gradient step size correction based on the source industry feature vector to determine the final mapping model and obtain the final target industry feature vector. This invention, by constructing a physical feature knowledge base, establishing inter-industry mapping relationships, and performing industry similarity evaluation, not only achieves adaptive improvement of cross-industry feature vectors and solves the data distribution offset problem, but also ensures the reliability of the migration through expert intervention thresholds.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent control technology, specifically a method for synergistic optimization of energy consumption and environmental factors in cleanrooms based on big data mining. Background Technology

[0002] Cleanrooms are widely used in production environments with strict requirements for air cleanliness, such as electronics manufacturing, biomedicine, food processing, and aerospace. Their core function is to maintain environmental indicators such as temperature, humidity, particulate matter concentration, and microbial count within specified ranges through air purification systems.

[0003] Currently, most cleanrooms still employ static control strategies based on setpoints, such as fixed temperature, humidity, and air volume. While this control method is simple and reliable, it lacks adaptability to dynamic environmental and production process changes, easily leading to "over-cleaning" or "energy waste."

[0004] With the development of IoT, cloud computing, and artificial intelligence technologies, especially the maturity of big data mining technology, new methods have been provided for optimizing the operation of complex systems. By conducting multi-dimensional analysis of data collected by a large number of sensors in cleanrooms, the nonlinear relationship between energy consumption and environmental factors can be revealed, potential patterns can be discovered, and these can be used to optimize control strategies.

[0005] Chinese invention patent CN118168130A discloses a high-level multi-channel automatic wind capture and delivery system, comprising a wind pressure monitoring module, a dynamic wind capture control module, an intelligent ventilation decision-making module, an airflow optimization module, and an airflow regulation module, all connected via communication. The wind pressure monitoring module collects wind pressure data in real time; the dynamic wind capture control module adjusts the wind capture outlets based on the monitoring data; and the intelligent ventilation decision-making module analyzes the data using machine learning algorithms to formulate the optimal ventilation strategy. The airflow optimization module optimizes the wind flow path, and the airflow regulation module automatically adjusts the airflow according to the intelligent ventilation strategy, controlling the fan output. This invention can intelligently and efficiently capture and utilize wind energy, maximizing energy utilization efficiency, reducing energy waste, and is environmentally friendly, contributing to sustainable development.

[0006] However, in the process of co-optimizing energy consumption and environmental factors in cleanrooms, the differences in physical mechanisms between different production scenarios make it difficult for models to be universally applicable. Taking the semiconductor and pharmaceutical industries as examples, semiconductor workshops use laminar flow air supply systems to maintain stable unidirectional airflow, while pharmaceutical workshops rely on turbulent flow dilution to achieve uniform purification of the space. This fundamental difference in airflow organization results in completely different dynamic response relationships in terms of heat and humidity load distribution, energy consumption characteristics, and cleanliness maintenance mechanisms. This not only causes a significant shift in data distribution but also makes it difficult to directly transfer and apply optimization models trained on data from a single industry. Summary of the Invention

[0007] The purpose of this invention is to provide a method for synergistic optimization of energy consumption and environmental factors in cleanrooms based on big data mining, so as to solve the problem of data distribution offset.

[0008] The technical solution of this invention is: a method for synergistic optimization of energy consumption and environmental factors in cleanrooms based on big data mining, comprising:

[0009] S1: Construct a physical feature knowledge base: Construct a physical feature knowledge base based on the source industry feature vector and the target industry feature vector. At the same time, obtain industry similarity based on the physical feature knowledge base, and improve the obtained target industry feature vector based on the industry similarity.

[0010] S2: Construct a hybrid model: Based on the industry similarity, determine the improvement method for the target industry feature vector, and obtain the final source industry feature vector, including:

[0011] S2.1: Determine the feature vector: Compare the industry feature similarity with a preset similarity threshold range, and determine the final target industry feature vector based on the comparison result, specifically as follows:

[0012] When the industry feature similarity is greater than the upper limit of the preset similarity threshold range, the target industry feature vector is the final target industry feature vector. When the industry feature similarity is within the preset similarity threshold range, steps S2.2-S3 are executed to construct a spatiotemporal graph convolutional network and obtain the final target industry feature vector. When the industry feature similarity is less than the lower limit of the similarity threshold range, adjustments are made by experts.

[0013] S2.2: Model Adjustment: By constructing the spatiotemporal graph convolutional network, the relative error corresponding to the initial target industry feature vector is obtained, and the gradient step size of the spatiotemporal graph convolutional network is corrected according to the relative error.

[0014] S3: Obtain the final target industry feature vector: Based on the source industry feature vector, pre-train the spatiotemporal graph convolutional network after gradient step correction, determine the final mapping model based on the pre-training results, and obtain the final target industry feature vector through the final mapping model.

[0015] Furthermore, obtaining industry similarity includes:

[0016] S1.1: Feature extraction: Using hot-wire anemometers and laser trackers, Reynolds number characteristic data and mainstream kinetic energy ratio in the semiconductor industry are obtained; using PIV testing and data analysis, turbulence intensity spectrum and spatial particulate matter concentration distribution in the pharmaceutical industry are obtained.

[0017] S1.2: Construct a knowledge base: Based on the Reynolds number characteristic data, the proportion of mainstream kinetic energy, the turbulence intensity spectrum, and the spatial particulate matter concentration distribution, construct a basic parameter table, and determine the mapping relationship between industries based on the basic parameter table;

[0018] S1.3: Industry Similarity Assessment: Based on the mapping relationship between the industries, obtain the source industry feature vector and the corresponding target industry feature vector, determine the industry feature similarity, and improve the target industry feature vector based on the industry feature similarity. The formula for obtaining the industry feature similarity is as follows:

[0019]

[0020] Where: S represents industry feature similarity, Φ s Let Φ be the source industry feature vector. t This represents the feature vector of the target industry.

[0021] Furthermore, the mapping relationships between industries were determined, including:

[0022] S1.2.1: Establish a basic parameter table: Use the Reynolds number characteristic data, the proportion of mainstream kinetic energy, the turbulence intensity spectrum, and the spatial particulate matter concentration distribution as the threshold range in the basic parameter table. At the same time, construct the basic parameter table according to the determined industry and parameter type.

[0023] S1.2.2: Perform industry relationship mapping: Based on the aforementioned basic parameter table, perform laminar-turbulent wind speed conversion and determine the mapping relationship between turbulent dissipation rate and Reynolds number, specifically as follows:

[0024]

[0025] Where: v t For the target turbulent wind speed, v l ε is the source laminar wind speed, D is the characteristic size, ε is the turbulent dissipation rate, and Re is the Reynolds number.

[0026] Furthermore, the industry feature similarity is compared with a preset similarity, and the target industry feature vector is improved based on the comparison result, specifically as follows:

[0027] If the industry feature similarity is less than the preset similarity, then the next step S2 is executed to improve the target industry feature vector; otherwise, the target industry feature vector is not improved.

[0028] Furthermore, the mapping relationship is verified through energy conservation and cleanliness equivalence, including:

[0029] W1: Verification: Based on the laminar kinetic energy density and turbulent kinetic energy density, obtain the energy ratio and perform energy conservation verification. At the same time, based on the set target mixing uniformity and the measured mixing uniformity, obtain the cleanliness ratio and perform cleanliness conservation verification.

[0030] W2: Adjusting the fan speed: Based on the difference between the target mixing uniformity and the measured mixing uniformity, determine the adjustment amount of the air supply volume. Simultaneously, based on the adjustment amount of the air supply volume, adjust the fan speed in stages. The formulas for adjusting the air supply volume and the fan speed are as follows:

[0031]

[0032] Where: ΔQ is the adjustment amount of the air supply volume, h is the adjustment coefficient, and U tar For the target mixing uniformity, U act To measure the mixing uniformity, Q0 is the baseline value of the air supply volume, ω i+1 Let ω be the fan speed corresponding to the (i+1)th stage. i The fan speed corresponding to the i-th stage;

[0033] W3: Adjust the basic parameter table: Based on the threshold range in the basic parameter table and the threshold values ​​of the last 10 successfully mapped data, obtain a new threshold range, specifically:

[0034]

[0035] Wherein: T new For the new threshold, T old The original threshold, This is the arithmetic mean of the most recent 10 successes. This is the historical average.

[0036] Furthermore, the energy ratio is compared with a preset energy ratio threshold range, and energy conservation is verified based on the comparison results, specifically as follows:

[0037] When the energy ratio is within the preset energy ratio threshold range, the mapping result conforms to energy conservation; otherwise, the mapping result does not conform to energy conservation, and step W3 is executed.

[0038] The formula for obtaining the energy ratio is as follows:

[0039]

[0040] in: E is the energy ratio. lam E is the laminar kinetic energy density. tur ρ is the turbulent kinetic energy density. s For air density, v t For the target turbulent wind speed, v l The source laminar flow velocity is TKE, and the percentage of turbulent kinetic energy is TKE.

[0041] Furthermore, the cleanliness ratio is compared with a preset cleanliness ratio threshold range, and based on the comparison results, cleanliness conservation verification is performed, specifically as follows:

[0042] When the cleanliness ratio is within the preset cleanliness ratio threshold range, the mapping result conforms to the cleanliness conservation; otherwise, the mapping result does not conform to the cleanliness conservation, and steps W2 and W3 are executed.

[0043] The formula for obtaining the cleanliness ratio is as follows:

[0044]

[0045] in: For cleanliness ratio, U tar For the target mixing uniformity, U act This represents the measured uniformity of mixing.

[0046] Furthermore, the gradient stride of the spatiotemporal graph convolutional network is corrected, including:

[0047] S2.2.1: Relaxation Factor Adjustment: Based on the total inflow corresponding to the source industry feature vector and the total outflow corresponding to the target industry feature vector, the relative error of the inflow and outflow mass flow is obtained, and the relaxation factor in the spatiotemporal graph convolutional network is adjusted accordingly, specifically as follows:

[0048]

[0049] Where: τ new τ is the adjusted relaxation factor. old Here, α is the relaxation factor before adjustment, b is the maximum allowable adjustment range, and E is the error sensitivity coefficient. mass This is relative error;

[0050] S2.2.2: Learning Rate Adjustment: Based on the target industry feature vector, obtain the momentum residual and adjust the learning rate in the spatiotemporal graph convolutional network, specifically as follows:

[0051]

[0052] Where: η new Here, η0 is the adjusted learning rate, r is the industry sensitivity coefficient, and E is the learning rate before adjustment. mom For momentum residuals;

[0053] S2.2.3: Constructing a dynamic graph: Using the coordinates of measurement points in the source industry feature vector as node coordinates, and based on the target industry feature vector, constructing a matrix between different nodes, specifically:

[0054]

[0055] Among them: A m，n Let be the matrix element between the m-th node and the n-th node, e be the base of the natural logarithm, and Cov(v) be the matrix element between the m-th node and the n-th node. m v n P represents the velocity covariance. m Let P be the spatial coordinates of the m-th node. n Let be the spatial coordinates of the nth node.

[0056] Furthermore, the final target industry feature vector is obtained, including:

[0057] S3.1: Model Pre-training: Based on the source industry feature vector, a pre-training loss function is set. Simultaneously, the source industry feature vector is used as the input to the spatiotemporal graph convolutional network after gradient step correction. The output obtains the corresponding target industry feature vector, and the corresponding feature retention rate is obtained. Specifically:

[0058]

[0059] Where: ω b To preserve the feature representation capability, W fix X is the frozen Chebyshev convolution kernel parameter matrix. new Input the feature matrix for the target industry;

[0060] S3.2: Determine the final model: Compare the feature retention rate with a preset retention rate threshold, and determine the final mapping model based on the comparison result, specifically as follows:

[0061] When the feature retention rate is greater than the preset retention rate threshold, the spatiotemporal graph convolutional network corresponding to the feature retention rate is the final mapping model; otherwise, repeat steps S2.2-S3.2 until the feature retention rate is greater than the preset retention rate threshold.

[0062] S3.3: Perform model prediction: Based on the target industry feature vector, set the training loss function, and use the source industry feature vector as the input of the final mapping model to obtain the final target industry feature vector.

[0063] This invention provides an improved method for the synergistic optimization of energy consumption and environmental factors in cleanrooms based on big data mining. Compared with existing technologies, it has the following improvements and advantages:

[0064] Firstly, this invention constructs a physical feature knowledge base, establishes inter-industry mapping relationships, and conducts industry similarity assessment. This not only achieves adaptive improvement of cross-industry feature vectors and solves the problem of data distribution offset, but also ensures the reliability of migration through expert intervention thresholds.

[0065] Secondly, this invention uses a spatiotemporal graph convolutional network to dynamically adjust model parameters, namely, to dynamically adjust the relaxation factor and learning rate. This allows the model to respond in real time to fluctuations in environmental factors (such as wind speed and particulate matter concentration), thereby improving control accuracy. At the same time, by constructing a dynamic graph, spatial correlation can be captured and airflow organization can be optimized.

[0066] Thirdly, this invention, through energy conservation and cleanliness equivalence verification, can ensure that energy consumption optimization does not exceed the cleanliness threshold, and at the same time, it can be linked to the fan speed for adjustment, thereby balancing response speed and stability through phased adjustment of air volume.

[0067] Fourthly, this invention freezes the convolution kernel using source industry data and obtains the feature retention rate. At the same time, it optimizes the loss function based on the target industry data. This not only reduces the dependence on the amount of target industry data and accelerates model convergence, but also ensures the effectiveness of transfer by using the retention rate threshold. Attached Figure Description

[0068] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0069] Figure 1 This is a flowchart illustrating the process of determining the target industry feature vector correction method in this invention;

[0070] Figure 2 This is a schematic diagram of the process for obtaining the final target industry feature vector in this invention;

[0071] Figure 3 This is an analysis chart of the feature retention rate in this invention. Detailed Implementation

[0072] 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 embodiments of the present invention, and not all embodiments. 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.

[0073] It should be noted that in the description of this invention, the terms "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0074] Furthermore, it should be understood that, for ease of description, the dimensions of the various components shown in the accompanying drawings are not drawn to actual scale; for example, the thickness or width of some layers may be exaggerated relative to other layers.

[0075] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined or described in one figure, it will not need to be discussed or described in detail in the description of the subsequent figures.

[0076] Example 1

[0077] refer to Figures 1-3 This embodiment provides a method for synergistic optimization of energy consumption and environmental factors in cleanrooms based on big data mining. This method specifically includes the following steps:

[0078] Step S1: Construct a physical feature knowledge base. This involves constructing a physical feature knowledge base based on the source industry feature vector and the target industry feature vector (specifically, the laminar flow features in the semiconductor industry and the turbulent flow features in the pharmaceutical industry). Then, based on this physical feature knowledge base, the industry similarity between different industries is obtained. Specifically:

[0079] Step S1.1: Feature Extraction. This involves using relevant instruments to obtain the source industry feature vector and the target industry feature vector. In this embodiment, the semiconductor industry and the pharmaceutical industry are used as examples for specific explanation. The semiconductor industry is the source industry, and the pharmaceutical industry is the target industry.

[0080] Reynolds number characteristic data and mainstream current energy ratio in the semiconductor industry were obtained using a hot-wire anemometer and laser tracker. Specifically, in the wafer processing area, a 5*5 grid of measuring points was set with the wafer transport path as the center, and the spacing between adjacent measuring points was set to 0.5m. The wind speed and turbulence intensity at the grid measuring points were obtained using a hot-wire anemometer to acquire Reynolds number characteristic data in the semiconductor industry.

[0081]

[0082] Where: Re is the Reynolds number, ρ is the fluid density, v is the characteristic velocity, L is the characteristic length, and μ is the fluid dynamic viscosity.

[0083] Furthermore, the velocity at each measuring point is decomposed into vectors, and the proportion of mainstream kinetic energy is obtained based on the velocity data after vector decomposition. Specifically:

[0084]

[0085] Among them: E ratio The proportion of main kinetic energy, v x The wind speed in the prevailing direction, v y v is the lateral wind speed component. z This represents the vertical wind speed component.

[0086] During the specific implementation process, the laminar flow data in the semiconductor workshop were as follows: fluid density 1.2 kg / m³ 3 Characteristic flow velocity: 0.45 m / s; characteristic length: 0.5 m; fluid dynamic viscosity: 1.8 × 10⁻⁶. -5 Pa\cdotps corresponds to a Reynolds number of approximately 15000. Further, based on the wind speed in the mainstream direction, the obtained mainstream kinetic energy is 5.12, and based on the wind speed in the mainstream direction, the magnitudes of the lateral and vertical wind speed components, the obtained total kinetic energy is 5.32, corresponding to a mainstream kinetic energy ratio of 96.2%.

[0087] Furthermore, through PIV testing and data analysis, the turbulence intensity spectrum and spatial particulate matter concentration distribution in the pharmaceutical industry are obtained. Specifically, within the filling line operating area, transient turbulence characteristics are captured at a sampling frequency of 500Hz, specifically:

[0088] E(k)=C k ε 2 / 3 k -5 / 3

[0089] Where: E(k) is the turbulent kinetic energy at wave number k, C k ε is the Kolmogorov constant, ε is the turbulent dissipation rate, and k is the wave number.

[0090] Furthermore, using a laser particle counter with a sampling flow rate of 1 cfm and a single sampling time of 1 minute, the particulate matter concentration at each sampling moment is obtained, and the standard deviation and average value of the particulate matter concentration are calculated. Simultaneously, based on the obtained standard deviation and average value of the particulate matter concentration, the spatial particulate matter concentration distribution is determined, specifically:

[0091]

[0092] Where: U is the mixing uniformity, σC The standard deviation of particulate matter concentration. This represents the average concentration of particulate matter.

[0093] During the implementation process, the particulate matter concentrations obtained from five different sampling points were 352, 367, 341, 358 and 365, respectively. The average particulate matter concentration was 356.6, the standard deviation of the particulate matter concentration was 9.8, and the corresponding mixing uniformity was 97.25%.

[0094] Step S1.2: Construct a knowledge base. This involves constructing a basic parameter table based on the Reynolds number characteristic data and mainstream kinetic energy ratio in the semiconductor industry, and the turbulent kinetic energy characteristic data and mixing uniformity in the pharmaceutical industry, obtained in Step S1.1. Simultaneously, based on the constructed basic parameter table, the mapping relationships between industries are determined. Specifically, as follows:

[0095] Step S1.2.1: Establish a basic parameter table. This involves using the Reynolds number characteristic data and mainstream kinetic energy ratio from the semiconductor industry, and the turbulent dissipation rate data and mixing uniformity from the pharmaceutical industry, obtained in Step S1.1, as threshold ranges in the basic parameter table. These are then combined with the corresponding industry and parameter type to construct the basic parameter table.

[0096] In the specific implementation process, based on the obtained Reynolds number characteristic data, the proportion of mainstream kinetic energy, the turbulent dissipation rate data, and the mixing uniformity, the basic parameter table is constructed as shown in Table 1 below:

[0097] Table 1: Basic Parameter Table

[0098]

[0099] Step S1.2.2: Perform industry relationship mapping. This involves converting laminar to turbulent wind speeds based on the data in the basic parameter table from step S1.2.1, and determining the mapping relationship between turbulent dissipation rate and Reynolds number. Specifically:

[0100]

[0101] Where: v t For the target turbulent wind speed, v l ε is the source laminar wind speed, D is the characteristic size, ε is the turbulent dissipation rate, and Re is the Reynolds number.

[0102] In the specific implementation process, the source laminar flow velocity in the semiconductor workshop was 0.48 m / s, the Reynolds number was 18000, and the characteristic dimension was 5 m. Therefore, the corresponding turbulent dissipation rate in the pharmaceutical workshop was 1.15 * 10⁻⁶. -3 m 2 / s 3The target turbulent wind speed is 1.24 m / s.

[0103] Step S1.3: Industry Similarity Assessment. Based on the industry relationship mapping in Step S1.2.2, the source industry feature vector and the corresponding target industry feature vector are obtained, and the industry feature similarity between the source and target industries is calculated. The obtained industry feature similarity is then compared with a preset similarity, and the target industry feature vector is improved based on the comparison results. Specifically:

[0104] If the obtained industry feature similarity is less than the preset similarity, then proceed to the next step S2 to improve the target industry feature vector. Otherwise, the obtained target industry feature vector is not improved and becomes the final target industry feature vector.

[0105] In this embodiment, the formula for obtaining industry feature similarity is as follows:

[0106]

[0107] Where: S represents industry feature similarity, Φ s Let Φ be the source industry feature vector. t This represents the feature vector of the target industry.

[0108] In the specific implementation process, the source industry is set as the semiconductor industry, and the target industry is set as the pharmaceutical industry. That is to say, the source industry feature vector is the Reynolds number and the proportion of mainstream kinetic energy in the semiconductor industry, and the target industry feature vector is the turbulent dissipation rate and mixing uniformity in the pharmaceutical industry.

[0109] Furthermore, if the source industry feature vector is set to [1.75, 0.963] and the target industry feature vector is set to [0.15, 0.927], the corresponding industry feature similarity is approximately 0.464. Meanwhile, the preset similarity in this embodiment is set to 0.6, so the target industry feature vector corresponding to this industry feature similarity needs further improvement.

[0110] Step S2: Construct a hybrid model. This involves comparing the industry feature similarity obtained in step S1.3 with a preset similarity threshold range, and determining the improvement method for the target industry feature vector based on the comparison results, thus obtaining the final target industry feature vector. Specifically:

[0111] Step S2.1: Determine the feature vector. This involves comparing the industry feature similarity obtained in step S1.3 with a preset similarity threshold range, and determining the final source industry feature vector based on the comparison result. Specifically:

[0112] When the obtained industry feature similarity is greater than the upper limit of the preset similarity threshold range, the target industry feature vector obtained through industry relationship mapping in step S1.2.2 is the final target industry feature vector. When the obtained industry feature similarity is within the preset similarity threshold range, steps S2.2-S3 are executed to construct a spatiotemporal graph convolutional network, and the final source industry feature vector is obtained through the constructed spatiotemporal graph convolutional network. When the obtained industry feature similarity is less than the lower limit of the similarity threshold range, expert intervention is performed to re-obtain the target industry feature vector.

[0113] Step S2.2: Model Adjustment. This involves obtaining the relative error corresponding to the initial target industry feature vector using the constructed spatiotemporal graph convolutional network, and adjusting the gradient stride of the spatiotemporal graph convolutional network based on the obtained relative error. Simultaneously, the adjusted spatiotemporal graph convolutional network is determined based on the adjusted gradient stride. Specifically:

[0114] Step S2.2.1: Relaxation Factor Adjustment. This involves obtaining the import / export quality flow difference based on the total import flow corresponding to the source industry feature vector and the total export flow corresponding to the target industry feature vector. Specifically:

[0115]

[0116] in: Let ρ be the difference in mass flow rate between the inlet and outlet, ρ be the fluid density, u be the velocity vector, and dA be the area vector of the infinitesimal element. For total import flow, This represents the total export volume.

[0117] Furthermore, based on the obtained difference in import and export quality flow rates and the total import flow rate corresponding to the source industry feature vector, the relative error between the two is determined, specifically as follows:

[0118]

[0119] Where: E mass This is a relative error. Let ρ be the difference in mass flow rate between inlet and outlet, ρ be the fluid density, u be the velocity vector, and dA be the area vector of the infinitesimal element.

[0120] Furthermore, based on the obtained relative error, the relaxation factor in the constructed spatiotemporal graph convolutional network is adjusted, specifically as follows:

[0121]

[0122] Where: τ new τ is the adjusted relaxation factor. oldHere, α is the relaxation factor before adjustment, b is the maximum allowable adjustment range, and E is the error sensitivity coefficient. mass This represents the relative error.

[0123] In the specific implementation process, when the obtained relative error is 0.4%, and according to actual needs, the error sensitivity coefficient is set to 20, the maximum allowable adjustment range is set to 0.05, and the corresponding relaxation factor adjustment amount is 0.0145.

[0124] Step S2.2.2: Learning rate adjustment. This involves obtaining the corresponding momentum residual based on the target industry feature vector, specifically:

[0125]

[0126] Among them: E mom Let V be the momentum residual and V be the velocity vector field. For the gradient operator, v d Let t be the kinematic viscosity and t be the time variable.

[0127] Furthermore, based on the obtained momentum residuals, the learning rate in the constructed spatiotemporal graph convolutional network is adjusted, specifically as follows:

[0128]

[0129] Where: η new Here, η0 is the adjusted learning rate, r is the industry sensitivity coefficient, and E is the learning rate before adjustment. mom This is the momentum residual.

[0130] In the specific implementation process, when the obtained momentum residual is 3*10 -4 m 2 / s 2 At the same time, based on actual needs, the industry sensitivity coefficient is set to 15, and the initial learning rate is 0.001. Therefore, the corresponding relaxation factor adjustment is 8.2 * 10^6. -4 .

[0131] Step S2.2.3: Construct a dynamic graph. The coordinates of the measurement points obtained by the laser tracker are used as node coordinates. These coordinates are then combined with the velocity covariance obtained through PIV testing to construct a matrix between different nodes. Specifically:

[0132]

[0133] Among them: A m,n Let be the matrix element between the m-th node and the n-th node, e be the base of the natural logarithm, and Cov(v) be the matrix element between the m-th node and the n-th node. m v n P represents the velocity covariance.m Let P be the spatial coordinates of the m-th node. n Let be the spatial coordinates of the nth node.

[0134] Step S3: Obtain the final target industry feature vector. This involves pre-training the spatiotemporal graph convolutional network constructed in step S2.2 based on the source industry feature vector, obtaining the final mapping model based on the pre-training results, and then combining this with the loss function of the target industry feature vector to obtain the final target industry feature vector. Specifically:

[0135] Step S3.1: Model Pre-training. This involves setting a pre-training loss function based on the source industry feature vectors, and using these source industry feature vectors as input to the spatiotemporal graph convolutional network constructed in Step S2.2. The output is the corresponding target industry feature vector. Notably, during pre-training, energy consumption and cleanliness in the semiconductor industry are used as comparative data for model pre-training, and the model is pre-trained according to the set pre-training loss function.

[0136] In this embodiment, the formula for obtaining the pre-training loss function is as follows:

[0137]

[0138] Where: L pre For the pre-training loss function, g K K is the energy consumption weighting coefficient, where K is the current energy consumption value. min K represents the lowest energy consumption in history. max This represents the highest energy consumption in history. U U is the cleanliness weighting coefficient, where U is the current cleanliness level. min For the minimum permissible cleanliness level, U max This represents the highest level of cleanliness in history.

[0139] Furthermore, during model pre-training, the input feature matrix of the target industry is combined with the model's convolution kernel parameter matrix to obtain the corresponding feature retention rate, specifically:

[0140]

[0141] Where: ω b To preserve the feature representation capability, W fix X is the frozen Chebyshev convolution kernel parameter matrix. new Input the feature matrix for the target industry.

[0142] Step S3.2: Determine the final model. This involves comparing the feature retention rate obtained in step S3.1 with a preset retention rate threshold, and determining the final mapping model based on the comparison result. Specifically:

[0143] When the obtained feature retention rate is greater than the preset retention rate threshold, the training model corresponding to that feature retention rate is the final mapping model. Otherwise, repeat steps S2.2-S3.2 until the obtained feature retention rate is greater than the preset retention rate threshold.

[0144] Step S3.3: Perform model prediction. That is, based on the final mapping model obtained in step S3.2, the source industry feature vector is used as the input of the final mapping model, and the corresponding target industry feature vector is output according to the training loss function set according to the target industry feature vector.

[0145] In this embodiment, the formula for obtaining the training loss function is as follows:

[0146] L fine =L pre +λ||ε act -ε tar || 2

[0147] Where: L fine Here, λ is the fine-tuning loss function, and L is the regularization coefficient. pre For the pre-training loss function, ε act To measure the turbulent dissipation rate, ε tar The target turbulent dissipation rate.

[0148] refer to Figure 3 , Figure 3 This is an analysis chart of the feature retention rate in this embodiment, by Figure 3 It can be seen that as the number of pre-training iterations increases, the feature retention rate gradually increases from the initial 70% to 98%, meaning that the spatiotemporal graph convolutional network can effectively learn and retain the key physical features of the source industry. Furthermore, at the 6th iteration, the feature retention rate exceeds the preset threshold, indicating that the model can be safely transferred to the target industry at this point. The curve also exhibits approximately linear growth, meaning that the model does not suffer from gradient vanishing or overfitting issues, maintaining high efficiency and stability during pre-training.

[0149] Example 2

[0150] This embodiment provides a method for synergistic optimization of energy consumption and environmental factors in cleanrooms based on big data mining. The specific implementation method is the same as in Embodiment 1, except that in step S1.2.2, during the industry relationship mapping process, the mapping relationship is verified through energy conservation and cleanliness equivalence. The invention will be illustrated below with specific examples of this embodiment.

[0151] In this embodiment, the mapping relationship is verified using energy conservation and cleanliness equivalence, and the final mapping result is determined based on the verification results. Specifically:

[0152] Step W1: Verification. This involves obtaining the kinetic energy density ratio based on the laminar and turbulent kinetic energy densities, thus determining the energy ratio for energy conservation verification. In other words, the obtained energy ratio is compared with a preset energy ratio threshold range, and the comparison result determines whether energy conservation has been achieved. Specifically:

[0153] If the obtained energy ratio is within the preset energy ratio threshold range, the obtained mapping result conforms to energy conservation. Otherwise, the obtained mapping result does not conform to energy conservation, and step W3 is executed.

[0154] In this embodiment, the formula for obtaining the energy ratio is as follows:

[0155]

[0156] in: E is the energy ratio. lam E is the laminar kinetic energy density. tur ρ is the turbulent kinetic energy density. s For air density, v t For the target turbulent wind speed, v l The source laminar flow velocity is TKE, and the percentage of turbulent kinetic energy is TKE.

[0157] Furthermore, based on the minimum requirements set by GMP specifications, a target mixing uniformity is set, and the measured mixing uniformity is obtained through spatial sampling calculations. Further, the ratio of the obtained target mixing uniformity to the measured mixing uniformity is calculated, i.e., the cleanliness ratio is determined, for cleanliness equivalence verification. In other words, the obtained cleanliness ratio is compared with a preset cleanliness ratio threshold range, and based on the comparison result, it is determined whether cleanliness equivalence has been achieved. Specifically:

[0158] When the obtained cleanliness ratio is within the preset cleanliness ratio threshold range, the obtained mapping result achieves the effect of cleanliness equivalence. Otherwise, the obtained mapping result does not achieve the effect of cleanliness equivalence, and steps W2 and W3 are executed.

[0159] In this embodiment, the formula for obtaining the cleanliness ratio is as follows:

[0160]

[0161] in: For cleanliness ratio, U tar For the target mixing uniformity, U act This represents the measured uniformity of mixing.

[0162] Step W2: Adjust the fan speed. That is, when it is determined in Step W1 that the mapping result does not achieve the equivalent cleanliness effect, obtain the current air supply volume baseline value, and determine the adjustment amount of the air supply volume based on the difference between the target mixing uniformity and the measured mixing uniformity. Specifically:

[0163] ΔQ=h×(U tar -U act )*Q0

[0164] Where: ΔQ is the adjustment amount of the air supply volume, h is the adjustment coefficient, and U tar For the target mixing uniformity, U act Q0 is the baseline value of the air supply volume for actual mixing uniformity measurement.

[0165] Furthermore, based on the determined adjustment amount of the air supply volume, the fan speed is adjusted in stages. Specifically, the fan speed is adjusted according to preset time intervals. It is worth noting that the specific number of adjustments to the fan speed can be set according to the actual needs of the scenario, so it will not be specifically described in this embodiment.

[0166] In this embodiment, the formula for adjusting the fan speed is as follows:

[0167]

[0168] Where: ω i+1 Let ω be the fan speed corresponding to the (i+1)th stage. i Let Q0 be the fan speed corresponding to the i-th stage, Q0 be the baseline value of the air supply volume, and ΔQ be the adjustment amount of the air supply volume.

[0169] Step W3: Adjust the basic parameter table. This involves combining the threshold ranges in the basic parameter table with the thresholds from the 10 most recent successful mappings to obtain a new threshold range. Specifically:

[0170]

[0171] Wherein: T new For the new threshold, T old The original threshold, This is the arithmetic mean of the most recent 10 successes. This is the historical average.

[0172] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for synergistic optimization of energy consumption and environmental factors in cleanrooms based on big data mining, characterized in that, Including: S1: Construct a physical feature knowledge base: Construct a physical feature knowledge base based on the source industry feature vector and the target industry feature vector. At the same time, obtain industry similarity based on the physical feature knowledge base, and improve the obtained target industry feature vector based on the industry similarity. S2: Construct a hybrid model: Based on the industry similarity, determine the improvement method for the target industry feature vector, and obtain the final source industry feature vector, including: S2.1: Determine the feature vector: Compare the industry feature similarity with a preset similarity threshold range, and determine the final target industry feature vector based on the comparison result, specifically as follows: When the industry feature similarity is greater than the upper limit of the preset similarity threshold range, the target industry feature vector is the final target industry feature vector. When the industry feature similarity is within the preset similarity threshold range, steps S2.2-S3 are executed to construct a spatiotemporal graph convolutional network and obtain the final target industry feature vector. When the industry feature similarity is less than the lower limit of the similarity threshold range, adjustments are made by experts. S2.2: Model Adjustment: By constructing the spatiotemporal graph convolutional network, the relative error corresponding to the initial target industry feature vector is obtained, and the gradient step size of the spatiotemporal graph convolutional network is corrected according to the relative error. S3: Obtain the final target industry feature vector: Based on the source industry feature vector, pre-train the spatiotemporal graph convolutional network after gradient step correction, determine the final mapping model based on the pre-training results, and obtain the final target industry feature vector through the final mapping model.

2. The method for synergistic optimization of energy consumption and environmental factors in cleanrooms based on big data mining as described in claim 1, characterized in that, To obtain industry similarity, including: S1.1: Feature extraction: Using hot-wire anemometers and laser trackers, Reynolds number characteristic data and mainstream kinetic energy ratio in the semiconductor industry are obtained; using PIV testing and data analysis, turbulence intensity spectrum and spatial particulate matter concentration distribution in the pharmaceutical industry are obtained. S1.2: Construct a knowledge base: Based on the Reynolds number characteristic data, the proportion of mainstream kinetic energy, the turbulence intensity spectrum, and the spatial particulate matter concentration distribution, construct a basic parameter table, and determine the mapping relationship between industries based on the basic parameter table; S1.3: Industry Similarity Assessment: Based on the mapping relationship between the industries, obtain the source industry feature vector and the corresponding target industry feature vector, determine the industry feature similarity, and improve the target industry feature vector based on the industry feature similarity. The formula for obtaining the industry feature similarity is as follows: Where: S represents industry feature similarity, Φ s Let Φ be the source industry feature vector. t This represents the feature vector of the target industry.

3. The method for synergistic optimization of energy consumption and environmental factors in cleanrooms based on big data mining as described in claim 2, characterized in that, The mapping relationships between industries were determined, including: S1.2.1: Establish a basic parameter table: Use the Reynolds number characteristic data, the proportion of mainstream kinetic energy, the turbulence intensity spectrum, and the spatial particulate matter concentration distribution as the threshold range in the basic parameter table. At the same time, construct the basic parameter table according to the determined industry and parameter type. S1.2.2: Perform industry relationship mapping: Based on the aforementioned basic parameter table, perform laminar-turbulent wind speed conversion and determine the mapping relationship between turbulent dissipation rate and Reynolds number, specifically as follows: Where: v t For the target turbulent wind speed, v l ε is the source laminar wind speed, D is the characteristic size, ε is the turbulent dissipation rate, and Re is the Reynolds number.

4. The method for synergistic optimization of energy consumption and environmental factors in cleanrooms based on big data mining as described in claim 2, characterized in that, The industry feature similarity is compared with a preset similarity, and the target industry feature vector is improved based on the comparison result, specifically as follows: If the industry feature similarity is less than the preset similarity, then the next step S2 is executed to improve the target industry feature vector; otherwise, the target industry feature vector is not improved.

5. A method for synergistic optimization of energy consumption and environmental factors in cleanrooms based on big data mining, as described in claim 2 or 3, characterized in that... The mapping relationship was verified using energy conservation and cleanliness equivalence, including: W1: Verification: Based on the laminar kinetic energy density and turbulent kinetic energy density, obtain the energy ratio and perform energy conservation verification. At the same time, based on the set target mixing uniformity and the measured mixing uniformity, obtain the cleanliness ratio and perform cleanliness conservation verification. W2: Adjusting the fan speed: Based on the difference between the target mixing uniformity and the measured mixing uniformity, determine the adjustment amount of the air supply volume. Simultaneously, based on the adjustment amount of the air supply volume, adjust the fan speed in stages. The formulas for adjusting the air supply volume and the fan speed are as follows: Where: ΔQ is the adjustment amount of the air supply volume, h is the adjustment coefficient, and U tar For the target mixing uniformity, U act To measure the mixing uniformity, Q0 is the baseline value of the air supply volume, ω i+1 Let ω be the fan speed corresponding to the (i+1)th stage. i The fan speed corresponding to the i-th stage; W3: Adjust the basic parameter table: Based on the threshold range in the basic parameter table and the threshold values ​​of the last 10 successfully mapped data, obtain a new threshold range, specifically: Wherein: T new For the new threshold, T old The original threshold, This is the arithmetic mean of the most recent 10 successes. This is the historical average.

6. The method for synergistic optimization of energy consumption and environmental factors in cleanrooms based on big data mining, as described in claim 5, is characterized in that... The energy ratio is compared with a preset energy ratio threshold range, and energy conservation verification is performed based on the comparison results. Specifically: When the energy ratio is within the preset energy ratio threshold range, the mapping result conforms to energy conservation; otherwise, the mapping result does not conform to energy conservation, and step W3 is executed. The formula for obtaining the energy ratio is as follows: in: E is the energy ratio. lam E is the laminar kinetic energy density. tur ρ is the turbulent kinetic energy density. s For air density, v t For the target turbulent wind speed, v l The source laminar flow velocity is TKE, and the percentage of turbulent kinetic energy is TKE.

7. The method for synergistic optimization of energy consumption and environmental factors in cleanrooms based on big data mining, as described in claim 5, is characterized in that... The cleanliness ratio is compared with a preset cleanliness ratio threshold range, and cleanliness conservation verification is performed based on the comparison results. Specifically: When the cleanliness ratio is within the preset cleanliness ratio threshold range, the mapping result conforms to the cleanliness conservation; otherwise, the mapping result does not conform to the cleanliness conservation, and steps W2 and W3 are executed. The formula for obtaining the cleanliness ratio is as follows: in: For cleanliness ratio, U tar For the target mixing uniformity, U act This represents the measured uniformity of mixing.

8. The method for synergistic optimization of energy consumption and environmental factors in cleanrooms based on big data mining as described in claim 1, characterized in that, The gradient stride of the spatiotemporal graph convolutional network is corrected, including: S2.2.1: Relaxation Factor Adjustment: Based on the total inflow corresponding to the source industry feature vector and the total outflow corresponding to the target industry feature vector, the relative error of the inflow and outflow mass flow is obtained, and the relaxation factor in the spatiotemporal graph convolutional network is adjusted accordingly, specifically as follows: Where: τ new τ is the adjusted relaxation factor. old Here, α is the relaxation factor before adjustment, b is the maximum allowable adjustment range, and E is the error sensitivity coefficient. mass This is relative error; S2.2.2: Learning Rate Adjustment: Based on the target industry feature vector, obtain the momentum residual and adjust the learning rate in the spatiotemporal graph convolutional network, specifically as follows: Where: η new Here, η0 is the adjusted learning rate, r is the industry sensitivity coefficient, and E is the learning rate before adjustment. mom For momentum residuals; S2.2.3: Constructing a dynamic graph: Using the coordinates of measurement points in the source industry feature vector as node coordinates, and based on the target industry feature vector, constructing a matrix between different nodes, specifically: Among them: A m，n Let be the matrix element between the m-th node and the n-th node, e be the base of the natural logarithm, and Cov(v) be the matrix element between the m-th node and the n-th node. m v n P represents the velocity covariance. m Let P be the spatial coordinates of the m-th node. n Let be the spatial coordinates of the nth node.

9. The method for synergistic optimization of energy consumption and environmental factors in cleanrooms based on big data mining as described in claim 1, characterized in that, The final target industry feature vector is obtained, including: S3.1: Model Pre-training: Based on the source industry feature vector, a pre-training loss function is set. Simultaneously, the source industry feature vector is used as the input to the spatiotemporal graph convolutional network after gradient step correction. The output obtains the corresponding target industry feature vector, and the corresponding feature retention rate is obtained. Specifically: Where: ω b To preserve the feature representation capability, W fix X is the frozen Chebyshev convolution kernel parameter matrix. new Input the feature matrix for the target industry; S3.2: Determine the final model: Compare the feature retention rate with a preset retention rate threshold, and determine the final mapping model based on the comparison result, specifically as follows: When the feature retention rate is greater than the preset retention rate threshold, the spatiotemporal graph convolutional network corresponding to the feature retention rate is the final mapping model; otherwise, repeat steps S2.2-S3.2 until the feature retention rate is greater than the preset retention rate threshold. S3.3: Perform model prediction: Based on the target industry feature vector, set the training loss function, and use the source industry feature vector as the input of the final mapping model to obtain the final target industry feature vector.

Images