Intelligent laboratory air quality control system and method

Abstract

The invention relates to the technical field of heating, ventilation and air conditioning control, in particular to an intelligent laboratory air quality control system and method.The intelligent laboratory air quality control system comprises a data acquisition module, an intelligent decision module, a fault diagnosis module and an execution control module; meanwhile, feedback type multi-target collaborative operation parameters are generated through a plurality of heterogeneous strategy models; the fault diagnosis module analyzes equipment health data through an AI health degree baseline model, and outputs diagnosis information as a dynamic constraint; and the execution control module fuses the compensation control sequence, the cooperative parameters and the diagnosis information, generates a final physical control instruction verified by a safety boundary rule, and drives a heating ventilation air conditioner execution mechanism. According to the invention, through a control strategy integrating prediction, optimization and diagnosis, the stability of the laboratory environment, the operation economy and the long-term operation reliability of the system are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of heating, ventilation and air conditioning control technology, specifically to an intelligent laboratory air quality control system and method. Background Technology

[0002] The primary design objective of HVAC systems in laboratory environments is to ensure personnel safety and environmental stability. To achieve this goal, conventional laboratory ventilation strategies require the use of 100% outdoor fresh air and the maintenance of a high air exchange rate (ACH) to effectively dilute and remove potential air pollutants.

[0003] One fundamental technical solution for achieving the above objectives is the constant air volume (CAV) system. CAV systems ensure a safety margin by continuously supplying a fixed, typically worst-case-condition-based, high airflow to the laboratory. The control logic of this system is relatively simple, and its operation is reliable. Its energy consumption is directly related to this constant high airflow; variable air volume (VAV) systems, developed on this basis, aim to improve energy efficiency. VAV systems can link ventilation volume to some direct physical state parameters (e.g., the opening degree of the fume hood's control window). Their control systems typically employ feedback control logic (such as a PID controller), meaning that when a sensor measures an environmental parameter (such as temperature) deviating from the setpoint, the system adjusts the actuators (such as valves) accordingly to correct the deviation.

[0004] The aforementioned existing technologies provide effective environmental protection solutions for conventional laboratories; however, in some advanced scientific research fields with extreme requirements for environmental stability, such as high-precision physics laboratories in optics, metrology, and quantum information, in addition to the conventional temperature, humidity and safety requirements, several more stringent and interrelated control objectives have emerged. For example, it is necessary to maintain stable temperature, ISO 5 level air cleanliness and VC-E level micro-vibration standards at the same time.

[0005] In these scenarios, the daily operation of the laboratory (e.g., the start-up and shutdown of specific scientific research equipment, the filling of cryogenic media, etc.) itself becomes a predictable source of disturbance that will have a combined impact on the above-mentioned multiple high-precision environmental parameters. Therefore, there is a technical problem to be solved in this field: how to actively intervene and compensate for disturbances before they have a substantial impact on the environment in order to achieve environmental stability.

[0006] Therefore, an intelligent laboratory air quality control system and method are proposed. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent laboratory air quality control system and method. Through a control strategy that integrates prediction, optimization, and diagnosis, it significantly improves the stability, operational economy, and long-term reliability of the laboratory environment. The system includes a data acquisition module, an intelligent decision-making module, a fault diagnosis module, and an execution control module. The intelligent decision-making module generates a feedforward compensation control sequence based on disturbance data and simultaneously generates feedback-based multi-objective collaborative operating parameters through multiple heterogeneous strategy models. The fault diagnosis module analyzes equipment health data using an AI health baseline model and outputs diagnostic information as dynamic constraints. The execution control module integrates the aforementioned compensation control sequence, collaborative parameters, and diagnostic information to generate a final physical control command verified by safety boundary rules, driving the HVAC actuators.

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

[0009] A smart laboratory air quality control system includes:

[0010] The data acquisition module collects state data from the physics laboratory, including three-dimensional temperature field data, suspended particle data, and structural vibration data; acquires disturbance source event data; and acquires vibration and pressure signals from the HVAC system.

[0011] The intelligent decision-making module inputs state data and disturbance source event data into the control prediction model for prediction, generating a compensation control sequence; it constructs an input vector from the state data; it simultaneously inputs the input vector into multiple heterogeneous strategy models for forward inference, integrates the parameter values ​​output by the multiple heterogeneous strategy models, and generates collaborative operation parameters.

[0012] The fault diagnosis module processes vibration and pressure signals to extract vibration spectrum features and pressure fluctuation pattern data; it analyzes the vibration spectrum features and pressure fluctuation pattern data through an AI health baseline model and outputs diagnostic information.

[0013] The execution control module integrates the compensation control sequence, the cooperative operation parameters, and the diagnostic information to generate physical control commands, and verifies the physical control commands through safety boundary rules to drive the actuators in the HVAC system.

[0014] Preferably, the data acquisition module acquires data by: deploying a temperature sensor array in a grid pattern along a three-dimensional spatial coordinate system within the laboratory to obtain three-dimensional temperature field data; deploying laser particle counters in the experimental equipment area, air supply vents, and return air vents to acquire suspended particle data; installing accelerometers on the building's load-bearing structure, instrument platform, and HVAC unit base to acquire structural vibration data; identifying and recording disturbance source event data such as door opening and closing, personnel movement, and equipment start-up and shutdown; and installing vibration and pressure sensors in the HVAC system to acquire vibration and pressure signals.

[0015] Preferably, the control prediction model includes: fusing the disturbance source event data and state data, and converting them into an initial state vector through an encoder network; inputting the initial state vector into a time series prediction model based on a Transformer architecture, and performing iterative time series extrapolation within the model at dynamically determined time steps to generate specific future environmental state change curves; determining a target compensation curve opposite to the target compensation based on the future environmental state change curves, and decomposing the target compensation curve into discrete target state points; for each discrete target state point, using the target state point and the current environmental state as input, querying a pre-trained offline neural network model for directly mapping state changes to control commands, and directly outputting the specific control command values ​​required to achieve the target state point using a non-iterative forward computation method; and combining all the solved specific control command values ​​in chronological order to generate the compensation control sequence.

[0016] Preferably, the process of generating collaborative operating parameters includes: constructing a state input vector from the three-dimensional temperature field data, suspended particle data, and structural vibration data; the multiple heterogeneous strategy models include models with three different optimization objectives: an energy consumption optimization model, a temperature field optimization model, and a particle sedimentation model; the multiple heterogeneous strategy models are independently trained based on different data subsets generated from all historical data through a bootstrap sampling method; the state input vector is simultaneously input into the multiple heterogeneous strategy models for parallel forward inference, and during the forward inference, parameter suggestion values ​​and a quantitative uncertainty index used to characterize the model's confidence in the current suggestion value are output; the multiple parameter suggestion values ​​output by the multiple heterogeneous strategy models are integrated through a dynamic weighted fusion strategy based on the inverse of uncertainty, and the fusion weight corresponding to each model is calculated according to the quantitative uncertainty index output by each model. The fusion weight of any model is inversely proportional to the quantitative uncertainty index output by the model, and the multiple parameter suggestion values ​​are weighted and summed using the fusion weight to generate collaborative operating parameters.

[0017] Preferably, the process of processing the vibration signal and pressure signal to extract vibration spectrum features and pressure fluctuation pattern data includes: performing a fast Fourier transform on the vibration signal to obtain the vibration spectrum; extracting the fundamental frequency and amplitude of each harmonic related to the rotational speed of the rotating component in the HVAC system, as well as the energy value at the preset bearing fault characteristic frequency, from the vibration spectrum; filling the sample into a fixed-length feature vector according to a preset dimensional order, as vibration spectrum features; calculating the root mean square value, kurtosis, and skewness of the pressure signal within a sliding time window and using them as statistical features; performing a continuous wavelet transform to extract the time-frequency energy distribution of the pressure transient event; and concatenating the data obtained after vectorizing the statistical features and the time-frequency energy distribution to form pressure fluctuation pattern data characterizing the pressure fluctuation pattern.

[0018] Preferably, the AI ​​health baseline model includes:

[0019] The input embedding layer transforms the extracted vibration spectrum features and pressure fluctuation pattern data into high-dimensional feature vectors.

[0020] The Transformer encoding layer receives the high-dimensional feature vector and models the intrinsic correlation between each feature dimension through an internal multi-head self-attention mechanism, and outputs a health prediction vector for the current state.

[0021] The anomaly detection layer generates an anomaly score by calculating the deviation between the health prediction vector and the actual input feature vector. When the anomaly score exceeds a dynamic threshold, the system is determined to be abnormal.

[0022] The root cause diagnosis layer is activated when an anomaly is detected. By analyzing the attention weights and model gradient information within the Transformer encoding layer, it calculates the contribution score of each input feature to the prediction bias, identifies the key feature with the highest contribution score, and matches the key feature with a preset physical fault feature signature library to output diagnostic information containing the specific physical root cause location.

[0023] Preferably, the process of driving the actuators of the HVAC system includes: fusing the compensation control sequence, the collaborative operation mode parameters, and diagnostic information to form a control intent, wherein the compensation control sequence serves as a feedforward adjustment benchmark, the collaborative operation mode parameters serve as a feedback optimization target, and the diagnostic information serves as a dynamic constraint condition; when the diagnostic information indicates that a component has a health risk, the fusion strategy is dynamically adjusted to reduce the operating load of the component; by querying the control instruction mapping table, the control intent is parsed into a specific physical control instruction sequence for each actuator in the HVAC system; according to preset equipment physical limits, process safety requirements, and energy consumption limit rules, the amplitude, rate of change, and combination logic of the physical control instruction sequence are checked item by item for safety; for instructions that exceed the safety boundary, the instruction parameters that exceed the single value limit are adjusted to the safety boundary value, and the logical safety of the entire instruction combination is rechecked; if the check passes, the corrected instruction is issued; if the check fails, the issuance of the instruction is intercepted and an alarm is triggered.

[0024] A smart laboratory air quality control method, comprising:

[0025] The system collects state data from a physics laboratory, including three-dimensional temperature field data, suspended particle data, and structural vibration data; acquires disturbance source event data; acquires vibration and pressure signals from the HVAC system; inputs the state data and disturbance source event data into a control prediction model for prediction, generating a compensation control sequence; constructs an input vector from the state data; simultaneously inputs the input vector into multiple heterogeneous strategy models for forward inference, integrates the parameter values ​​output by the multiple heterogeneous strategy models, and generates collaborative operation parameters; processes the vibration and pressure signals to extract vibration spectrum features and pressure fluctuation pattern data; analyzes the vibration spectrum features and pressure fluctuation pattern data using an AI health baseline model, outputting diagnostic information; fuses the compensation control sequence, the collaborative operation pattern vector, and the diagnostic information to generate physical control commands, and verifies the physical control commands using safety boundary rules to drive the actuators within the HVAC system.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] 1. By introducing a feedforward, model-predictive compensation control mechanism, this invention can proactively intervene before foreseeable disturbances actually occur, thereby significantly suppressing fluctuations in environmental parameters and providing higher environmental stability compared to traditional feedback control.

[0028] 2. By adopting an adaptive collaborative optimization mechanism based on "expert integration" and "uncertainty perception", this invention can make dynamic and reliable trade-offs among multiple conflicting control objectives. The uncertainty-based fusion strategy reduces the risk of misjudgment caused by the defects of a single model, enabling the system to make more robust and reasonable control decisions under complex and ever-changing working conditions.

[0029] 3. By introducing AI-based, deep-level physical fault diagnosis capabilities, this invention can identify early performance degradation or minor anomalies in equipment that are difficult for conventional systems to detect online. This predictive maintenance capability helps to address faults before they escalate into major problems, thereby improving the long-term operational reliability of the system and providing clear guidance for maintenance work due to its accurate root cause localization. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of an intelligent laboratory air quality control system provided in an embodiment of the present invention;

[0031] Figure 2 This is a schematic diagram of the process for generating a compensation control sequence according to an embodiment of the present invention;

[0032] Figure 3 This is a schematic diagram of the AI ​​health baseline model diagnostic process provided in an embodiment of the present invention;

[0033] Figure 4 This is a schematic flowchart of an intelligent laboratory air quality control method provided in an embodiment of the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] Please see Figures 1 to 4 This invention provides an intelligent laboratory air quality control system and method, the technical solution of which is as follows:

[0036] Example 1:

[0037] A smart laboratory air quality control system, the specific structure of which is as follows: Figure 1 As shown, it includes:

[0038] The data acquisition module collects state data from the physics laboratory, including three-dimensional temperature field data, suspended particle data, and structural vibration data; acquires disturbance source event data; and acquires vibration and pressure signals from the HVAC system.

[0039] The intelligent decision-making module inputs state data and disturbance source event data into the control prediction model for prediction, generating a compensation control sequence; it constructs an input vector from the state data; it simultaneously inputs the input vector into multiple heterogeneous strategy models for forward inference, integrates the parameter values ​​output by the multiple heterogeneous strategy models, and generates collaborative operation parameters.

[0040] The fault diagnosis module processes vibration and pressure signals to extract vibration spectrum features and pressure fluctuation pattern data; it analyzes the vibration spectrum features and pressure fluctuation pattern data through an AI health baseline model and outputs diagnostic information.

[0041] The execution control module integrates the compensation control sequence, the cooperative operation parameters, and the diagnostic information to generate physical control commands, and verifies the physical control commands through safety boundary rules to drive the actuators in the HVAC system.

[0042] Furthermore, the data acquisition module's data acquisition process includes: deploying a temperature sensor array in a grid pattern along a three-dimensional spatial coordinate system within the laboratory to acquire three-dimensional temperature field data; deploying laser particle counters in the experimental equipment area, air supply vents, and return air vents to acquire suspended particle data; installing accelerometers on the building's load-bearing structure, instrument platform, and HVAC unit base to acquire structural vibration data; identifying and recording disturbance source event data such as door opening and closing, personnel movement, and equipment start-up and shutdown; and installing vibration and pressure sensors in the HVAC system to acquire vibration and pressure signals.

[0043] Specifically, the system's data acquisition process is achieved through a multimodal sensing module, configured to construct a comprehensive dynamic dataset of the laboratory. To achieve precise sensing of the internal temperature field of the laboratory, this module deploys an array of multiple temperature sensors in a high-density grid along a three-dimensional spatial coordinate system within the laboratory. To monitor and ensure the air cleanliness of the laboratory in real time, this embodiment deploys multi-channel laser particle counters at key locations such as the experimental equipment area, air supply vents, and return air vents. These counters can simultaneously monitor the particle concentration of multiple key particle size channels to ensure that the air cleanliness in key areas is always maintained above the preset standard. To assess in real time the structural vibrations caused by the external environment and the operation of the HVAC system itself, this embodiment installs high-sensitivity triaxial accelerometers on the building's key load-bearing structures and precision instrument platforms. Their measurement resolution is sufficient to capture micro-vibrations caused by the external environment and internal equipment, to assess in real time whether the vibration state of the laboratory floor meets the vibration standard requirements of precision instruments. To collect non-periodic disturbance events that can be used for prediction, this embodiment identifies and records them through various methods. Door opening and closing events are collected by status sensors installed at all entrances and exits of the laboratory; personnel activity events are collected using space occupancy sensors to obtain real-time data on the number and spatial distribution of personnel while protecting privacy; equipment start-up and shutdown events are accurately identified by the power monitoring unit based on the on / off state and changes in current; this embodiment performs in-depth data collection on the HVAC system itself, where vibration signals are captured by installing vibration sensors at the bearing housings of key rotating components such as the main blower and circulating water pump. Pressure signals are captured by installing dynamic pressure sensors at key nodes of the main pipeline to capture pressure transients caused by airflow pulsation or valve operation.

[0044] By constructing a high-fidelity dynamic panorama of the laboratory through multi-dimensional and multi-scale data, it can not only perceive the subtle changes in the environment such as temperature field, cleanliness and micro-vibration in real time, but also provide a complete and reliable decision-making basis for subsequent advanced intelligent algorithms such as predictive feedforward control, adaptive optimization and health diagnosis through in-depth capture of disturbance source events and equipment status.

[0045] Further, the control prediction model includes: fusing the disturbance source event data and state data, and transforming them into an initial state vector through an encoder network; inputting the initial state vector into a time series prediction model based on a Transformer architecture, and performing iterative time series extrapolation within the model at dynamically determined time steps to generate specific future environmental state change curves; determining a target compensation curve opposite to the target compensation curve based on the future environmental state change curves, and decomposing the target compensation curve into discrete target state points; for each discrete target state point, using the target state point and the current environmental state as input, querying a pre-trained offline neural network model for directly mapping state changes to control commands, and directly outputting the specific control command values ​​required to achieve the target state point using a non-iterative forward computation method; combining all the solved specific control command values ​​in chronological order to generate the compensation control sequence, the specific process being as follows. Figure 2 As shown.

[0046] The system first fuses data from two sources: one is disturbance source event data obtained from the Laboratory Information Management System (LIMS) via an API interface, which can be constructed into a vector containing event type (e.g., "liquid nitrogen filling"), key parameters (e.g., "50 liters"), and scheduled execution time; the other is real-time laboratory current state data collected from a multimodal sensor network, which is a high-dimensional state vector containing the current three-dimensional temperature field, suspended particle concentration, and structural vibration readings. These two sets of heterogeneous data are jointly input into a multimodal encoder network. This encoder network has parallel processing branches; for example, embedding layers are used to process categorized event types, and fully connected layers are used to process numerical event parameters and state data. The processing results of each branch are finally concatenated and transformed into a unified, for example, 512-dimensional initial state vector, which is a complete mathematical description of "the present moment when something is about to happen".

[0047] The initial state vector is then input into a time series prediction model based on the Transformer architecture. This model utilizes its core self-attention mechanism to dynamically analyze the intrinsic correlation and importance of different dimensions of information in the initial state vector (such as the type of disturbance event, the current temperature gradient, etc.). Before the inference begins, a small sub-network inside the model adaptively determines an optimal inference time step Δt based on the characteristics of the disturbance event (such as the duration and intensity of its impact). For example, for a rapid and strong disturbance like liquid nitrogen filling, Δt may be determined to be 5 seconds; for a slow disturbance like solar radiation, Δt may be determined to be 60 seconds.

[0048] Subsequently, the model performs iterative forward extrapolation at this dynamic time step: first, it predicts the state at time t+Δt based on the initial state, and then uses the predicted state at time t+Δt as the new input to predict the state at time t+2Δt. This process is repeated until the entire preset prediction time domain (e.g., the next 30 minutes) is covered. Finally, a decoder network transforms a series of high-dimensional state vectors generated within the model into multiple future environmental state change curves with definite physical units.

[0049] After obtaining the predicted environmental state change curve, the system generates a target compensation curve with the opposite effect through numerical inversion and decomposes it into a series of discrete target state points. To calculate the specific control command required to achieve each target state point, this embodiment uses a pre-trained offline inverse dynamic neural network model. The training method of this inverse model (in this embodiment, it is a multilayer perceptron MLP with three hidden layers) is as follows: the input X of its training dataset is the actual environmental state change in history (state (t+1) - state (t)), and the label Y of the training data is the actual control command (command (t)) at the moment that caused this change. In this way, the model directly learns the nonlinear mapping relationship of inferring the cause from the result. During real-time operation, the system feeds each target state point (i.e., a desired state change) as input into this trained inverse model. The model can directly output a multidimensional control vector containing multiple specific control command values ​​(e.g., {VAV-1 target air volume: 350m³ / h, heating coil target power: 2.1kW, ...}) through a single non-iterative forward calculation.

[0050] Finally, the system will sort and combine the multi-dimensional control vectors calculated for all discrete target state points according to their corresponding timestamps. The resulting compensation control sequence is a structured time-series instruction queue, which contains precise instructions that need to be issued to different actuators at different times in the future. This sequence is then submitted to the edge computing and execution control node for final security verification and execution.

[0051] In this embodiment, the encoder network comprises two parallel branches: one branch processes categorical information such as event type in the disturbance source event data through an embedding layer; the other branch uses a network consisting of three fully connected layers to process numerical information such as sensor readings. The number of neurons in each layer is 256, 128, and 64, respectively, and ReLU is used as the activation function. The outputs of the two branches are concatenated and finally mapped to a 512-dimensional initial state vector through a fully connected layer. The time series prediction model based on the Transformer architecture consists of six stacked encoder layers with a model dimension of 512. Each encoder layer contains an 8-head multi-head self-attention mechanism and a feedforward network with 2048 nodes. The inverse dynamic neural network model used for direct mapping is a multilayer perceptron with three hidden layers, each containing 512, 256, and 128 neurons, respectively, using the ReLU activation function, and trained offline using an Adam optimizer with a learning rate of 0.001.

[0052] By combining forward-looking time-series prediction with non-iterative inverse mapping, a control shift from passive response to active intervention is achieved. This enables the calculation of highly customized, time-precisely changing compensation control sequences before the actual occurrence of disturbances, thereby suppressing environmental fluctuations to the greatest extent with minimal adjustment costs and ensuring the ultimate stability of the laboratory environment.

[0053] Further, the process of generating collaborative operating parameters includes: constructing a state input vector from the three-dimensional temperature field data, suspended particle data, and structural vibration data; the multiple heterogeneous strategy models include models with three different optimization objectives: an energy consumption optimization model, a temperature field optimization model, and a particle sedimentation model; the multiple heterogeneous strategy models are independently trained based on different data subsets generated from all historical data through a bootstrap sampling method; the state input vector is simultaneously input into the multiple heterogeneous strategy models for parallel forward inference, and during the forward inference, parameter suggestion values ​​and a quantitative uncertainty index used to characterize the model's confidence in the current suggestion value are output; the multiple parameter suggestion values ​​output by the multiple heterogeneous strategy models are integrated through a dynamic weighted fusion strategy based on the inverse of uncertainty; based on the quantitative uncertainty index output by each model, a fusion weight corresponding to each model is calculated, the fusion weight of any model is inversely proportional to the quantitative uncertainty index output by the model, and the multiple parameter suggestion values ​​are weighted and summed using the fusion weight to generate collaborative operating parameters.

[0054] The multiple heterogeneous policy models in this embodiment are specifically composed of three independent deep neural networks driven by different optimization objectives:

[0055] Energy consumption optimization model: The goal of this model is to minimize the instantaneous total power of the HVAC system. Its training data labels focus more on the system's power consumption, so it tends to output control parameters that prioritize energy saving. In this embodiment, it adopts a lightweight multilayer perceptron (MLP) architecture.

[0056] Temperature field optimization model: The goal of this model is to minimize the gradient and drift over time of the three-dimensional temperature field in the laboratory. Its training data labels focus more on the reading variance of a high-precision temperature sensor array. To better handle spatial data, it adopts a neural network architecture containing convolutional layers (CNN) to understand the temperature field in the form of images.

[0057] Particle sedimentation model: The goal of this model is to maximize the removal efficiency of suspended particles in the air. Its training data labels focus more on the rate of change of laser particle counter readings at key locations. It adopts a neural network architecture containing recurrent layers (RNN) to better capture the time-series dynamics of airflow and particle motion.

[0058] To ensure model heterogeneity, these three models not only have different neural network architectures, but their respective training datasets are also based on different subsets of data generated from all historical data using bootstrap sampling.

[0059] To achieve confidence-based decision fusion, each expert model must quantify its own uncertainty regarding the proposed parameter values ​​while outputting them. This embodiment employs Monte Carlo Dropout technology to achieve this; during the real-time forward inference phase, when a state input vector is fed into an expert model, the model does not perform a single computation. Instead, it performs N independent forward computations (e.g., N=50) consecutively while Dropout is activated (i.e., randomly deactivating some neurons). This results in a distribution of N slightly different parameter proposal values. The mean of this distribution is used as the final parameter proposal value for that model; the variance or standard deviation of this distribution is used as its quantified uncertainty index. A larger variance indicates greater model divergence, i.e., higher uncertainty. After obtaining the parameter proposal values ​​and their corresponding uncertainty indices from all three expert models, the system begins to execute the final fusion decision.

[0060] In terms of specific fusion strategies, the system adopts a dynamic weighting method based on the confidence level of each model. The core of this method is to evaluate the quantitative uncertainty index attached to the output of each heterogeneous strategy model when it outputs its suggested value in real time. For models with high uncertainty index values ​​(i.e., the model lacks "confidence" in the current suggestion), the system will dynamically assign them a lower fusion weight. Conversely, models with lower uncertainty index values ​​and higher "confidence" will obtain a relatively higher weight, thus having a greater say in the final decision. The final collaborative operation mode parameters are the result of weighted summation of the parameter suggested values ​​output by each model and these dynamically calculated weights.

[0061] In a specific embodiment, the structural parameters of the multiple heterogeneous strategy models are as follows: The temperature field optimization model adopts a convolutional neural network (CNN) architecture containing three layers of three-dimensional convolutions (Conv3D); the input three-dimensional temperature field data is processed into a 32×32×8 tensor, the first convolutional layer uses 16 5×5×3 convolutional kernels with a stride of 2; the second and third layers both use 32 3×3×3 convolutional kernels with a stride of 1; all convolutional layers are followed by ReLU activation functions and batch normalization layers; the particle sedimentation model adopts a stacked two-layer long short-term memory network (LSTM), whose hidden state dimension is set to 128 to effectively capture the time-series dynamics of suspended particle concentration changes; the energy consumption optimization model adopts a relatively simple four-layer fully connected network with 256, 128, 64 and 32 neurons respectively.

[0062] Furthermore, before integrating the parameter values ​​output by multiple heterogeneous strategy models, the intelligent decision-making module performs the following steps: providing the parameter adjustment suggestion output by one of the heterogeneous strategy models as input to all other heterogeneous strategy models, predicting the estimated impact of implementing the parameter adjustment suggestion on other optimization objectives; rejecting the parameter adjustment suggestion based on whether any indicator in the estimated impact exceeds a preset catastrophic threshold, and then integrating the results.

[0063] Before integrating the parameter values ​​output by multiple heterogeneous strategy models, this invention introduces a suggestion-impact pre-evaluation mechanism based on cross-validation between models to prevent a single model from outputting potentially destructive control suggestions in extreme cases. Specifically, when the temperature field optimization model outputs a significant control parameter adjustment suggestion, the suggestion is not immediately weighted and merged; instead, it is temporarily stored by the system and used as input to query the energy consumption optimization model and the particle sedimentation model to predict "what impact will this adjustment have on the total energy consumption and suspended particle concentration of the system?" If any other model predicts that its core indicators will deteriorate beyond a preset catastrophic threshold, the system will directly reject the initial adjustment suggestion and trigger a replanning process based on the current state. This mechanism adds a "rational review" step to the entire collaborative decision-making process by establishing a rigid balance between models, ensuring that the final integrated output parameters take into account the bottom-line safety of all objectives.

[0064] By using an integrated decision engine composed of multiple specialized models and employing a dynamic fusion strategy based on model self-uncertainty perception, the collaborative optimization of multiple conflicting objectives is achieved. This mechanism can automatically adjust the weight of each specialized objective according to the specific scenario, ensuring decision robustness while keeping the control focus consistent with the laboratory's most important task at present.

[0065] Furthermore, the process of processing the vibration signal and pressure signal to extract vibration spectrum features and pressure fluctuation pattern data includes: performing a fast Fourier transform on the vibration signal to obtain the vibration spectrum; extracting the fundamental frequency and amplitude of each harmonic related to the rotational speed of the rotating component in the HVAC system, as well as the energy value at the preset bearing fault characteristic frequency, from the vibration spectrum; filling the sample into a fixed-length feature vector according to a preset dimensional order, as vibration spectrum features; calculating the root mean square value, kurtosis, and skewness of the pressure signal within a sliding time window and using them as statistical features; performing a continuous wavelet transform to extract the time-frequency energy distribution of the pressure transient event; and concatenating the data obtained after vectorizing the statistical features and the time-frequency energy distribution to form pressure fluctuation pattern data characterizing the pressure fluctuation pattern.

[0066] Specifically, the raw, high-frequency time-series signals acquired from the multimodal sensing module are transformed into structured, fixed-length feature vectors for subsequent analysis by the AI ​​health baseline model. The internal processing flow of this unit is divided into two parallel parts: one for vibration signals and one for pressure signals. For vibration signal processing, the system first acquires high-frequency sampled raw time-series signals from accelerometers installed on key rotating components (such as the main circulating water pump bearing housing). These signals are segmented into fixed-length data segments, and after applying window functions (such as the Hanning window) to reduce spectral leakage, a Fast Fourier Transform (FFT) is performed to convert them from the time domain to the frequency domain to obtain a high-resolution vibration spectrum. Subsequently, the feature extraction step accurately extracts two types of key information from this spectrum: one is the fundamental frequency and the amplitude of its harmonics corresponding to the equipment rotation speed; the other is the energy values ​​at multiple specific fault characteristic frequencies pre-calculated based on the equipment bearing model. Ultimately, the extracted harmonic amplitudes and fault energy values ​​are filled into a fixed-length feature vector according to a pre-defined dimensional order. This vector represents the vibration spectrum characteristics that comprehensively characterize the health status of the equipment. For pressure signal processing, the system employs a dual analysis path to capture its complex dynamics. First, within a sliding time window, the raw signals from the dynamic pressure sensors in the main pipeline are statistically analyzed to calculate indicators such as root mean square, kurtosis, and skewness, characterizing the macroscopic form of pressure fluctuations. To capture pressure transients caused by non-stationary events such as valve opening and closing, the system performs continuous wavelet transform on the same data window, decomposing the signal into a two-dimensional time-frequency spectrum that clearly shows the energy distribution of the transient event in the time and frequency dimensions. Finally, this two-dimensional time-frequency spectrum is flattened into a one-dimensional vector and concatenated with the previously calculated statistical feature values ​​to form a high-dimensional fusion feature vector that fully describes the pressure dynamics within the pipeline, serving as the final pressure fluctuation pattern.

[0067] By combining frequency domain and time-frequency domain analysis, structured feature vectors that are highly sensitive to early and subtle faults can be extracted from the high-noise raw sensor signals. This not only provides high-quality, high-information-density input for the accurate diagnosis of subsequent AI models, but also significantly improves the entire system's ability to identify potential physical risks from massive amounts of dynamic data.

[0068] Furthermore, the AI ​​health baseline model includes:

[0069] The input embedding layer transforms the extracted vibration spectrum features and pressure fluctuation pattern data into high-dimensional feature vectors.

[0070] The Transformer encoding layer receives the high-dimensional feature vector and models the intrinsic correlation between each feature dimension through an internal multi-head self-attention mechanism, and outputs a health prediction vector for the current state.

[0071] The anomaly detection layer generates an anomaly score by calculating the deviation between the health prediction vector and the actual input feature vector. When the anomaly score exceeds a dynamic threshold, the system is determined to be abnormal.

[0072] The root cause diagnosis layer is activated when an anomaly is detected. By analyzing the attention weights and model gradient information within the Transformer encoding layer, it calculates the contribution score of each input feature to the prediction bias, identifies the key feature with the highest contribution score, and matches this key feature with a pre-defined physical fault feature signature library. The output includes diagnostic information containing the specific physical root cause location. The specific process is as follows: Figure 3 As shown.

[0073] The input embedding layer is responsible for the final preprocessing of the incoming, structured vibration spectrum feature vector and pressure fluctuation pattern feature vector to adapt to the input requirements of the subsequent Transformer model. First, the two feature vectors are concatenated into a single, high-dimensional combined feature vector. Then, this combined vector is linearly mapped through a fully connected layer to adjust its dimension to match the internal working dimension of the Transformer encoder (e.g., 512 dimensions). Crucially, a sinusoidal position encoding vector is added to this feature vector to give each feature in the sequence (e.g., a specific frequency cell) its unique and absolute position information, thus compensating for the lack of temporal awareness in the Transformer model itself.

[0074] The Transformer encoding layer employs a masked autoencoder self-supervised learning approach for training. During training, the system uses only HVAC system data from confirmed healthy states. Before each input of a combined feature vector, a subset of its dimensions (e.g., 15%) is randomly masked by setting their values ​​to zero. This corrupted vector is fed into the Transformer encoding layer, which uses its internal multi-head self-attention mechanism to learn the complex, non-linear relationships between the various feature dimensions in the health data. The model's training objective is to accurately predict and reconstruct the original values ​​of those masked dimensions. In real-time operation, the unmasked, complete feature vector is input to the trained Transformer encoding layer. This layer processes the input vector based on its learned health pattern knowledge and outputs a health prediction vector.

[0075] The anomaly detection layer is responsible for quantifying the degree of system anomaly based on the output of the encoding layer. This is achieved by calculating the root mean square error between the actual input feature vector and the health prediction vector output by the Transformer encoding layer, generating a real-time anomaly score. Simultaneously, the system calculates a dynamically changing health status threshold based on the statistical distribution (e.g., mean and three standard deviations) of all anomaly scores from confirmed healthy operation over a past period (e.g., 24 hours). When the real-time calculated anomaly score significantly exceeds this dynamic threshold, the system determines that the HVAC system has malfunctioned.

[0076] The root cause diagnosis layer is activated after the anomaly detection layer issues an alarm. Its goal is to achieve interpretable AI and accurately locate the root cause of the problem. This embodiment uses a SHAP-based algorithm, which is used to analyze and calculate how much each dimension (i.e., each frequency bin and each statistical feature) of the input feature vector contributes to the output that caused the huge prediction bias (high anomaly score).

[0077] After the calculation is completed, the system will obtain a contribution score vector with the same dimension as the input vector. The system will identify several key features with the highest contribution scores (for example, the third harmonic amplitude of the fundamental frequency and the kurtosis of the pressure signal have the highest contribution scores). Finally, the system will combine these key features and match them with the patterns preset in the physical fault feature signature library (for example, the combined pattern is highly consistent with the fault signature of slight imbalance of wind turbine blades), thereby outputting the final diagnostic information.

[0078] Preferably, the present invention implements the physical fault feature signature library as a JSON-based rule database that can be configured by operation and maintenance personnel. Each entry in the library corresponds to a known physical fault. The entry contains a unique fault ID, a fault name description, and one or more matching rules consisting of a "feature-condition-threshold" triple. When multiple calculated key contribution features simultaneously satisfy all matching rules under a certain fault entry, the system determines that the specific fault has occurred. For example, the fault signature entry for "slight imbalance of wind turbine blades" can be defined as: {"fault_id": "F001", "ault_name": "slight imbalance of wind turbine blades", "rules": [{"feature": "vibration spectrum - fundamental frequency 3rd harmonic amplitude", "condition": ">", "threshold": 0.5}, {"feature": "pressure signal - kurtosis", "condition": "in_range", "threshold": [3.5, 4.5]}]}. This structured definition makes the basis for fault diagnosis clear and is easy to expand and maintain in the future.

[0079] Through Transformer's self-supervised learning, the health paradigm of equipment can be deeply understood without a large number of fault samples; its dynamic threshold can effectively adapt to changes in operating conditions and reduce false alarms; by using the SHAP algorithm to make the AI ​​decision-making process transparent, faults can be quantified and traced back to specific physical characteristics, providing a clear and reliable basis for predictive maintenance and realizing the improvement from passive alarm to proactive diagnosis.

[0080] Furthermore, the process of driving the actuators of the HVAC system includes: fusing the compensation control sequence, collaborative operation mode parameters, and diagnostic information to form a control intent; using the compensation control sequence as a feedforward adjustment benchmark, the collaborative operation mode parameters as a feedback optimization target, and the diagnostic information as a dynamic constraint; dynamically adjusting the fusion strategy to reduce the operating load of the component when the diagnostic information indicates a health risk; parsing the control intent into a specific physical control command sequence for each actuator within the HVAC system by querying the control command mapping table; performing item-by-item safety checks on the amplitude, rate of change, and combination logic of the physical control command sequence based on preset equipment physical limits, process safety requirements, and energy consumption limit rules; adjusting the command parameters exceeding the safety boundary to the safety boundary value for commands exceeding the single value limit, and re-checking the logical safety of the entire command combination; if the check passes, issuing the corrected command; if the check fails, intercepting the issuance of the command and triggering an alarm.

[0081] Specifically, within each control cycle, the node receives data packets from the cloud platform for three different roles: a compensation control sequence as a reference for feedforward adjustment, cooperative operation mode parameters as a feedback optimization target, and diagnostic information reporting the system's health status. In the initial healthy state, the node uses the compensation control sequence as the basis for the open-loop control plan, while simultaneously running a local PID controller to perform closed-loop fine adjustment using the cooperative operation mode parameters as the real-time setpoint, thereby responding to small real-time changes while executing long-term plans.

[0082] The core of this process lies in its ability to dynamically adjust the fusion strategy based on diagnostic information. For example, when a node receives diagnostic information such as: {Component ID: VAV-03, Fault mode: Vacuum valve actuator response lag, Risk level: Medium, Recommendation: Avoid frequent or small-scale adjustments}, its fusion strategy will be dynamically adjusted. The system will reduce the weight of fine-tuning the airflow of VAV-03 in the collaborative operation mode parameters and actively compensate by increasing the total supply air pressure of the main air handling unit or fine-tuning other healthy VAV terminals. The essence of this adjustment is to ensure the overall environmental stability and operational safety at the expense of local optimization and energy efficiency.

[0083] The merged "control intent" (e.g., {AHU supply air temperature: 18.5℃, VAV-03 damper opening: 45%, ...}) needs to be parsed into physical instructions that the actuator can understand. This is done by querying a control instruction mapping table, which is a pre-set database containing the conversion relationship from physical targets to specific control signals. For example, "AHU supply air temperature 18.5℃" is parsed as "chilled water regulating valve opening signal: 5.2V"; "VAV-03 damper opening 45%" is parsed as "damper actuator pulse signal: send 90 valve opening pulses".

[0084] In the final moments before an instruction is issued, the system performs a rigorous, item-by-item safety check. The first layer of check examines whether the value and rate of change of each instruction in the sequence exceed the physical limits of the device. After all individual values ​​have been corrected to safe ranges, the second layer of check examines whether there are any logical conflicts in the entire instruction combination, such as an erroneous instruction to simultaneously perform maximum cooling and heating. Only after both layers of checks pass will the corrected instruction sequence be finally issued to the actuator. If the second layer of logical check fails, the instruction will be intercepted and a high-level alarm will be triggered to ensure the absolute operational safety of the system.

[0085] The actuators driving the HVAC system further include: an actuator physical model based on the physical response time, action energy consumption curves, and system thermal inertia parameters of each actuator; receiving a sequence of physical control commands and treating the sequence of physical control commands as a multi-objective state within a future time window; solving the optimal execution path problem with the optimization objectives of reaching the target state in the shortest time and minimizing total execution energy consumption; performing time-sequence rearrangement and delivery time optimization on each command in the physical control command sequence to generate an optimized execution queue; and driving the actuators in the HVAC system according to the order and timestamps defined in the execution queue.

[0086] Specifically, to address the discrepancy between ideal control commands and the dynamic characteristics (such as delay and inertia) of physical actuators, and to further reduce instantaneous energy consumption and mechanical wear during system adjustment, the execution control module integrates a physical model-based control command timing optimization engine before ultimately driving the actuators. This engine incorporates a physical model of each actuator, such as valves and dampers, precisely defining its complete action response time (e.g., 45 seconds for the VAV-01 damper to go from 0% to 100%), the energy consumption curves at different opening degrees, and the system thermal inertia caused by its actions (e.g., 90 seconds for the supply air temperature to stabilize after the chilled water valve reaches its target opening). When the engine receives a set of target physical control commands generated by the upper-level decision module (e.g., "Set the VAV-01 opening to 80% and increase the main fan frequency to 45Hz within 10 seconds"), it does not immediately issue them. Instead, it performs timing rearrangement and path optimization on these commands by solving a constrained programming problem with the goal of minimizing total execution energy consumption. For example, after calculation, the engine found that if "the main fan frequency is first increased to 45Hz, and after waiting for 2 seconds for the pipeline static pressure to stabilize, the VAV-01 air valve is then driven to open", the overall instantaneous power consumption impact and mechanical stress are much smaller than the two actions being executed simultaneously. Therefore, the engine will generate an optimized execution queue with precise timestamps and strictly follow the timing defined by the queue to smoothly issue instructions to the physical actuators one by one.

[0087] By optimizing the physical execution timing of instructions, instantaneous power consumption surges and equipment mechanical wear can be significantly reduced. It avoids action conflicts between actuators, making the system adjustment process smoother and more stable, thereby effectively extending equipment life and improving the overall operational efficiency of the system.

[0088] Furthermore, the security boundary rules are configurable rule bases. The rule bases dynamically load security rule configuration files in JSON format corresponding to the current working scenario of the laboratory obtained from the laboratory information management system, and adjust the specific limits and logic of the verification.

[0089] The security boundary rules are not a fixed, globally effective set of rules, but rather a configurable rule base dynamically associated with the current working scenario of the laboratory. Specifically, the system obtains the current state of the laboratory in real time through an interface with the Laboratory Information Management System (LIMS), such as "Standby_Mode" and "High_Sensitivity_Optical_Experiment_Mode". Each mode corresponds to an independent security rule configuration file encoded in JSON (JavaScript Object Notation) format. When the system enters the "High_Sensitivity_Optical_Experiment_Mode", it loads the corresponding JSON configuration file, which explicitly defines more stringent security boundaries, such as tightening the upper limit of the fan frequency change rate from 5Hz / s to 0.5Hz / s and adding a "valve_cycle_prohibited" logical lock. This scenario-aware security boundary makes the security protection capability of this invention no longer static, but can be precisely and coded adaptively adjusted according to the actual needs of the experimental process, achieving refined protection for core scientific research tasks.

[0090] By integrating feedforward, feedback, and diagnostic information, a resilient control mode with forward planning and real-time fine-tuning capabilities is constructed. This mode can dynamically adjust the strategy to ensure system stability when components malfunction, and a multi-level verification mechanism ensures that the final output instructions to the actuator are both safe and logically sound.

[0091] By integrating prediction, optimization, and diagnostic control strategies, this system offers a significant comprehensive improvement to high-precision laboratory environmental control. It introduces model-based predictive feedforward compensation control, which proactively anticipates and intervenes in the substantial impact of disturbances on the environment, achieving higher environmental stability than traditional feedback control. Simultaneously, the system collaboratively decides using multiple heterogeneous models focusing on different objectives such as energy consumption, temperature field, and cleanliness, employing a dynamic fusion strategy with uncertainty awareness to reliably weigh multiple conflicting objectives, resulting in more robust and reasonable control decisions and improved operational economy. Furthermore, the system possesses AI-based weak fault diagnosis capabilities, enabling online identification of early equipment performance degradation and guiding predictive maintenance through precise root cause analysis, effectively enhancing the long-term operational reliability of the system.

[0092] Example 2:

[0093] To improve the automatic air conditioning capabilities in a constant temperature and humidity intelligent physics laboratory, this invention introduces an intelligent laboratory air quality control method, the specific process of which is as follows: Figure 4 As shown.

[0094] First, three-dimensional temperature and humidity field data are collected in a grid-like manner within the laboratory. High-precision composite sensors are also used to collect local microenvironmental data around key equipment such as the core optical platform and lasers. Simultaneously, structural vibration and airborne particle concentration are collected, and vibration and pressure signals from the fans and pumps of the main air handling unit are obtained during operation. The pre-determined experimental plan is actively retrieved from the Laboratory Information Management System (LIMS) via a program interface, and operations such as "turning on a high-power laser" are interpreted as future disturbance source event data carrying specific time and power parameters.

[0095] Next, after receiving the disturbance event data from LIMS, the method combines this data with the current laboratory state data and inputs it into a time series prediction model based on the Transformer architecture. The model will proactively deduce the specific impact curve of the disturbance event on indoor temperature and humidity over a future period. Subsequently, based on this prediction curve, the method generates a time-series compensation control sequence through a pre-trained inverse dynamic neural network model. This sequence can instruct the HVAC system to intervene and adjust in advance and smoothly before the disturbance actually occurs, thereby actively offsetting the upcoming environmental fluctuations.

[0096] Meanwhile, the real-time collected laboratory status data is used to form an input vector, which is then simultaneously input into multiple heterogeneous strategy models trained with different optimization objectives (such as lowest energy consumption, most stable temperature and humidity, and highest cleanliness). Through a dynamic weighting strategy based on uncertainty perception, the method integrates the suggested values ​​output by multiple models to generate a set of collaborative operating parameters. These parameters can automatically make the most reasonable dynamic trade-off between multiple conflicting objectives such as energy saving and environmental stability, depending on whether the laboratory is currently in standby or experimental state.

[0097] Subsequently, vibration and pressure signals collected from HVAC equipment are continuously processed, and feature vectors that can characterize subtle performance changes of the equipment are extracted through methods such as Fast Fourier Transform and Continuous Wavelet Transform. Then, the method analyzes these feature vectors through an AI health baseline model. This model has mastered the paradigm of the equipment in a healthy state through self-supervised learning, so it can generate anomaly scores by calculating the deviation between the current features and the health paradigm, and output diagnostic information containing specific physical root causes when anomalies occur, such as "early scaling has occurred in a certain transducer of the humidifier".

[0098] Finally, the aforementioned generated compensation control sequence (feedforward), cooperative operating parameters (feedback), and diagnostic information (constraints) are integrated to form the final control intent. For example, the diagnostic information will enable the method to proactively reduce the operating load of the faulty component during control. Then, the method resolves the control intent into physical control commands for each actuator by querying a mapping table. Before the command is issued, a safety boundary rule module will strictly verify the amplitude, rate of change, and combination logic of the command. Only after ensuring absolute safety will the final command be issued to drive the actuators in the HVAC system to complete precise air quality control.

[0099] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent laboratory air quality control system, characterized in that, include: The data acquisition module collects state data from the physics laboratory, including three-dimensional temperature field data, suspended particle data, and structural vibration data; acquires disturbance source event data; and acquires vibration and pressure signals from the HVAC system. The intelligent decision-making module inputs state data and disturbance source event data into the control prediction model for prediction, generating a compensation control sequence; it constructs an input vector from the state data; it simultaneously inputs the input vector into multiple heterogeneous strategy models for forward inference, integrates the parameter values ​​output by the multiple heterogeneous strategy models, and generates collaborative operation parameters. The fault diagnosis module processes vibration and pressure signals to extract vibration spectrum features and pressure fluctuation pattern data; it analyzes the vibration spectrum features and pressure fluctuation pattern data through an AI health baseline model and outputs diagnostic information. The execution control module integrates the compensation control sequence, the cooperative operation parameters, and the diagnostic information to generate physical control commands, and verifies the physical control commands through safety boundary rules to drive the actuators in the HVAC system.

2. The intelligent laboratory air quality control system according to claim 1, characterized in that, The data acquisition module's data acquisition process includes: deploying a temperature sensor array in a grid pattern along a three-dimensional spatial coordinate system within the laboratory to acquire three-dimensional temperature field data; deploying laser particle counters in the experimental equipment area, air supply vents, and return air vents to acquire suspended particle data; installing accelerometers on the building's load-bearing structure, instrument platform, and HVAC unit base to acquire structural vibration data; identifying and recording disturbance source event data such as door opening and closing, personnel movement, and equipment start-up and shutdown; and installing vibration and pressure sensors in the HVAC system to acquire vibration and pressure signals.

3. The intelligent laboratory air quality control system according to claim 1, characterized in that, The control prediction model includes: fusing the disturbance source event data and state data, and converting them into an initial state vector through an encoder network; inputting the initial state vector into a time series prediction model based on a Transformer architecture, and performing iterative time series extrapolation within the model at dynamically determined time steps to generate specific future environmental state change curves; determining a target compensation curve opposite to the target compensation curve based on the future environmental state change curves, and decomposing the target compensation curve into discrete target state points; for each discrete target state point, using the target state point and the current environmental state as input, querying a pre-trained offline neural network model for directly mapping state changes to control commands, and directly outputting the specific control command values ​​required to achieve the target state point using a non-iterative forward computation method; and combining all the solved specific control command values ​​in chronological order to generate the compensation control sequence.

4. The intelligent laboratory air quality control system according to claim 1, characterized in that, The process of generating collaborative operating parameters includes: constructing a state input vector from the three-dimensional temperature field data, suspended particle data, and structural vibration data; the multiple heterogeneous strategy models include models with three different optimization objectives: an energy consumption optimization model, a temperature field optimization model, and a particle sedimentation model; the multiple heterogeneous strategy models are independently trained based on different data subsets generated from all historical data using a bootstrap sampling method; the state input vector is simultaneously input into the multiple heterogeneous strategy models for parallel forward inference, and during the forward inference, parameter suggestion values ​​and a quantitative uncertainty index used to characterize the model's confidence in the current suggestion value are output; the multiple parameter suggestion values ​​output by the multiple heterogeneous strategy models are integrated using a dynamic weighted fusion strategy based on the inverse of uncertainty; based on the quantitative uncertainty index output by each model, a fusion weight corresponding to each model is calculated, the fusion weight of any model is inversely proportional to the quantitative uncertainty index output by the model, and the multiple parameter suggestion values ​​are weighted and summed using the fusion weight to generate collaborative operating parameters.

5. The intelligent laboratory air quality control system according to claim 1, characterized in that, The process of processing vibration and pressure signals to extract vibration spectrum features and pressure fluctuation pattern data includes: performing a fast Fourier transform on the vibration signal to obtain the vibration spectrum; extracting the fundamental frequency and amplitude of each harmonic related to the rotational speed of the rotating component in the HVAC system, as well as the energy value at the preset bearing fault characteristic frequency, from the vibration spectrum; filling the sample into a fixed-length feature vector according to a preset dimensional order, as vibration spectrum features; calculating the root mean square value, kurtosis, and skewness of the pressure signal within a sliding time window and using them as statistical features; performing a continuous wavelet transform to extract the time-frequency energy distribution of the pressure transient event; and concatenating the data obtained after vectorizing the statistical features and the time-frequency energy distribution to form pressure fluctuation pattern data characterizing the pressure fluctuation pattern.

6. The intelligent laboratory air quality control system according to claim 1, characterized in that, The AI ​​health baseline model includes: The input embedding layer transforms the extracted vibration spectrum features and pressure fluctuation pattern data into high-dimensional feature vectors. The Transformer encoding layer receives the high-dimensional feature vector and models the intrinsic correlation between each feature dimension through an internal multi-head self-attention mechanism, and outputs a health prediction vector for the current state. The anomaly detection layer generates an anomaly score by calculating the deviation between the health prediction vector and the actual input feature vector. When the anomaly score exceeds a dynamic threshold, the system is determined to be abnormal. The root cause diagnosis layer is activated when an anomaly is detected. By analyzing the attention weights and model gradient information within the Transformer encoding layer, it calculates the contribution score of each input feature to the prediction bias, identifies the key feature with the highest contribution score, and matches the key feature with a preset physical fault feature signature library to output diagnostic information containing the specific physical root cause location.

7. The intelligent laboratory air quality control system according to claim 1, characterized in that, The process of driving the actuators of the HVAC system includes: fusing the compensation control sequence, collaborative operation mode parameters, and diagnostic information to form a control intent; using the compensation control sequence as a feedforward adjustment benchmark, the collaborative operation mode parameters as a feedback optimization target, and the diagnostic information as a dynamic constraint; dynamically adjusting the fusion strategy to reduce the operating load of the component when the diagnostic information indicates a health risk; parsing the control intent into a specific physical control command sequence for each actuator in the HVAC system by querying the control command mapping table; performing item-by-item safety checks on the amplitude, rate of change, and combination logic of the physical control command sequence according to preset equipment physical limits, process safety requirements, and energy consumption limit rules; for commands exceeding safety boundaries, adjusting the command parameters exceeding the single value limit to the safety boundary value, and re-checking the logical safety of the entire command combination; if the check passes, issuing the corrected command; if the check fails, intercepting the issuance of the command and triggering an alarm.

8. A method for intelligent laboratory air quality control, characterized in that: The system collects state data from a physics laboratory, including three-dimensional temperature field data, suspended particle data, and structural vibration data; acquires disturbance source event data; acquires vibration and pressure signals from the HVAC system; inputs the state data and disturbance source event data into a control prediction model for prediction, generating a compensation control sequence; constructs an input vector from the state data; simultaneously inputs the input vector into multiple heterogeneous strategy models for forward inference, integrates the parameter values ​​output by the multiple heterogeneous strategy models, and generates collaborative operation parameters; processes the vibration and pressure signals to extract vibration spectrum features and pressure fluctuation pattern data; analyzes the vibration spectrum features and pressure fluctuation pattern data using an AI health baseline model, outputting diagnostic information; fuses the compensation control sequence, the collaborative operation pattern vector, and the diagnostic information to generate physical control commands, and verifies the physical control commands using safety boundary rules to drive the actuators within the HVAC system.

Images