Digital twinborn-based laboratory heating and ventilation system predictive analysis method and system

Abstract

The invention relates to the technical field of clean laboratory environments, and discloses a predictive analysis method and system of a laboratory heating and ventilation system based on digital twinning. The method comprises the following steps: acquiring a plurality of first target data and a first reference value in a first time period, and respectively calculating first trend data and first spectrum data through a first trend algorithm and a first conversion algorithm; acquiring a plurality of second target data and a second reference value in a second time period, and calculating second trend data and second spectrum data; then, the difference between the first trend data and the second trend data is calculated as first state information based on the digital twin platform, and the difference frequency band between the first spectrum data and the second spectrum data is calculated as second state information. And then, matching corresponding equipment state data according to the state information, calculating a similarity value of the two, and if the similarity value is smaller than a preset equipment reference similarity value, sending out an equipment early warning prompt. According to the invention, hardware linkage can be effectively promoted, accurate prediction of the equipment state is realized, and the equipment fault risk is reduced.

Description

Technical Field

[0001] The present application relates to the technical field of clean laboratory environments, and in particular to a predictive analysis method and system for laboratory HVAC systems based on digital twins. Background Art

[0002] A clean laboratory is a laboratory that, through specialized design, construction, and equipment configuration, controls indoor air pollutants such as dust particles and microorganisms to below a certain level to meet the environmental cleanliness requirements of specific experiments or production processes. A laboratory requires a variety of equipment to maintain a clean working environment. Therefore, the equipment must operate 24 / 7 and be monitored by intelligent hardware.

[0003] Intelligent laboratory hardware refers to hardware devices and systems that utilize sensors, the Internet of Things, and other technologies to intelligently monitor, control, and manage various laboratory equipment, instruments, and environments. Currently, sensor technology is relatively mature, capable of accurately measuring various physical, chemical, and biological quantities in the laboratory, such as temperature, humidity, air pressure, gas concentration, light intensity, and pH. Furthermore, sensors are becoming increasingly precise, stable, and reliable, with increasingly smaller sizes and lower power consumption, meeting the needs of diverse laboratory environments and application scenarios. These intelligent hardware devices enable automated data collection and real-time status monitoring, aiming to improve laboratory operational efficiency, safety, and reliability, reduce labor costs and experimental errors, and provide researchers with a more convenient and efficient experimental environment.

[0004] However, intelligent hardware in laboratories involves multiple hardware brands and subsystems, such as sensor systems, communication systems, and control systems, resulting in a high level of system complexity. During the integration process, data from each subsystem is often collected and aggregated in the backend for centralized display. However, the data between subsystems is relatively independent, making it impossible to achieve cross-hardware linkage and predict device status across subsystems. Summary of the Invention

[0005] In order to enable the hardware to be linked and the hardware of the subsystems to predict the equipment status, this application provides a predictive analysis method and system for the laboratory HVAC system based on digital twins.

[0006] In the first aspect, the present application provides a predictive analysis method for a laboratory HVAC system based on digital twins, which adopts the following technical solutions:

[0007] A predictive analysis method for laboratory HVAC systems based on digital twins includes the following steps:

[0008] Acquiring a plurality of first target data and a plurality of first reference values within a first time period;

[0009] Calculating first trend data using a first trend algorithm based on the plurality of first target data, and calculating first spectrum data using a first conversion algorithm based on the plurality of first reference values;

[0010] Acquiring a plurality of second target data and a plurality of second reference values within a second time period;

[0011] Calculating second trend data using a first trend algorithm based on the plurality of second target data; calculating second spectrum data using a first conversion algorithm based on the plurality of second reference values;

[0012] Calculate the difference between the first trend data and the second trend data based on the digital twin platform as first state information, and calculate the difference frequency band between the first spectrum data and the second spectrum data as second state information;

[0013] Match the first device status data based on the first status information, match the second device status data based on the second status information, calculate the similarity value between the first device status data and the second device status data, and issue a device warning prompt if the similarity value is less than a preset device reference similarity value.

[0014] By adopting these technical solutions, the digital twin intelligent management and control platform can monitor the health of equipment in real time by collecting key operating parameters such as temperature, vibration, and noise. This enables remote, real-time monitoring of cleanrooms, promptly detecting and warning of environmental anomalies and equipment failures, safeguarding the living environment of experimental animals and improving the accuracy and reliability of experimental data.

[0015] Optionally, based on a temperature sensor and a vibration sensor installed on the working equipment, the temperature sensor is electrically connected to a first edge computing unit, the vibration sensor is electrically connected to a second edge computing unit, and both the first edge computing unit and the second edge computing unit are remotely connected to the digital twin platform;

[0016] The method further comprises the steps of:

[0017] The first edge computing unit acquires a plurality of first target data and a plurality of second target data from the temperature sensor, and calculates the first trend data and the second trend data by using a first trend algorithm, wherein the first trend algorithm is used to calculate an average fluctuation value of the plurality of values;

[0018] The second edge computing unit obtains a plurality of first reference values and a plurality of second reference values from the vibration sensor, and calculates the first spectrum data and the second spectrum data by using a first conversion algorithm, wherein the first conversion algorithm is used to convert a plurality of values in the time domain into spectrum data in the frequency domain;

[0019] The digital twin platform obtains the first trend data, the second trend data, the first spectrum data and the second spectrum data, and calculates the difference or average difference between the first trend data and the second trend data as the first state information; calculates the energy difference between the corresponding frequency bands in the first spectrum data and the second spectrum data, and calculates the continuous frequency bands in which the energy difference is greater than the preset energy reference value as the second state information.

[0020] By adopting this technical solution, the edge computing unit is an AI edge server. This AI edge server analyzes data, promptly identifies abnormal conditions, and sends warning information to operators on the control platform. This allows for personnel and spare parts to be adjusted and prepared before actual equipment failures occur. This ensures maximum utilization of equipment lifespan and better manages personnel and spare parts inventory, achieving an optimal balance between practicality and cost-effectiveness. For laboratory cleanrooms, a digital twin model is established that incorporates environmental parameters and equipment operating status.

[0021] Optionally, the method for calculating the similarity value between the first device status data and the second device status data further includes the following sub-steps:

[0022] Introducing a first dynamic gain coefficient for the first device state data, the first dynamic gain coefficient being dynamically adjusted according to a nonlinear attenuation function K1(t)=e^(-αt) based on the duration t of the device being uncalibrated, where α is a preset attenuation coefficient;

[0023] A second dynamic gain coefficient is introduced for the second device status data. The second dynamic gain coefficient is dynamically adjusted according to the inverse proportional function K2(P)=β / (P+γ) based on the real-time monitored working device power P, where β and γ are power compensation parameters;

[0024] Based on the first dynamic gain coefficient and the second dynamic gain coefficient, a weighted matching index of the fluctuation amplitude of the device status data is calculated:

[0025] Match=Σ[K1(t)×ΔS1(i)-K2(P)×ΔS2(i)]^2;

[0026] Wherein ΔS1(i) and ΔS2(i) are the i-th fluctuation components of the first device status data and the second device status data respectively;

[0027] Similarity value = 1 / (1+Match).

[0028] By adopting the above technical solution, the larger the value of the similarity value, the higher the degree of similarity. Accurately match the device status data: By matching the fluctuation of the first device status data with the change amplitude of the second device data, it is possible to gain a more accurate insight into the relationship between the status of the devices. This feature helps to quickly and accurately judge the synergy of the operating status of different devices in complex equipment systems, thereby providing a solid data foundation for the overall optimized operation of the equipment. Gain value that adapts to dynamic changes: The setting of the first gain value becoming smaller as the uncalibrated time increases, and the second gain value decreasing as the power of the working equipment increases, enables this patented technology to dynamically adapt to different operating conditions of the equipment. When the equipment is operated for a long time without calibration, the change in the first gain value can effectively reduce the impact of errors accumulated over time on the accuracy of the data; and the second gain value is adjusted according to the equipment power, which can always maintain accurate measurement of the equipment status data when the equipment power fluctuates, greatly improving the reliability of equipment status monitoring under different working conditions.

[0029] Optionally, based on an image sensor and a vibration sensor installed on the working equipment, the image sensor is electrically connected to a first edge computing unit, the vibration sensor is electrically connected to a second edge computing unit, and both the first edge computing unit and the second edge computing unit are remotely connected to the digital twin platform;

[0030] The method further comprises the steps of:

[0031] The first edge computing unit acquires a plurality of first target data and a plurality of second target data from the image sensor, and calculates the first trend data and the second trend data by using a first trend algorithm, wherein the first trend algorithm is used to calculate a target state value corresponding to a preset template in the plurality of image data;

[0032] The second edge computing unit obtains a plurality of first reference values and a plurality of second reference values from the vibration sensor, and calculates the first spectrum data and the second spectrum data by using a first conversion algorithm, wherein the first conversion algorithm is used to convert a plurality of values in the time domain into spectrum data in the frequency domain;

[0033] The digital twin platform obtains the first trend data, the second trend data, the first spectrum data and the second spectrum data, and calculates the difference or average difference between the first trend data and the second trend data as the first state information; calculates the energy difference between the corresponding frequency bands in the first spectrum data and the second spectrum data, and calculates the continuous frequency bands in which the energy difference is greater than the preset energy reference value as the second state information.

[0034] By adopting the above technical solution, based on the image sensor and vibration sensor installed on the working equipment, the acquired data is processed by the first and second edge computing units electrically connected thereto, that is, the first edge computing unit calculates the target state value of the image data through a specific trend algorithm to obtain trend data, and the second edge computing unit converts the vibration sensor data from the time domain to the frequency domain to obtain spectrum data. By detecting and comparing the device status in the image, the differences between different parameters on the working equipment are compared. The digital twin platform calculates status information based on these data, which can realize effective analysis of the multi-dimensional data of the working equipment, timely discover abnormal trends and potential faults during the operation of the equipment, and realize remote real-time monitoring and early warning of the equipment, which helps to ensure the stable operation of the equipment, improve work efficiency and reliability, and make full use of edge computing to reduce data transmission pressure and improve the overall performance of the system.

[0035] Optionally, the method further comprises the following steps:

[0036] Collect the power parameters of the three-phase current, voltage and power factor of the working equipment in real time, and extract the time domain feature vector of the power parameters based on the sliding time window;

[0037] A dynamic threshold algorithm is used to perform anomaly detection on the time domain feature vector. The dynamic threshold algorithm calculates the anomaly index according to the following formula:

[0038] ;

[0039] Among them, X i (t) is the i-th power parameter at time t, μ i (t), σ i (t) are the mean and standard deviation of the sliding window in the first N minutes, ω i is the preset weight coefficient;

[0040] When the abnormal index exceeds a preset adaptive threshold, the frequency domain harmonic component of the corresponding power parameter is extracted and matched with the fault mode feature corresponding to the frequency domain harmonic component in the historical energy consumption database to generate an abnormal fluctuation spectrum;

[0041] The abnormal fluctuation spectrum is input into a pre-trained deep residual network model to output a prediction result, which includes a predicted value of equipment health and a predicted interval of remaining life. The prediction result is associated with a timestamp generated for the prediction result and then stored in a blockchain evidence database.

[0042] By employing these technical solutions, the energy monitoring system uses monitoring equipment to accurately monitor energy usage in every room and device, conducts statistical analysis of energy consumption anomalies, and identifies inappropriate energy allocation plans within the system. Comprehensive energy monitoring of all systems, through AI-powered learning and analysis, not only generates optimized energy utilization plans but also, based on historical energy consumption data, implements contingency plans for timely and seasonal energy consumption peaks.

[0043] Optionally, the method further comprises the following steps:

[0044] Synchronously collect heterogeneous data of physical entities through multi-source sensors, including stress and strain distribution data of mechanical structures, three-dimensional geometric point cloud data, and turbulence intensity spectrum data and multiphase flow pressure field data of fluid systems;

[0045] Construct a finite element model generator based on a graph neural network, input the stress and strain distribution data and the geometric point cloud data into the finite element model generator, and optimize the mesh division through iteration:

[0046] ;

[0047] Among them, K i is the element stiffness matrix, u i is the node displacement vector, f i is the external force vector acting on the node, TV(u) is the total variation regularization term, the above parameters are calculated through the stress-strain distribution data and the geometric point cloud data, and λ is the equilibrium coefficient;

[0048] A fluid dynamics model generator is constructed using a spatiotemporal convolutional network. The turbulence intensity spectrum data and the multiphase flow pressure field data are mapped to a fluid dynamics grid, and the residual of the Navier-Stokes equation is solved by a differentiable solver:

[0049] ;

[0050] Where v is the fluid velocity vector, t is the time, is the kinematic viscosity, p is the pressure; (v·∇)v is the convection term; is the viscous diffusion term; ∇p is the pressure gradient;

[0051] The finite element model provides output grid data, which has a matching data interface for coupling with the fluid dynamics grid. The finite element model is used to update the dynamic grid boundary conditions in the fluid dynamics grid for dynamic correction of R NS ;

[0052] Deploy an online model update engine to receive data streams from multiple source sensors for physical entity updates in real time, and dynamically adjust the weight parameters of the finite element model generator and the fluid dynamics model generator through an adaptive gradient descent algorithm:

[0053] ;

[0054] in, is the measured data at time t; M t is the model prediction value; JS is the Jensen-Shannon divergence, α and β are the multi-objective optimization weight coefficients; η is the learning rate; θ t+1 is the new weight parameter of the finite element model and the fluid dynamics model after the tth iteration update; θ t is the weight parameter at the current time t; ∇ θ is the gradient operator with respect to the weight parameter θ.

[0055] By adopting the above technical solutions, we study methods for constructing digital twin models for different complex systems, such as automatically generating accurate models based on machine learning and deep learning algorithms, including finite element models of mechanical structures and computational fluid dynamics models of fluid systems.

[0056] Optionally, the method further comprises the following steps:

[0057] Labeling data containing sensitive attributes in digital twin systems;

[0058] Use a preset asymmetric encryption algorithm to encrypt the marked sensitive data in real time;

[0059] Verify the identity of the access request through a biometric-based verification algorithm;

[0060] Based on the identity verification results, periodic incremental backups are performed on verified encrypted sensitive data.

[0061] By adopting the above technical solutions, we will study data security and privacy protection technologies in digital twin systems to prevent data leakage, tampering, and malicious attacks. For example, encryption algorithms and access control technologies will be used to protect data security. We will also build a security protection system for digital twin models to ensure the integrity and reliability of the models and prevent them from being illegally used or destroyed.

[0062] Optionally, the method further comprises the following steps:

[0063] Decrypt the target sensitive data using a decryption algorithm corresponding to the asymmetric encryption algorithm to generate a data integrity check value;

[0064] Inputting the data integrity check value and the characteristic vector of the trend data into a preset neural network verification model to verify the logical consistency of the digital twin model;

[0065] When the logical consistency verification fails, a multi-level warning mechanism is triggered: first, a device-level alarm is sent to the local terminal, and if there is no response within the preset time, a system-level alarm is sent to the central platform.

[0066] By employing the above technical solutions, security audits are conducted on user operations within the digital twin system and various events generated by the system, recording relevant operations and event information in a log. Subsequent analysis of the audit logs allows for timely detection of abnormal behavior and potential security vulnerabilities, and targeted measures are taken to repair and prevent them. For digital twin models, digital signatures, hash checksums, and other technologies are used to regularly perform integrity checks on the model. If the model is illegally modified or damaged, an alert is issued and recovery measures are taken, ensuring the integrity and reliability of the model and preventing its illegal exploitation.

[0067] In a second aspect, the present application provides a predictive analysis system for laboratory HVAC systems based on digital twins, which adopts the following technical solutions:

[0068] A predictive analysis system for a laboratory HVAC system based on a digital twin, comprising a processor, wherein the processor executes the steps of the predictive analysis method for a laboratory HVAC system based on a digital twin as described in any one of the above.

[0069] In a third aspect, the present application provides a storage medium that adopts the following technical solution:

[0070] A storage medium having a program stored therein, wherein the program, when executed by a processor, implements the steps of any one of the above-mentioned predictive analysis methods for a laboratory HVAC system based on a digital twin.

[0071] In summary, this application includes at least one of the following beneficial technical effects:. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 This is a step-by-step diagram of the predictive analysis method for laboratory HVAC systems based on digital twins.

[0073] Figure 2 This is a calculation step diagram based on the temperature sensor and vibration sensor installed on the working equipment, the first state information and the second state information. DETAILED DESCRIPTION

[0074] Embodiments of the present application are described in detail below, examples of which are illustrated in the accompanying drawings.

[0075] Throughout this specification, reference to the terms "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples" means that the specific features, structures, materials, or characteristics described in conjunction with the embodiments or examples are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0076] The present application embodiment discloses a predictive analysis method for a laboratory HVAC system based on digital twins, referring to Figure 1 , including the following steps:

[0077] In the first time period, such as the hour from 9 a.m. to 10 a.m., multiple first target data and multiple first reference values are obtained. The first target data comes from the values of temperature sensors evenly distributed in key parts of the HVAC system. These temperature sensors record data every 1 minute, and 60 first target data are obtained in this hour. The first reference value comes from the value of a vibration sensor or a sound sensor, which is used to measure sound data. For example, a vibration sensor is installed near the fan of the HVAC system. It collects vibration data at a frequency of 100 times per second. Similarly, a large amount of vibration data is obtained as the first reference value in this hour; a sound sensor is set near the air duct to collect sound data generated by the operation of the fan and the air flow through the air duct.

[0078] Based on these 60 first target data points, a first trend algorithm is used to calculate first trend data. This algorithm can be, for example, a moving average method. By processing this data, the current temperature change trend is determined, such as a temperature increase of 0.2°C every 10 minutes. For multiple first reference values, first spectrum data is calculated using a first conversion algorithm. This first conversion algorithm uses FFT (Fast Fourier Transform). Taking vibration data as an example, FFT analysis is performed on the collected vibration signal, converting the time domain signal into a frequency domain signal. This results in first spectrum data showing the distribution of vibration energy at different frequencies. A significant vibration energy peak may be observed at 500Hz.

[0079] In the second time period, assuming it is from 9 to 10 am on the next day, multiple second target data and multiple second reference values are acquired again. Similarly, the temperature sensor and the vibration and sound sensors collect data according to the previous collection frequency and method.

[0080] Based on the multiple second target data, the first trend algorithm calculates second trend data. For example, the temperature shows a trend of increasing by 0.3°C every 10 minutes. Based on the multiple second reference values, the first conversion algorithm calculates second spectrum data. If FFT analysis is performed again on the vibration data, it may be found that the vibration energy peak at 500Hz has increased, and a new energy peak has appeared at 1000Hz.

[0081] Based on the digital twin platform, the difference between the first and second trend data is calculated as the first state information. This difference clearly indicates the difference in temperature trends between the two time periods, such as an accelerated temperature rise rate. The difference in frequency band between the first and second spectrum data is also calculated as the second state information. This second state information reflects the difference in vibration frequency changes between the two time periods, such as the newly added 1000Hz frequency energy peak.

[0082] Based on the first state information, the first device state data is matched within the digital twin platform. For example, in the digital twin model, when the temperature rise rate exceeds a certain threshold, the corresponding device may have a heat dissipation problem, and data representing the degree of abnormality of this device is matched. Based on the second state information, the second device state data is matched. For example, if the energy of certain frequencies in the vibration spectrum increases abnormally, the corresponding device may have problems such as loose components. Data representing the degree of abnormality of the device is matched within the digital twin platform. Finally, the similarity value between the first device state data and the second device state data is calculated. If the similarity value is greater than the preset device reference similarity value, such as the preset similarity value is 0.7, when the calculated similarity value is 0.8, an equipment warning prompt is issued to remind staff that there may be hidden dangers of failure in the HVAC system equipment.

[0083] The digital twin intelligent management and control platform boasts powerful capabilities. It monitors the health of equipment in real time by collecting key operating parameters such as temperature, vibration, and noise. For example, in a cleanroom housing experimental animals, the platform monitors the temperature of HVAC equipment in real time. If the temperature exceeds the range suitable for the animals, it combines vibration and noise data to determine whether the equipment is functioning properly. If an equipment anomaly is detected, a timely warning can be issued to prevent adverse environmental conditions or equipment failures from impacting the animals' living environment. This ensures the accuracy and reliability of experimental data and prevents deviations in experimental results due to environmental factors.

[0084] Reference Figure 2 In the construction of the laboratory's working equipment monitoring system, an efficient and intelligent data collection and analysis architecture has been formed based on temperature sensors and vibration sensors installed in key parts of the working equipment. The temperature sensor is electrically connected to the first edge computing unit through a dedicated circuit, and the vibration sensor is electrically connected to the second edge computing unit in the same way. Both the first edge computing unit and the second edge computing unit establish a remote connection with the remote digital twin platform via a high-speed, stable wireless network. For example, on the air-conditioning unit equipment in a large clean laboratory, temperature sensors are installed in parts that are prone to heat, such as the compressor housing and motor windings, to accurately measure the real-time temperature of key parts of the equipment; vibration sensors are installed near parts that are prone to vibration, such as fan bearings and compressor pistons, to capture vibration conditions during equipment operation.

[0085] The method further includes the following detailed steps:

[0086] The first edge computing unit acquires multiple first target data and multiple second target data from the temperature sensor in real time. Assuming that the temperature sensor collects data once per minute during the first time period from 9:00 AM to 10:00 AM, the first edge computing unit will acquire 60 first target data points. During the second time period from 9:00 AM to 10:00 AM the next day, the first edge computing unit will also acquire 60 second target data points. The first edge computing unit uses the first trend algorithm to conduct an in-depth analysis of this data to calculate the first trend data and the second trend data. The first trend algorithm here is a tool specifically used to calculate the average fluctuation value of multiple values. The average fluctuation value can be calculated in two ways. One method is to calculate the average absolute value of the increase per unit time. For example, if the temperature rises from an initial 25°C to 28°C within the first hour, then the average absolute value of the increase per unit time (1 hour) is (28-25) ÷ 1 = 3°C / hour. Another method is to calculate the absolute value of the increase over N consecutive data points. Assuming N is 10 consecutive data points, with the temperature increasing from 25.1°C in the first data point to 25.6°C in the tenth data point, then the absolute value of the increase over these 10 data points is 25.6-25.1 = 0.5°C. Both methods can reflect the growth trend.

[0087] The second edge computing unit obtains multiple first reference values and multiple second reference values from the vibration sensor. During the first hour, the vibration sensor collected vibration data at a frequency of 100 times per second, resulting in a total of 360,000 first reference values. The same number of second reference values was obtained during the same time period on the second day. The second edge computing unit processes this data using a first conversion algorithm to calculate first and second spectrum data. The core function of the first conversion algorithm is to convert multiple values in the time domain into spectrum data in the frequency domain. Taking vibration data as an example, the first conversion algorithm (such as the Fast Fourier Transform (FFT)) can convert vibration amplitude change data, originally presented on a time axis, into spectrum data showing the distribution of vibration energy at different frequencies. This clearly identifies the frequency bands where the device experiences significant vibration energy, such as a significant vibration energy peak at 500Hz.

[0088] After acquiring the first trend data, second trend data, first spectrum data, and second spectrum data, the digital twin platform conducts in-depth analysis. It calculates the difference or average difference between the first and second trend data as first state information. For example, if the first trend data shows an average temperature increase of 2°C per hour and the second trend data shows an average temperature increase of 3°C per hour, the difference of 1°C or the average difference ((3-2) ÷ 2 = 0.5°C) between the two becomes first state information, which intuitively reflects the difference in temperature trends between the two time periods. The digital twin platform also calculates the energy difference between corresponding frequency bands in the first and second spectrum data and uses consecutive frequency bands where the energy difference exceeds a preset energy reference value as second state information. For example, if the preset energy reference value is 100, and the energy value of the 500Hz frequency band in the first spectrum data is 150 and the energy value of the same frequency band in the second spectrum data is 200, the energy difference is 50. If multiple consecutive frequency bands show similar energy difference changes greater than the preset value, these consecutive frequency bands constitute second state information, which reflects the changes in the vibration frequency characteristics of the device between the two time periods.

[0089] The edge computing unit here utilizes an advanced AI edge server with powerful built-in AI algorithms and data analysis models. The AI edge server enables fast and efficient local data analysis, eliminating the need to transmit all data to the cloud, significantly reducing data transmission latency. For example, when analyzing temperature and vibration data, the AI edge server can detect an abnormally accelerated temperature rise or the appearance of a new, abnormally high-energy frequency band in the vibration spectrum. Furthermore, it promptly sends an alert to operators on the control platform via a reliable communication link. For example, in a chip manufacturing laboratory cleanroom, if the AI edge server detects an abnormally high temperature rise or an abnormal frequency band in the vibration spectrum of a lithography machine, it sends an alert to the operator before actual equipment failure occurs, such as before lithography accuracy is significantly affected. This allows personnel to immediately coordinate and prepare personnel and spare parts, placing maintenance personnel on standby and preparing any parts that may need replacement. This approach not only ensures maximum equipment utilization over its lifespan, preventing premature equipment failure due to unexpected failures, but also allows for better control of personnel and spare parts inventory, preventing resource waste caused by excessive stockpiling, thus achieving an optimal balance between practicality and cost-effectiveness. Furthermore, for laboratory cleanrooms, the digital twin platform integrates environmental parameters (such as temperature, humidity, and air pressure) with equipment operating conditions (such as temperature, vibration, and speed) to create a highly realistic digital twin model. This model reflects the actual state of the laboratory cleanroom in real time, providing strong support for efficient laboratory operation and precise management.

[0090] The method for calculating the similarity value between the first device status data and the second device status data further includes the following sub-steps:

[0091] A first dynamic gain coefficient is introduced for the first device status data. The first dynamic gain coefficient is dynamically adjusted according to the nonlinear attenuation function K1(t)=e^(-αt) based on the duration t of the device being uncalibrated, where α is a preset attenuation coefficient. For example, in a certain laboratory, after a set of key reaction equipment has been running for a long time, the measurement accuracy of its temperature sensors, pressure sensors and other equipment may gradually decrease due to various factors, and this decrease will continue before recalibration. Assuming that α is set to 0.01, when the duration t of the device being uncalibrated reaches 100 hours, the first dynamic gain coefficient K1(100)=e^(-0.01×100)=e^(-1)≈0.368. This means that as the uncalibrated time continues to increase, the first dynamic gain coefficient will decrease exponentially, thereby effectively reducing the impact of data errors caused by long-term uncalibrated equipment on the overall device status judgment, making the subsequent analysis of the first device status data closer to the actual situation.

[0092] A second dynamic gain coefficient is introduced for the second device status data. The second dynamic gain coefficient is dynamically adjusted according to the inverse proportional function K2(P)=β / (P+γ) based on the real-time monitored working device power P, where β and γ are power compensation parameters. Taking a large-scale high-precision processing equipment in a laboratory as an example, the power of the equipment will change greatly under different working modes. When the equipment is performing fine processing, the power P may be low. Assuming β is 100 and γ is 50, if the power P is 100W at this time, then the second dynamic gain coefficient K2(100)=100 / (100+50)=2 / 3; when the equipment is performing rough processing, the power is greatly increased to 300W. At this time, the second dynamic gain coefficient K2(300)=100 / (300+50)=2 / 7. This second dynamic gain coefficient adjusted in real time according to the equipment power can always accurately measure the equipment status data when the equipment power fluctuates, avoid data measurement deviation caused by power changes, and provide a guarantee for accurate analysis of the equipment status.

[0093] Based on the first dynamic gain coefficient and the second dynamic gain coefficient, the weighted matching index of the fluctuation amplitude of the device status data is calculated:

[0094] Match=Σ[K1(t)×ΔS1(i)-K2(P)×ΔS2(i)]^2.

[0095] Where ΔS1(i) and ΔS2(i) are the i-th fluctuation components of the first device status data and the second device status data, respectively. Assume that at a certain moment, the fluctuation components of the first device status data ΔS1(1) = 0.5 and ΔS1(2) = 0.3, the fluctuation components of the second device status data ΔS2(1) = 0.4 and ΔS2(2) = 0.2, the first dynamic gain coefficient K1(t) = 0.5, and the second dynamic gain coefficient K2(P) = 0.6.

[0096] Then, Match = [0.5 × 0.5 - 0.6 × 0.4]^2 + [0.5 × 0.3 - 0.6 × 0.2]^2 = (0.25 - 0.24)^2 + (0.15 - 0.12)^2 = 0.001. This formula incorporates two dynamic gain coefficients, which take into account the effects of the device's uncalibrated duration and real-time power, into the calculation of the device status data fluctuation amplitude. This allows the matching index to more comprehensively and accurately reflect the relationship between device status data.

[0097] Finally, the similarity value = 1 / (1+Match). Based on the Match value calculated above, the similarity value at this time = 1 / (1+0.001) = 0.999. The larger the similarity value, the higher the degree of similarity. This method of accurately matching device status data can more accurately understand the relationship between the status of devices by matching the fluctuation of the first device status data with the change amplitude of the second device data. For example, in an automated production experimental system containing multiple devices, this method can quickly and accurately determine the synergy of the operating status of different devices. When the temperature change of one device is highly similar to the vibration change of another device, it may mean that there is some potential connection between them, such as because they share the same cooling system or mutual influence on the mechanical structure. This feature provides a solid data foundation for the overall optimization operation of the equipment. Engineers can use these similarity values to determine whether the equipment is operating normally and whether the operating parameters of the equipment need to be adjusted.

[0098] The first gain value decreases as the uncalibrated time increases, and the second gain value decreases as the power of the working equipment increases, allowing this patented technology to dynamically adapt to the different operating conditions of the equipment. When the equipment has been running for a long time without calibration, the change in the first gain value can effectively reduce the impact of errors accumulated over time on data accuracy. Just like the chemical production laboratory equipment mentioned earlier, as the uncalibrated time increases, the first dynamic gain coefficient decreases, avoiding the interference of inaccurate measurement data on the judgment of the equipment status; the second gain value is adjusted according to the equipment power, which can always maintain an accurate measurement of the equipment status data when the equipment power fluctuates. For example, the processing equipment in the electronic manufacturing laboratory can accurately reflect the equipment status through the second dynamic gain coefficient regardless of how the power changes, greatly improving the reliability of equipment status monitoring under different working conditions.

[0099] In other implementations, a data acquisition and analysis system based on image and vibration sensors is used in HVAC laboratories to efficiently monitor and accurately manage operating equipment. These sensors are installed on key operating equipment, such as air conditioning compressors and fans, to capture real-time operating status information.

[0100] The image sensor and vibration sensor are electrically connected to the first and second edge computing units, respectively. This electrical connection ensures that data collected by the sensors can be quickly and stably transmitted to the edge computing units for preliminary processing. The first and second edge computing units are remotely connected to the digital twin platform via a high-speed network, and the digital twin platform performs in-depth analysis and processing of the data from the edge computing units.

[0101] The first edge computing unit handles the crucial task of processing image sensor data. In HVAC labs, image sensors can be installed on the exterior of air conditioning units, inside air ducts, on heat exchanger surfaces, and elsewhere to capture the equipment's appearance, component operating status, and airflow.

[0102] The first edge computing unit obtains multiple first target data and multiple second target data from the image sensor. For example, within a specific time period, the image sensor captures images of the device at certain time intervals. The first target data may be a series of images captured between 9:00 AM and 10:00 AM, while the second target data may be images captured between 2:00 PM and 3:00 PM.

[0103] The first edge computing unit processes this image data using the first trend algorithm. The core of this algorithm is to calculate the target state value corresponding to a preset template from multiple image data. The preset template is constructed based on the image features of the device under normal operating conditions. Taking the heat exchanger of an air conditioning unit as an example, the preset template may include features such as the neat arrangement of the heat exchanger fins and the lack of dirt accumulation on the surface. The first trend algorithm analyzes each image, extracts features related to the preset template, such as the tilt angle of the fins and the area covered by dirt, and converts these features into target state values.

[0104] By processing the first target data and the second target data separately, the first edge computing unit generates first trend data and second trend data. For example, the first trend data shows that the average tilt angle of the heat exchanger fins is 5 degrees in the morning, while the second trend data shows that the average tilt angle changes to 8 degrees in the afternoon. This indicates that the condition of the heat exchanger fins may have changed, possibly becoming loose or deformed.

[0105] The second edge computing unit is dedicated to processing data from vibration sensors. Vibration sensors are typically installed in key parts of equipment, such as compressor casings and fan bearings, to monitor vibration during operation.

[0106] The second edge computing unit obtains multiple first reference values and multiple second reference values from the vibration sensor. These reference values are the amplitudes of vibration signals collected by the vibration sensor at different time periods. For example, the first reference value is the vibration amplitude collected within the first 30 minutes after the device is started, while the second reference value is collected within 30 minutes after the device has been running continuously for 2 hours.

[0107] The second edge computing unit uses the first conversion algorithm to convert multiple values in the time domain into frequency domain spectrum data. The time domain data reflects the changes in the vibration signal over time, while the frequency domain data shows the distribution of the vibration signal at different frequencies. A commonly used conversion algorithm is the Fast Fourier Transform (FFT). Using the FFT algorithm, the second edge computing unit converts the time domain vibration signal collected by the vibration sensor into frequency domain spectrum data, generating first and second spectrum data.

[0108] For example, the first spectrum data shows that during the equipment startup phase, the vibration signal is mainly concentrated at two frequencies, 50Hz and 100Hz. However, the second spectrum data shows that after the equipment has been running continuously for two hours, a significant vibration peak appears at a frequency of 200Hz. This may indicate that a component of the equipment has failed during operation, causing the vibration frequency to change.

[0109] The digital twin platform receives the first and second trend data from the first edge computing unit, as well as the first and second spectrum data from the second edge computing unit. By comprehensively analyzing this data, the digital twin platform can gain a deeper understanding of the device's operating status.

[0110] The digital twin platform first calculates the difference or average difference between the first and second trend data as primary status information. For example, in the case of the trend data for the tilt angle of a heat exchanger's fins, the calculated difference is 3 degrees, which becomes the primary status information. This information intuitively reflects the state changes of the equipment over time, helping operators determine potential equipment issues.

[0111] The digital twin platform also calculates the energy difference between the corresponding frequency bands in the first spectrum data and the second spectrum data. For example, for a frequency band with a frequency of 50Hz, the energy value in the first spectrum data is 100 joules, the energy value in the second spectrum data is 150 joules, and the energy difference is 50 joules. When the energy difference is greater than the preset energy reference value, the digital twin platform will further calculate the continuous frequency bands in which these energy differences are greater than the reference value as the second state information. Assuming that the preset energy reference value is 30 joules, and the energy difference in the 50Hz-100Hz frequency band is greater than 30 joules, then the continuous frequency band of 50Hz-100Hz constitutes the second state information.

[0112] By using primary and secondary status information, the digital twin platform can comprehensively and accurately assess the operating status of equipment. If the primary status information indicates a significant change in a certain equipment parameter, and the secondary status information indicates an abnormal frequency band in the vibration spectrum, then the equipment may be faulty and require prompt inspection and maintenance. In this way, the data acquisition and analysis system based on image sensors and vibration sensors, combined with the intelligent processing capabilities of the digital twin platform, provides strong support for HVAC laboratory equipment management, ensuring stable and efficient equipment operation.

[0113] The method further comprises the steps of:

[0114] During equipment operation, the energy monitoring system collects real-time power parameters such as the three-phase current, voltage, and power factor of the operating equipment. These parameters reflect the equipment's power usage. For example, in a large HVAC laboratory, a continuously running motor has three-phase currents I1, I2, and I3, a voltage U, and a power factor cosφ. The system processes this data using a sliding time window, which acts like a moving observation frame, continuously capturing data within a certain time range. This method extracts time-domain feature vectors corresponding to the power parameters. For example, if the sliding time window is set to 10 minutes, the window slides forward by one minute every minute, continuously collecting power parameter data for these 10 minutes. Representative time-domain feature vectors are then extracted from this data. These feature vectors contain information about the temporal changes in the power parameters.

[0115] The dynamic threshold algorithm is used to detect anomalies on the extracted time domain feature vector. The dynamic threshold algorithm is based on the formula Calculate abnormal indicators. Taking multiple devices in a large HVAC laboratory as an example, this could be the current parameter of the i-th device at time t (or other power parameters such as voltage and power factor).

[0116] X i (t) is the i-th power parameter at time t, μ i (t), σ i (t) are the mean and standard deviation of the sliding window in the first N minutes (assuming N = 15 minutes), ω i is a preset weighting factor. For example, a higher weighting factor is assigned to the current parameter of a critical device to highlight its importance in anomaly detection. When the calculated anomaly indicator A(t) exceeds the preset adaptive threshold, it indicates that the device's power parameters have experienced abnormal fluctuations.

[0117] When the abnormality index exceeds the preset adaptive threshold, the system will further extract the frequency domain harmonic components of the corresponding power parameters. Taking the frequency conversion equipment in a large HVAC laboratory as an example, its frequency domain harmonic components are within a certain range during normal operation. Once the equipment fails, such as aging of the internal components of the inverter, the frequency domain harmonic components will change. The system matches the extracted frequency domain harmonic components with the fault mode characteristics corresponding to the frequency domain harmonic components in the historical energy consumption database. The historical energy consumption database stores the power parameter characteristics of various devices under different fault conditions. Through matching, an abnormal fluctuation spectrum is generated. For example, if the harmonic components of the equipment match the fault mode characteristics of the motor winding short circuit in the database, a corresponding abnormal fluctuation spectrum will be generated, which intuitively shows the specific situation of the equipment abnormality.

[0118] The generated abnormal fluctuation spectrum is input into a pre-trained deep residual network model. The deep residual network model is an intelligent model trained on a large amount of data and capable of analyzing abnormal fluctuation spectra. The model outputs predictions, including a predicted value for device health and a predicted range for remaining life. For example, after model analysis, the predicted health value of a particular device is 70%, indicating that the device is currently in a relatively healthy state but has certain potential risks. The predicted remaining life is also given as 3-6 months, prompting operations personnel to pay close attention to the device or arrange for repairs within this timeframe. The prediction results are then linked to the timestamp of the generated prediction results and stored in a blockchain evidence database. The blockchain evidence database is tamper-proof, ensuring the authenticity and reliability of the data and facilitating subsequent query and traceability.

[0119] The energy monitoring system uses monitoring equipment to precisely monitor energy usage in each area and device. For example, in a HVAC laboratory, the system can monitor the power consumption of air conditioning, lighting, and other equipment in each HVAC area. Statistical analysis is performed on abnormal energy consumption to identify inappropriate aspects of the system's energy allocation plan. For example, if the air conditioning energy consumption in a certain HVAC area is found to be abnormally high during off-peak hours, analysis may reveal that the air conditioning temperature control system is improperly configured. Comprehensive energy monitoring of all systems, through AI-powered learning and analysis, not only generates optimized energy utilization plans but also, based on historical energy consumption data, implements contingency plans for timely and seasonal energy consumption peaks. For example, based on energy consumption data from previous summers, it can predict peak electricity consumption on certain days this summer. Energy allocation strategies can be adjusted in advance, and backup power supplies can be appropriately deployed to ensure efficient energy utilization and stable equipment operation.

[0120] The method further comprises the steps of:

[0121] In the HVAC laboratory, multi-source sensors work together to collect heterogeneous data of physical entities. Heterogeneous data includes stress and strain distribution data of mechanical structures, three-dimensional geometric point cloud data, and turbulence intensity spectrum data and multiphase flow pressure field data of fluid systems. For example, for the laboratory's air-conditioning units, in terms of mechanical structure, stress and strain sensors are installed on the key components of the compressor to simultaneously collect stress and strain distribution data. When the compressor is running at high speed, these sensors can accurately capture the stress changes on the components, helping engineers understand whether the components have fatigue damage due to long-term high-load operation. 3D laser scanning equipment is used to obtain 3D geometric point cloud data of the air-conditioning unit, accurately depicting the unit's external contours, pipeline layout, and internal space structure, providing an accurate geometric information basis for subsequent airflow simulation and structural optimization.

[0122] In terms of fluid systems, multiple turbulence intensity sensors and spectrum analysis equipment are placed in the air duct to collect turbulence intensity spectrum data. This data can provide a detailed reflection of the turbulent characteristics of the airflow in the duct, such as the degree of airflow turbulence, the generation and development of vortices, etc., which is crucial for optimizing air duct design and reducing airflow noise and energy loss. Pressure sensors are installed near components such as heat exchangers and humidifiers to collect multiphase flow pressure field data and monitor the pressure changes of multiphase fluids such as air and water during heat exchange and humidification processes, ensuring that the system operates stably under appropriate pressure conditions and safeguarding the heat and moisture treatment effect of the HVAC system.

[0123] Taking the frame structure of a large modular air conditioning unit in the HVAC laboratory as an example, a finite element model generator based on a graph neural network was constructed. Stress and strain distribution data and geometric point cloud data were input into the finite element model generator. The stress and strain distribution data records the stress and strain generated by factors such as vibration and pressure during the operation of the unit. The geometric point cloud data accurately represents the shape, dimensions, and connection relationships of the frame components.

[0124] The finite element model is obtained by iteratively optimizing the mesh division, and the calculation is performed according to the following formula:

[0125] .

[0126] Among them, K i is the element stiffness matrix, u i is the node displacement vector, TV(u) is the total variation regularization term, and λ is the equilibrium coefficient. i Determines the ability of each unit in the frame structure to resist deformation, just like the strength of each beam in a building. iThis describes the displacement of the frame nodes under load, such as the sway amplitude of the nodes under fan vibration excitation. The external force vector fi represents the external force acting on the nodes, such as the vibration force generated by fan operation. The total variation regularization term TV(u) ensures the stability and smoothness of the model and prevents unreasonable sudden changes. The balance coefficient λ adjusts the weights of each term in the optimization process, allowing the model to more accurately simulate the mechanical behavior of the frame in actual operation, predict the structural stability of the frame, avoid deformation and fracture caused by long-term stress, and ensure the safe and reliable operation of the air conditioning unit.

[0127] In the finite element model generator based on graph neural networks, the correspondence between various parameters and stress and strain distribution data and geometric point cloud data is as follows:

[0128] Element stiffness matrix K i :Geometric point cloud data provides the geometric shape and spatial distribution of the structure, which is used to determine the geometric parameters of the unit (such as node coordinates, unit size), which directly affects K i The construction of K is done by calculating the stiffness matrix based on the material properties and geometric shape. The stress-strain distribution data reflects the mechanical properties of the material (such as elastic modulus and Poisson's ratio), and is combined with the geometric data to determine K. i The numerical value of .

[0129] Node displacement vector u i :Geometric point cloud data is divided into grids to generate node positions, u i is the displacement calculation result of these nodes. The stress and strain distribution data indirectly affects u i , because the mechanical properties of the material (reflected by stress and strain data) determine the stiffness matrix K i , and then through equation K i u i =f i Solve for the displacement.

[0130] External force vector f i :The stress and strain distribution data directly provide the external force information (such as mechanical load, thermal load) acting on the node, which is used to define f i The numerical value of .

[0131] Total variation regularization term TV(u): Geometric point cloud data is used to construct a regularization term based on node adjacency (defined by the geometry) to constrain the smoothness of the displacement field and avoid mesh distortion. Stress and strain distribution data implicitly contain physical constraints (such as material continuity). The regularization term optimizes the displacement field by balancing residual errors with physical plausibility.

[0132] The balance coefficient λ is a hyperparameter determined by experiments or optimization algorithms and is used to adjust the residual term ||K i u i -f iThe weight of ||2 and the regularization term TV(u) ensures that the model achieves a balance between geometric accuracy (relying on point cloud data) and physical rationality (relying on stress and strain data).

[0133] Geometric point cloud data dominates K i The geometric part and the spatial constraints of TV(u), the support node displacement u i Stress and strain distribution data provide material properties and external force information, which directly affects K i The material part and f i and indirectly constrain u through the mechanical equilibrium equation i The parameter λ acts as a global coordination factor to ensure that the physical and geometric properties of the two types of data work together in the optimization process.

[0134] For the HVAC laboratory's air handling system, a spatiotemporal convolutional network was used to construct a fluid dynamics model generator. Taking the simulation of airflow within an air duct as an example, turbulence intensity spectrum data and multiphase flow pressure field data were mapped to CFD, a fluid dynamics grid. Turbulence intensity spectrum data reflects the turbulent characteristics of the airflow within the duct, such as vortex frequency and intensity distribution, which affect air delivery efficiency and mixing. Pressure field data, on the other hand, displays the pressure at different locations within the duct, playing a key role in determining the flow direction and velocity of the airflow.

[0135] The Navier-Stokes equation residuals are solved using a differentiable solver using the formula:

[0136] .

[0137] The fluid velocity vector v describes the speed and direction of the airflow; time t is used to analyze the dynamic changes of the airflow over time; the kinematic viscosity Reflects the viscous properties of air; pressure p determines the pressure-driven state of the airflow. (v·∇)v is the convection term; is the viscous diffusion term; ∇p is the pressure gradient. Solving the residual of this equation allows for more accurate simulation of air flow, heat transfer, and mass transfer within the duct. This helps engineers optimize the shape, size, and layout of the duct, reduce airflow resistance, improve the energy efficiency of the air handling system, and ensure that laboratory temperature, humidity, and air quality remain stable and meet standards.

[0138] Cross-scale coupling of finite element models and CFD is based on structure-fluid interaction:

[0139] The finite element model calculates the deformation or vibration of the mechanical structure and generates the displacement field u i ;

[0140] The displacement field is input into CFD as a dynamic boundary condition to update the flow field grid, affecting the velocity v and pressure p of the fluid.

[0141] Data transfer interface: Use interpolation algorithms, such as radial basis functions, to map the displacement data of the finite element model nodes to the corresponding positions of each CFD mesh; through coupled solvers, such as Co-Simulation, realize two-way real-time data exchange between FEM and CFD.

[0142] Real-time data assimilation: Deploy an online update engine to receive multi-source sensor data (such as real-time pressure, flow rate), and adjust CFD model parameters (such as , boundary conditions); combined with FEM structural state feedback (such as material aging), the flow-consolidation boundary conditions are dynamically modified.

[0143] Multi-objective optimization: The objective function considers both the fluid residual ||RNs|| and the structural residual ||K i u i -f i ||, and balance the convergence of the two through the weight coefficients α and β.

[0144] The integration of finite element models and CFD is essentially a synergy of multi-physics coupling and data-driven development. Finite element models provide structural deformation data, which serves as dynamic boundary conditions for CFD. Turbulence intensity spectra and multiphase flow pressure field data are embedded in the CFD solution through feature extraction and machine learning. An online update mechanism ensures real-time consistency between the model and the physical entity. This enables a full-dimensional digital twin of the HVAC system, from structural health to fluid performance, supporting fault early warning and energy efficiency optimization.

[0145] In more detail, the meshing data provided by the finite element model provides a consistent data interface for coupling with CFD. Therefore, the finite element model can update the CFD dynamic mesh boundary conditions, dynamically correct Rns, and realize fluid-structure coupling simulation.

[0146] In fluid-structure interaction simulations, updating the CFD dynamic mesh boundary conditions and dynamically correcting the Navier-Stokes residuals (RNS) in conjunction with the finite element model can be achieved by following these steps:

[0147] Adopt explicit coupling: solve the finite element model and CFD independently in steps, and achieve interaction by alternating data transfer.

[0148] Displacement interpolation: The node displacement u calculated by the finite element model is i Mapping to CFD boundary mesh nodes via radial basis functions (RBF) or nearest neighbor interpolation ensures displacement continuity at the structure-fluid interface, avoiding geometry mismatches.

[0149] Dynamic Mesh: The fluid domain is treated as an elastic medium and the internal mesh is adjusted according to the boundary displacement.

[0150] Adaptive time step: Dynamically adjust the time step according to the rate of change of residuals, such as reducing the step size when the residual increases.

[0151] Relaxation Iteration: Introducing Sub-Relaxation Factor Control the variable update speed and improve the convergence stability. The formula is: ;

[0152] Data-driven correction: The flow field data measured by the sensor, such as pressure and velocity, are used as constraints, and the residual terms are corrected through adjoint optimization to make the simulation results close to the real data.

[0153] Coupling solution: Based on the established finite element model and CFD, define the interface, such as the pipe wall. The finite element model solves the structural displacement u i Output the interface displacement. Update the CFD dynamic mesh boundary conditions and transfer the displacement to the fluid domain. CFD solves the updated flow field, which includes velocity v and pressure p, and calculates the fluid load ff. Apply the fluid load ff as an external force to the finite element model, and proceed to the next time step.

[0154] Convergence judgment: Monitor the structural residual and fluid residual simultaneously until both are below the set threshold.

[0155] For example, consider the vibration analysis of fan blades in HVAC systems: a finite element model calculates the blade's vibration displacement under airflow. The dynamic mesh updates the flow field shape around the blade in CFD. The CFD simulates the updated airflow pressure distribution and feeds it back into the finite element model as a dynamic load. This process is iterated until the vibration and flow field reach a steady state, allowing the blade fatigue life to be predicted. By combining the finite element model with CFD, dynamically updating mesh boundary conditions, and correcting RNS residuals, fluid-structure interaction simulation can accurately simulate the interaction between structural deformation and fluid response. Data-driven and machine learning enhance the model's dynamic adaptability, providing a reliable tool for fault warning and design optimization in areas such as HVAC systems.

[0156] In the HVAC laboratory, an online model update engine is deployed to monitor the operating status of equipment such as air conditioning units in real time. For example, an air conditioning unit's compressor receives real-time data updates from various sensors on the compressor, including data such as the compressor casing temperature measured by a temperature sensor, vibration amplitude and frequency detected by a vibration sensor, and operating current feedback from a current sensor.

[0157] The weight parameters of the finite element model and the fluid dynamics model are dynamically adjusted by the adaptive gradient descent algorithm according to the formula:

[0158] ;

[0159] in, is the measured data at time t, such as the actual temperature value; Mt is the model prediction value; Jensen-Shannon divergence (JS) measures the difference in the distribution of measured data and model prediction data, judging the accuracy of the model prediction. The multi-objective optimization weight coefficients α and β adjust the optimization direction of the model under different objectives, such as assigning different weights when focusing on temperature changes and vibration anomalies. η is the learning rate, which controls the step size of the model parameter update to ensure that the model can quickly and accurately adapt to changes in the compressor operating state. t+1 is the new weight parameter of the finite element model and the fluid dynamics model after the tth iteration update; θ t is the weight parameter at the current time t; ∇ θ is a gradient operator with respect to the weight parameter θ. By continuously adjusting the model weight parameters, the model can promptly reflect the actual operating conditions of the compressor, detecting potential faults such as component wear and poor lubrication in advance. This provides a scientific basis for equipment maintenance and management, ensuring the stable operation of HVAC laboratory equipment.

[0160] The method further comprises the steps of:

[0161] Data containing sensitive attributes in the digital twin system should be labeled. A large amount of data of various types exists in the digital twin system of an HVAC laboratory. Data containing sensitive attributes warrants special attention. For example, data collected by high-precision temperature and humidity sensors in the laboratory reflects the precise temperature and humidity changes in the experimental environment. This data is crucial for experiments that are extremely sensitive to environmental conditions (such as biological sample cultivation and high-precision chemical reaction experiments), and therefore falls into the category of sensitive data. Another example is the operating parameters of HVAC equipment during certain special experiments, such as the start and stop times of the compressor and the frequency adjustment range. Leakage of this information could affect the normal progress of the experiment or even cause it to fail, and thus also has sensitive attributes.

[0162] A specially designed tagging algorithm screens the data in the system on a category-by-category basis. When the algorithm detects that a certain type of data matches a pre-defined sensitive attribute, it automatically tags it. For example, if a set of data is identified as real-time temperature and humidity monitoring values within a specific experimental area, the "Sensitive - Temperature and Humidity Experimental Data" tag is immediately added to the data's metadata to facilitate subsequent targeted processing.

[0163] Use a preset asymmetric encryption algorithm to encrypt the marked sensitive data in real time. Use a preset asymmetric encryption algorithm, such as the RSA algorithm, to encrypt the marked sensitive data in real time. In the digital twin system of the HVAC laboratory, data generation and transmission are continuous. When temperature and humidity data with sensitive tags are transmitted from the sensor to the system server, the system automatically calls the encryption module. Using an asymmetric encryption algorithm, a corresponding public key and private key pair is generated for each set of sensitive data. The public key is used to encrypt the data, and the encrypted ciphertext is transmitted within the system's internal network. For example, a set of sensitive data representing the current precise temperature value in the laboratory is encrypted using the public key before transmission, turning it into a string of garbled ciphertext. Only authorized devices or users with the corresponding private key can decrypt the ciphertext and restore it to the actual temperature value, effectively preventing the data from being stolen and illegally viewed during transmission and storage.

[0164] Access requests are verified using a biometric-based authentication algorithm. In HVAC laboratory scenarios, personnel entering the laboratory to operate the digital twin system may include researchers, operations and maintenance engineers, and others. Taking fingerprint recognition as an example, each authorized personnel enters their fingerprint into the system upon initial access and registration. When researchers wish to access sensitive laboratory environment data stored in the digital twin system, they must capture their fingerprint using a specific authentication device (such as a fingerprint-enabled access control system or computer login device). The system compares the captured fingerprint with pre-stored authorized fingerprints and calculates the similarity using a complex biometric recognition algorithm. Only when the similarity reaches a preset high threshold (e.g., above 95%) is the identity verified and access to the relevant sensitive data is permitted. Furthermore, biometric recognition technologies such as facial recognition and iris recognition can also be applied to this identity verification process to further enhance the accuracy and security of authentication.

[0165] Based on the identity verification results, periodic incremental backups are performed on verified encrypted sensitive data. Based on the identity verification results, periodic incremental backups are performed on verified encrypted sensitive data. In HVAC laboratories, the amount of encrypted sensitive data can be very large, such as equipment operating parameters and experimental environment monitoring data accumulated over a long period of time. To ensure data security and recoverability, the system performs incremental backups at preset time periods (such as 2 a.m. every day). Incremental backups only back up data that has changed since the last backup, rather than backing up all data repeatedly. For example, on a given day, only the operating parameters of some HVAC equipment change. The system intelligently identifies these changed data and backs them up along with the associated encryption keys to a dedicated data storage device (such as a secure off-site server). This not only saves storage resources but also enables quick and accurate restoration to the most recent correct state in the event of data loss or corruption, ensuring the normal operation of the digital twin system and data integrity.

[0166] In the digital twin system of the HVAC laboratory, data security and privacy protection are crucial research topics. In addition to the encryption algorithms mentioned above, access control technology is also an important means of protecting data security. For example, different user roles are set up in the system, such as super administrators, ordinary researchers, operation and maintenance personnel, etc., and different permissions are assigned to each role. The super administrator has the highest authority and can operate all data and functions of the system; ordinary researchers can only access some temperature and humidity, equipment operating status and other data related to their own experiments; operation and maintenance personnel are mainly responsible for monitoring equipment operating parameters and adjusting some configurations, and can only access data related to equipment operation and maintenance. Through this strict access control strategy, the access scope of different personnel to data is limited to prevent unauthorized access and tampering of data.

[0167] At the same time, a security protection system for the digital twin model is constructed to ensure the integrity and reliability of the model from multiple levels. At the network level, firewalls, intrusion detection systems, and other equipment are used to monitor and filter network traffic entering the digital twin system in real time to prevent external malicious network attacks, such as denial of service attacks (DoS) and port scanning. At the system level, the digital twin system's software is regularly scanned and repaired for vulnerabilities, and security patches are updated to prevent hackers from exploiting software vulnerabilities to illegally use or damage the model. At the physical level, hardware equipment such as servers is ensured to be placed in a secure computer room environment, and access control, monitoring, and other measures are implemented to prevent physical equipment from being illegally accessed or damaged. By comprehensively applying these technologies and measures, a comprehensive security protection barrier for the HVAC laboratory's digital twin system is established, providing solid data security guarantees for the laboratory's scientific research work and the stable operation of equipment.

[0168] The method further comprises the steps of:

[0169] In the HVAC laboratory's digital twin system, data integrity and model logical consistency must be guaranteed. When processing or analyzing sensitive target data, the decryption algorithm corresponding to the asymmetric encryption algorithm must first be used to decrypt and verify it.

[0170] For example, consider the temperature and humidity data from a key laboratory area. This data is encrypted and protected using an asymmetric encryption algorithm during transmission and storage. When researchers need to view the real-time temperature and humidity data for this area, the system invokes the corresponding decryption algorithm to decrypt the encrypted data. During the decryption process, the system also generates a data integrity check value. For example, a hash algorithm (such as SHA-256) is used to calculate the decrypted data, generating a unique hash value as the data integrity check value. This hash value can be used to verify whether the data has been tampered with during transmission and storage.

[0171] Next, the data integrity check value and the eigenvectors of the trend data are input into a pre-set neural network verification model to verify the logical consistency of the digital twin model. The eigenvectors of the trend data are extracted from historical data and reflect the trends and patterns of data changes over time. For example, in a HVAC laboratory, temperature and humidity data typically exhibit certain periodic and seasonal variations. By analyzing historical temperature and humidity data, eigenvectors representing these variations can be extracted. The neural network verification model learns the relationship between these eigenvectors and the data integrity check value to determine whether the digital twin model accurately reflects the actual physical system. If the temperature and humidity trends in the digital twin model significantly differ from the actual collected data, or if the data integrity check value differs from the expected result, then there may be a problem with the logical consistency of the digital twin model.

[0172] When logical consistency verification fails, the system triggers a multi-level warning mechanism to promptly identify and resolve the issue. For example, if the energy consumption predicted by the digital twin model differs significantly from the actual energy consumption data collected for an air conditioning unit in the laboratory, and the neural network verification model confirms that the logical consistency does not hold, the system will take immediate action.

[0173] The system will send a device-level alarm to the local terminal. The local terminal can be a monitoring screen at the laboratory site, a handheld device of the operation and maintenance personnel, etc. The alarm information will detail the device with the problem (such as a specific air-conditioning unit), the type of problem (such as abnormal energy consumption), and the possible impact. After receiving the device-level alarm, the operation and maintenance personnel need to respond within a preset time (such as 15 minutes). If there is no response from the operation and maintenance personnel within the preset time, the system will consider the problem to be serious and require a higher level of intervention. At this time, a system-level alarm will be sent to the central platform. The central platform is usually monitored by laboratory managers or technical experts. Based on the system-level alarm information, they can organize professionals to conduct in-depth investigation and processing to ensure that the problem is resolved in a timely manner to avoid greater impact on the normal operation of the laboratory.

[0174] Auditing user operations within the digital twin system and all system-generated events is a crucial measure for ensuring system security. In the HVAC laboratory, user operations include controlling equipment, querying and modifying data, and system events include starting and stopping equipment, as well as fault alarms. The system records all these operations and events in a detailed audit log.

[0175] For example, when researchers modify the temperature and humidity settings in a certain experimental area, the system records the time of the operation, the operator's identity, the settings before and after the modification, and other information. Similarly, when a device issues a fault alarm, the system records the time of the fault, the type of fault, and the relevant parameters of the device. By analyzing these audit logs, abnormal behavior and potential security vulnerabilities can be detected in a timely manner. For example, if a user is found to frequently operate a device during non-working hours, or if a device frequently issues fault alarms within a short period of time, it may indicate an abnormal situation. In response to these abnormal situations, system administrators can further investigate the cause and take targeted measures to repair and prevent them, such as strengthening user authority management and performing maintenance and inspections on equipment.

[0176] Digital twin models are regularly verified for integrity using technologies such as digital signatures and hash checks. Digital signatures ensure the model's provenance and integrity. For example, when creating a digital twin model, a digital signature is generated using an encryption algorithm and stored with the model. During the model integrity check, the system verifies the validity of the digital signature. If the digital signature is invalid, it indicates that the model may have been illegally modified.

[0177] Hash verification verifies the integrity of the model by calculating its hash value. For example, the code and data of the digital twin model are hashed regularly, and the calculated hash value is compared with the previously saved hash value. If the two hash values are inconsistent, it means that the model may have been modified. Once it is discovered that the model has been illegally modified or damaged, the system will promptly issue an alarm and take recovery measures. For example, the original version of the model can be restored from a backup, or the model can be retrained and optimized to ensure the integrity and reliability of the model, prevent the model from being illegally used, and ensure the normal operation of the HVAC laboratory digital twin system and the accuracy of the experimental data.

[0178] An embodiment of the present application also discloses a predictive analysis system for a laboratory HVAC system based on a digital twin, including a processor that executes the steps of the predictive analysis method for a laboratory HVAC system based on a digital twin as described in any one of the above.

[0179] An embodiment of the present application also discloses a storage medium, in which a program is stored. When the program is executed by a processor, the steps of the predictive analysis method of the laboratory HVAC system based on digital twin are implemented.

[0180] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A predictive analysis method for laboratory HVAC systems based on digital twins, characterized in that: The steps include: Acquiring a plurality of first target data and a plurality of first reference values within a first time period; Calculating first trend data using a first trend algorithm based on the plurality of first target data, and calculating first spectrum data using a first conversion algorithm based on the plurality of first reference values; Acquiring a plurality of second target data and a plurality of second reference values within a second time period; Calculating second trend data using a first trend algorithm based on the plurality of second target data; Calculating second spectrum data using a first conversion algorithm according to the plurality of second reference values; Calculate the difference between the first trend data and the second trend data based on the digital twin platform as first state information, and calculate the difference frequency band between the first spectrum data and the second spectrum data as second state information; Match the first device status data based on the first status information, match the second device status data based on the second status information, calculate the similarity value between the first device status data and the second device status data, and issue a device warning prompt if the similarity value is less than a preset device reference similarity value.

2. The predictive analysis method for laboratory HVAC systems based on digital twins according to claim 1 is characterized in that: Based on a temperature sensor and a vibration sensor installed on the working equipment, the temperature sensor is electrically connected to a first edge computing unit, the vibration sensor is electrically connected to a second edge computing unit, and the first edge computing unit and the second edge computing unit are both remotely connected to the digital twin platform; The method further comprises the steps of: The first edge computing unit acquires a plurality of first target data and a plurality of second target data from the temperature sensor, and calculates the first trend data and the second trend data by using a first trend algorithm, wherein the first trend algorithm is used to calculate an average fluctuation value of the plurality of values; The second edge computing unit obtains a plurality of first reference values and a plurality of second reference values from the vibration sensor, and calculates the first spectrum data and the second spectrum data by using a first conversion algorithm, wherein the first conversion algorithm is used to convert a plurality of values in the time domain into spectrum data in the frequency domain; The digital twin platform obtains the first trend data, the second trend data, the first spectrum data, and the second spectrum data, and calculates a difference or an average difference between the first trend data and the second trend data as first state information; The energy difference between the first spectrum data and the energy difference between the corresponding frequency bands in the second spectrum data is calculated, and the continuous frequency bands in which the energy difference is greater than a preset energy reference value are calculated as the second state information.

3. The predictive analysis method for laboratory HVAC systems based on digital twins according to claim 1, characterized in that: The method for calculating the similarity value between the first device status data and the second device status data further includes the following sub-steps: Introducing a first dynamic gain coefficient for the first device state data, the first dynamic gain coefficient being dynamically adjusted according to a nonlinear attenuation function K1(t)=e^(-αt) based on the duration t of the device being uncalibrated, where α is a preset attenuation coefficient; A second dynamic gain coefficient is introduced for the second device status data. The second dynamic gain coefficient is dynamically adjusted according to the inverse proportional function K2(P)=β / (P+γ) based on the real-time monitored working device power P, where β and γ are power compensation parameters; Based on the first dynamic gain coefficient and the second dynamic gain coefficient, a weighted matching index of the fluctuation amplitude of the device status data is calculated: Match=Σ[K1(t)×ΔS1(i)-K2(P)×ΔS2(i)]^2; Wherein ΔS1(i) and ΔS2(i) are the i-th fluctuation components of the first device status data and the second device status data respectively; Similarity value = 1 / (1+Match).

4. The predictive analysis method for laboratory HVAC systems based on digital twins according to claim 1, characterized in that: Based on an image sensor and a vibration sensor installed on the working equipment, the image sensor is electrically connected to a first edge computing unit, the vibration sensor is electrically connected to a second edge computing unit, and the first edge computing unit and the second edge computing unit are both remotely connected to the digital twin platform; The method further comprises the steps of: The first edge computing unit acquires a plurality of first target data and a plurality of second target data from the image sensor, and calculates the first trend data and the second trend data by using a first trend algorithm, wherein the first trend algorithm is used to calculate a target state value corresponding to a preset template in the plurality of image data; The second edge computing unit obtains a plurality of first reference values and a plurality of second reference values from the vibration sensor, and calculates the first spectrum data and the second spectrum data by using a first conversion algorithm, wherein the first conversion algorithm is used to convert a plurality of values in the time domain into spectrum data in the frequency domain; The digital twin platform obtains the first trend data, the second trend data, the first spectrum data, and the second spectrum data, and calculates a difference or an average difference between the first trend data and the second trend data as first state information; The energy difference between the first spectrum data and the energy difference between the corresponding frequency bands in the second spectrum data is calculated, and the continuous frequency bands in which the energy difference is greater than a preset energy reference value are calculated as the second state information.

5. The predictive analysis method for laboratory HVAC systems based on digital twins according to claim 1, characterized in that: The method further comprises the steps of: Collect the power parameters of the three-phase current, voltage and power factor of the working equipment in real time, and extract the time domain feature vector of the power parameters based on the sliding time window; A dynamic threshold algorithm is used to perform anomaly detection on the time domain feature vector. The dynamic threshold algorithm calculates the anomaly index according to the following formula: ； Among them, X i (t) is the i-th power parameter at time t, μ i (t), σ i (t) are the mean and standard deviation of the sliding window in the first N minutes, ω i is the preset weight coefficient; When the abnormal index exceeds a preset adaptive threshold, the frequency domain harmonic component of the corresponding power parameter is extracted and matched with the fault mode feature corresponding to the frequency domain harmonic component in the historical energy consumption database to generate an abnormal fluctuation spectrum; The abnormal fluctuation spectrum is input into a pre-trained deep residual network model to output a prediction result, which includes a predicted value of equipment health and a predicted interval of remaining life. The prediction result is associated with a timestamp generated for the prediction result and then stored in a blockchain evidence database.

6. The predictive analysis method for laboratory HVAC systems based on digital twins according to claim 1, characterized in that: The method further comprises the steps of: Synchronously collect heterogeneous data of physical entities through multi-source sensors, including stress and strain distribution data of mechanical structures, three-dimensional geometric point cloud data, and turbulence intensity spectrum data and multiphase flow pressure field data of fluid systems; Construct a finite element model generator based on graph neural network, input the stress and strain distribution data and the geometric point cloud data into the finite element model generator, and obtain the finite element model by iteratively optimizing the mesh division: ； Among them, K i is the element stiffness matrix, u i is the node displacement vector, f i is the external force vector acting on the node, TV(u) is the total variation regularization term, the above parameters are calculated through the stress-strain distribution data and the geometric point cloud data, and λ is the equilibrium coefficient; A fluid dynamics model generator is constructed using a spatiotemporal convolutional network. The turbulence intensity spectrum data and the multiphase flow pressure field data are mapped to a fluid dynamics grid, and the residual of the Navier-Stokes equation is solved by a differentiable solver: ； Where v is the fluid velocity vector, t is the time, is the kinematic viscosity, p is the pressure; (v·∇)v is the convection term; is the viscous diffusion term; ∇p is the pressure gradient; The finite element model provides output grid data, which has a matching data interface for coupling with the fluid dynamics grid. The finite element model is used to update the dynamic grid boundary conditions in the fluid dynamics grid for dynamic correction of R NS ; Deploy an online model update engine to receive data streams from multiple source sensors for physical entity updates in real time, and dynamically adjust the weight parameters of the finite element model generator and the fluid dynamics model generator through an adaptive gradient descent algorithm: ； in, is the measured data at time t; M t is the model prediction value; JS is the Jensen-Shannon divergence, α and β are the multi-objective optimization weight coefficients; η is the learning rate; θ t+1 is the new weight parameter in the finite element model and fluid dynamics model after the tth iteration update; θ t is the weight parameter at the current time t; ∇ θ is the gradient operator with respect to the weight parameter θ.

7. The predictive analysis method for laboratory HVAC systems based on digital twins according to claim 1, characterized in that: The method further comprises the steps of: Labeling data containing sensitive attributes in digital twin systems; Use a preset asymmetric encryption algorithm to encrypt the marked sensitive data in real time; Verify the identity of the access request through a biometric-based verification algorithm; Based on the identity verification results, periodic incremental backups are performed on verified encrypted sensitive data.

8. The predictive analysis method for laboratory HVAC systems based on digital twins according to claim 7, characterized in that: The method further comprises the steps of: Decrypt the target sensitive data using a decryption algorithm corresponding to the asymmetric encryption algorithm to generate a data integrity check value; Inputting the data integrity check value and the characteristic vector of the trend data into a preset neural network verification model to verify the logical consistency of the digital twin model; When the logical consistency verification fails, a multi-level warning mechanism is triggered: first, a device-level alarm is sent to the local terminal, and if there is no response within the preset time, a system-level alarm is sent to the central platform.

9. A predictive analysis system for laboratory HVAC systems based on digital twins, characterized in that: It includes a processor that executes the steps of the predictive analysis method for a laboratory HVAC system based on a digital twin as described in any one of claims 1 to 8.

10. A storage medium, characterized in that: The medium stores a program, which, when executed by a processor, implements the steps of the predictive analysis method for a laboratory HVAC system based on a digital twin as described in any one of claims 1 to 8.

Images