Intelligent laboratory air quality control system and method

Abstract

The present application relates to the HVAC control technical field, specifically to a kind of intelligent laboratory air quality control system and method, including data acquisition module, intelligent decision module, fault diagnosis module and execution control module, intelligent decision module is based on disturbance data generation feedforward compensation control sequence, simultaneously through multiple heterogeneous strategy model generates feedback multi-objective collaborative operation parameter;Fault diagnosis module passes through AI health degree baseline model analysis equipment health data, exports diagnostic information as dynamic constraint;Execution control module fuses the above-mentioned compensation control sequence, collaborative parameter and diagnostic information, generates the final physical control instruction verified by safety boundary rule, drives HVAC actuator.The present application is fused by prediction, optimization and diagnosis control strategy, significantly improves the stability of laboratory environment, operating economy and the reliability of long-term operation of system.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of heating, ventilation and air conditioning control technology, in particular to an intelligent laboratory air quality control system and method. BACKGROUND

[0002] The primary design goal of a heating, ventilation and air conditioning (HVAC) system in a laboratory environment 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 maintain a high air change rate (ACH) to effectively dilute and remove potential air pollutants.

[0003] One basic technical solution to achieve the above goal is a constant air volume (CAV) system. A CAV system ensures safety margins by continuously supplying a fixed, usually higher air volume designed for the worst-case scenario, to the laboratory. The control logic of this system is relatively simple and reliable in operation. Its energy consumption is directly related to this constant high air volume; a variable air volume (VAV) system developed on this basis aims to improve energy use efficiency. A VAV system can link ventilation volume to some direct physical state parameters (e.g., the opening degree of a fume hood adjustment window). Its control system usually uses feedback control logic (such as a PID controller), i.e., when a sensor measures that an environmental parameter (such as temperature) deviates from the set point, the system will adjust the actuator (such as a valve) accordingly to correct the deviation.

[0004] The above existing technology provides an effective environmental protection solution for conventional laboratories; however, in some advanced scientific research fields with extreme requirements for environmental stability, such as high-precision physical laboratories in optics, metrology, quantum information, etc., in addition to the conventional temperature and humidity and safety requirements, there are also multiple more stringent and interrelated control objectives, for example, the need to maintain temperature stability, 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 and stop of specific scientific research equipment, the filling of low-temperature medium, etc.) itself becomes a predictable disturbance source that will have a combined effect on the above-mentioned multiple high-precision environmental parameters, therefore, there is a technical problem to be solved in the field: how to actively intervene and compensate before the disturbance has a substantial impact on the environment, to achieve environmental stability.

[0006] To this end, an intelligent laboratory air quality control system and method are proposed. SUMMARY

[0007] The purpose of the present application is to provide an intelligent laboratory air quality control system and method, which significantly improves the stability of the laboratory environment, the operation economy and the reliability of long-term operation of the system through the control strategy of the integration of prediction, optimization and diagnosis. It 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 a feedback multi-objective collaborative operation parameter through multiple heterogeneous strategy models. The fault diagnosis module analyzes equipment health data through an AI health degree baseline model and outputs diagnostic information as a dynamic constraint. The execution control module integrates the above compensation control sequence, collaborative parameters and diagnostic information to generate a final physical control instruction that has passed the safety boundary rule check, which drives the heating, ventilation and air conditioning (HVAC) actuator.

[0008] To achieve the above purpose, the present application provides the following technical solutions:

[0009] An intelligent laboratory air quality control system, comprising:

[0010] A data acquisition module acquires state data of a physical laboratory, including three-dimensional temperature field data, suspended particle data and structural vibration data; acquires disturbance source event data; and acquires vibration signals and pressure signals of a heating, ventilation and air conditioning (HVAC) system.

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

[0012] A fault diagnosis module processes the vibration signals and the pressure signals to extract vibration frequency spectrum features and pressure fluctuation mode data; analyzes the vibration frequency spectrum features and the pressure fluctuation mode data through an AI health degree baseline model to output diagnostic information.

[0013] An execution control module integrates the compensation control sequence, the collaborative operation parameters and the diagnostic information to generate a physical control instruction, and checks the physical control instruction through a safety boundary rule to drive the actuator in the HVAC system.

[0014] Preferably, the data acquisition module acquires data by: deploying a temperature sensor array in a three-dimensional space coordinate grid in the laboratory to obtain three-dimensional temperature field data; deploying a laser particle counter at the experimental equipment area, air supply outlet and air return inlet to collect suspended particle data; installing an acceleration sensor on the building load-bearing structure, instrument platform and HVAC unit base to collect structural vibration data; identifying and recording disturbance source event data such as opening and closing doors, personnel walking, and equipment starting and stopping; and installing a vibration sensor and a pressure sensor on the HVAC to obtain vibration signals and pressure signals.

[0015] Preferably, the control prediction model includes: fusing the disturbance source event data and the state data, and converting the fused data 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 iteratively performing time series deduction inside the model according to a dynamically determined time step to generate a specific future environmental state change curve; determining a target compensation curve opposite to the target compensation according to the future environmental state change curve, and decomposing the target compensation curve into discrete target state points; for each discrete target state point, inputting the target state point and the current environmental state as inputs, querying a pre-off-line trained neural network model for directly mapping state changes to control instructions, and directly outputting specific control instruction values required to achieve the target state point in a non-iterative forward calculation manner, and combining all the solved specific control instruction values in chronological order to generate the compensation control sequence.

[0016] Preferably, the process of generating the collaborative operation 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 an energy consumption optimization model, a temperature field optimization model and a particle deposition model with three different optimization objectives; 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 reasoning, and during the forward reasoning, parameter suggestion values and a quantitative uncertainty index representing the confidence of the model in the current suggestion values are output; multiple parameter suggestion values output by the multiple heterogeneous strategy models are integrated through a dynamic weighted fusion strategy based on the inverse of the uncertainty, 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 the collaborative operation parameters.

[0017] Preferably, the processing of the vibration signal and the pressure signal to extract the vibration spectrum feature and the pressure fluctuation mode data comprises: performing fast Fourier transform on the vibration signal to obtain a vibration spectrum, extracting the amplitude of the fundamental frequency and each order harmonic related to the rotating component speed of the HVAC system and the energy value at the preset bearing fault characteristic frequency from the vibration spectrum, and filling the extracted features into a fixed-length feature vector according to a preset dimension order as the vibration spectrum feature; calculating the root mean square value, kurtosis and skewness of the pressure signal in a sliding time window as statistical features, performing continuous wavelet transform to extract the time-frequency energy distribution of the pressure transient event, and splicing the statistical features and the time-frequency energy distribution to obtain the pressure fluctuation mode data representing the pressure fluctuation mode.

[0018] Preferably, the AI health baseline model comprises:

[0019] An input embedding layer is used to convert the extracted vibration spectrum feature and the pressure fluctuation mode data into a high-dimensional feature vector.

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

[0021] An anomaly determination layer calculates the deviation between the health prediction vector and the actual input feature vector to generate an anomaly score, and determines that an anomaly occurs when the anomaly score exceeds a dynamic threshold.

[0022] A root cause diagnosis layer is activated when an anomaly is determined to occur, calculates the contribution score of each input feature to the prediction deviation by analyzing the attention weight and model gradient information in the Transformer encoding layer, identifies the key feature with the highest contribution score, and matches the key feature with a preset physical fault feature signature library to output diagnosis information containing specific physical root cause positioning.

[0023] Preferably, the process of driving the actuators of the HVAC system comprises: fusing the compensation control sequence as a feedforward regulation reference, the cooperative operation mode parameters as a feedback optimization target, and the diagnostic information as a dynamic constraint condition, to form a control intention, and dynamically adjusting the fusion strategy to reduce the operation load of the components when the diagnostic information indicates that the components have health risks; resolving the control intention into a specific physical control instruction sequence for each actuator in the HVAC system by querying a control instruction mapping table; performing item-by-item safety checks on the amplitude, rate of change, and combination logic of the physical control instruction sequence according to preset device physical limits, process safety requirements, and energy consumption limit rules, adjusting the instruction parameters that exceed the single-item numerical limits to the safety boundary values for instructions that exceed the safety boundaries, and rechecking the logical safety of the entire instruction combination, issuing the corrected instructions if the checking is passed, and intercepting the issuance of the instructions and triggering an alarm if the checking is not passed.

[0024] An intelligent laboratory air quality control method comprises:

[0025] Collecting state data of a physical laboratory, the state data comprising three-dimensional temperature field data, suspended particle data, and structural vibration data; acquiring disturbance source event data; acquiring vibration signals and pressure signals of the HVAC system; inputting the state data and the disturbance source event data into a control prediction model for prediction to generate a compensation control sequence; constructing an input vector from the state data; inputting the input vector into multiple heterogeneous strategy models for forward reasoning, integrating the parameter values output by the multiple heterogeneous strategy models to generate cooperative operation mode parameters; processing the vibration signals and the pressure signals to extract vibration frequency spectrum features and pressure fluctuation mode data; analyzing the vibration frequency spectrum features and the pressure fluctuation mode data through an AI health degree baseline model to output diagnostic information; fusing the compensation control sequence, the cooperative operation mode parameters, and the diagnostic information to generate physical control instructions, and checking the physical control instructions through safety boundary rules to drive the actuators in the HVAC system.

[0026] Compared with the prior art, the present application has the following beneficial effects:

[0027] 1. By introducing a feedforward-based model prediction compensation control mechanism, the present application can actively 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 the adaptive cooperative optimization mechanism based on "expert integration" and "uncertainty perception", the application can dynamically and reliably balance multiple conflicting control objectives. The fusion strategy based on uncertainty reduces the risk of misjudgment due to defects in a single model, enabling more stable and reasonable control decisions in complex and variable working conditions.

[0029] 3. By introducing AI-based weak fault diagnosis capability at a deep physical level, the application can identify early performance degradation or minor abnormalities in equipment that conventional systems cannot perceive. This predictive maintenance capability helps address issues before they become major problems, improving long-term system reliability and providing clear guidance for maintenance work due to its precise root cause positioning. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 An intelligent laboratory air quality control system structure schematic diagram is provided for the embodiments of the application.

[0031] Figure 2 A flowchart for generating a compensation control sequence is provided for the embodiments of the application.

[0032] Figure 3 An AI health degree baseline model diagnosis flowchart is provided for the embodiments of the application.

[0033] Figure 4 An intelligent laboratory air quality control method flowchart is provided for the embodiments of the application. DETAILED DESCRIPTION

[0034] The technical solutions in the embodiments of the application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the application, not all embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0035] Please refer to Figures 1 to 4 The application provides an intelligent laboratory air quality control system and method, and the technical solutions are as follows:

[0036] Embodiment one:

[0037] An intelligent laboratory air quality control system, as shown in Figure 1 , includes:

[0038] The data acquisition module acquires state data of the physical laboratory, the state data including three-dimensional temperature field data, suspended particle data and structure vibration data; acquires disturbance source event data; acquires vibration signals and pressure signals of the heating and ventilation air conditioner;

[0039] The intelligent decision module inputs the state data and the disturbance source event data into a control prediction model for prediction to generate a compensation control sequence; constructs an input vector from the state data; simultaneously inputs the input vector into a plurality of heterogeneous strategy models for forward reasoning, integrates parameter values output by the plurality of heterogeneous strategy models to generate cooperative operation parameters;

[0040] The fault diagnosis module processes the vibration signals and the pressure signals, extracts vibration frequency spectrum features and pressure fluctuation mode data; analyzes the vibration frequency spectrum features and the pressure fluctuation mode data through an AI health degree baseline model to output diagnosis information;

[0041] The execution control module fuses the compensation control sequence, the cooperative operation parameters and the diagnosis information to generate physical control instructions, checks the physical control instructions through a safety boundary rule, and drives an execution mechanism in the heating and ventilation air conditioner.

[0042] Further, the data acquisition module acquires data in the following process: deploying a temperature sensor array along a three-dimensional space coordinate system in the laboratory to acquire three-dimensional temperature field data; deploying laser particle counters in the experimental equipment area, the air supply port and the return air port to acquire suspended particle data; installing acceleration sensors on the building bearing structure, the instrument platform and the heating and ventilation air conditioner base to acquire structure vibration data; identifying and recording disturbance source event data such as opening and closing doors, personnel walking and equipment starting and stopping; installing vibration sensors and pressure sensors on the heating and ventilation air conditioner to acquire vibration signals and pressure signals.

[0043] Specifically, the data acquisition process of the system is realized by a multi-modal sensing module, which is configured to build a complete information laboratory dynamic dataset; in order to realize the accurate perception of the temperature field inside the laboratory, the module is deployed in a high-density grid along the three-dimensional spatial coordinate system in the laboratory, and an array composed of multiple temperature sensors is deployed; in order to monitor and protect the air cleanliness of the laboratory in real time, multiple channel laser particle counters are deployed at key positions such as the experimental equipment area, air supply port and return air port. The counter can monitor the particle concentration of multiple key particle size channels at the same time, so as to ensure that the air cleanliness of the key area is always maintained above the preset level standard; in order to evaluate the structure vibration caused by the external environment and the operation of the HVAC system itself in real time, high-sensitivity three-axis acceleration sensors are installed on the key bearing structure of the building and the precision instrument platform. The measurement resolution is sufficient to capture the micro-vibration caused by the external environment and the internal equipment, so as to evaluate in real time whether the vibration state of the laboratory floor meets the vibration standard requirements of precision instruments; in order to collect non-periodic disturbance events that can be used for prediction, the embodiment is identified and recorded in multiple ways. The door opening and closing event is collected by the state sensor installed at all entrances and exits of the laboratory; the personnel activity event uses the space occupancy sensor to obtain real-time data of the number of personnel and the spatial position distribution under the premise of protecting privacy; the equipment start-stop event is accurately identified by the power monitoring unit according to the on-off and change of current; the embodiment deeply collects data of the HVAC system itself, in which the vibration signal is captured by installing vibration sensors on the bearing seat positions of key rotating parts such as the main air supply fan and the circulating water pump. The pressure signal is captured by installing dynamic pressure sensors at the key nodes of the main pipeline to capture the pressure transients caused by airflow pulsation or valve action.

[0044] A high-fidelity laboratory dynamic panorama is built by multi-dimensional and multi-scale data, which can not only realize real-time perception of the fine changes of the environment such as temperature field, cleanliness and micro-vibration, but also provide complete and reliable decision basis for subsequent predictive feedforward control, adaptive optimization and health diagnosis of advanced intelligent algorithms through deep capture of disturbance source events and equipment state.

[0045] Further, the control prediction model comprises: fusing the disturbance source event data and state data, and converting 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 iteratively deducing time series inside the model according to a dynamically determined time step, to generate a specific future environmental state change curve; determining a target compensation curve opposite to the target compensation according to the future environmental state change curve, and decomposing the target compensation curve into discrete target state points; for each discrete target state point, taking the target state point and the current environmental state as input, querying a pre-off-line trained neural network model for directly mapping state changes to control instructions, directly outputting specific control instruction values required to achieve the target state point in a non-iterative forward calculation manner, and combining all solved specific control instruction values in chronological order to generate the compensation control sequence, and the specific process is as shown in Figure 2

[0046] The system first fuses data from two sources: one is disturbance source event data obtained from a laboratory information management system (LIMS) through an API interface, which can be constructed as a vector containing event type (such as "liquid nitrogen filling"), key parameters (such as "50 liters"), and scheduled execution time; the other is current state data of the laboratory collected in real time from a multi-modal sensor network, which is a high-dimensional state vector containing current three-dimensional temperature field, suspended particle concentration, and structure vibration readings; the two groups of heterogeneous data are input into a multi-modal encoder network. The encoder network has parallel processing branches, for example, using embedding layers to process categorical event types, using fully connected layers to process numerical event parameters and state data, and the processing results of each branch are finally spliced and converted into a unified initial state vector, for example, 512-dimensional, which is a complete mathematical description of "what is happening at this moment".

[0047] The initial state vector is then input into a time series prediction model based on a Transformer architecture. The model can dynamically analyze the internal relationship and importance of different dimension information (such as the type of disturbance event and the current temperature gradient) in the initial state vector using its core self-attention mechanism. Before deduction begins, a small subnetwork inside the model will adaptively determine an optimal deduction time step Δt according to the characteristics of the disturbance event (such as the influence duration and intensity), for example, for liquid nitrogen filling, a fast and strong disturbance, Δt may be determined as 5 seconds; for solar radiation, a slow disturbance, Δt may be determined as 60 seconds.

[0048] ​Subsequently, the model iteratively performs forward inference on this dynamic time step: first, it predicts the state at time t+At based on the initial state, then takes the predicted state at time t+At as the new input to predict the state at time t+2At, and so on, until the entire prediction time domain (e.g., 30 minutes in the future) is covered. Finally, a decoder network converts the series of high-dimensional state vectors generated by the model into a curve of the future environmental state change with explicit physical units.

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

[0050] Finally, the system sorts and combines the multi-dimensional control vectors calculated for all discrete target state points according to their corresponding timestamps, and the final generated compensation control sequence is a structured time-sequential instruction queue containing accurate instructions to be issued to different execution mechanisms at different times in the future. This sequence is then submitted to the edge computing and execution control node for final security check and execution.

[0051] In this embodiment, the encoder network includes two parallel branches: one branch processes the categorical information such as event type in the perturbation source event data through an embedding layer; the other branch processes the numerical information such as sensor readings using a network composed of three fully connected layers with 256, 128 and 64 neurons respectively, and ReLU as the activation function. The outputs of the two branches are concatenated and finally mapped to an initial state vector of 512 dimensions through a fully connected layer. The time series prediction model based on the Transformer architecture is composed of 6 encoder layers stacked together, with the model dimension set to 512. Each encoder layer internally includes an 8-head multi-head self-attention mechanism and a feedforward network with 2048 nodes. The inverse dynamic neural network model for direct mapping is a three-layer hidden layer multilayer perceptron with 512, 256 and 128 neurons respectively in each hidden layer, using ReLU activation function, and using the Adam optimizer with a learning rate of 0.001 for offline training.

[0052] By combining forward-looking time series prediction with non-iterative inverse mapping, a control transition from passive response to active intervention is achieved, which can calculate a highly customized, time-varying compensation control sequence before the disturbance actually occurs, thereby minimizing the adjustment cost and maximizing the suppression of environmental fluctuations to ensure the ultimate stability of the laboratory environment.

[0053] Further, the process of generating the cooperative operation parameter includes: constructing a state input vector from the three-dimensional temperature field data, suspended particle data and structure vibration data; the plurality of heterogeneous strategy models include models with three different optimization objectives: energy consumption optimization model, temperature field optimization model and particle deposition model; the plurality of heterogeneous strategy models are independently trained based on different data subsets generated from all historical data by the bootstrap sampling method; the state input vector is simultaneously input into the plurality of heterogeneous strategy models for parallel forward inference, and during the forward inference, parameter suggestion values and quantitative uncertainty indicators representing the confidence of the current suggestion values of the models are output; the plurality of parameter suggestion values output by the plurality of heterogeneous strategy models are integrated by a dynamic weighting fusion strategy based on the inverse of uncertainty, the fusion weights corresponding to each model are calculated according to the quantitative uncertainty indicators output by each model, and the fusion weight of any model is inversely proportional to the quantitative uncertainty indicator output by the model, and the plurality of parameter suggestion values are weighted and summed using the fusion weights to generate the cooperative operation parameter.

[0054] The plurality of heterogeneous strategy 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 total instantaneous power of the HVAC system, and its training data labels are more focused on the electrical energy consumption of the system, so it tends to output control parameters that prioritize energy saving. In this embodiment, it adopts a lightweight multi-layer 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, and its training data labels are more focused on the variance of the readings of the 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 an image.

[0057] Particle settling model: The goal of this model is to maximize the removal efficiency of suspended particles in the air, and its training data labels are more focused on the rate of change of readings of the key position laser particle counter. It adopts a neural network architecture containing recurrent layers (RNN) to better capture the time series dynamics of air flow and particle motion.

[0058] To ensure the heterogeneity of the models, not only are the neural network architectures different, but the training data sets of each model are also different, based on the sub-sets of data generated from the entire historical data through the bootstrap sampling method.

[0059] To achieve confidence-based decision fusion, each expert model must quantify its own uncertainty about the suggestion of the output parameter value at the same time. This embodiment uses the Monte Carlo Dropout technique to achieve this; during the real-time forward inference phase, when a state input vector is fed into a certain expert model, the model does not perform only one calculation. In fact, it will perform N (for example, N = 50) independent forward calculations continuously in the state of activated Dropout (i.e., randomly inactivating part of the neurons); this will result in a distribution of N slightly different parameter suggestion values. The mean of this distribution is taken as the final parameter suggestion value of the model; the variance or standard deviation of the distribution is taken as the quantified uncertainty indicator, and the larger the variance, the greater the divergence of the model, i.e., the higher the uncertainty; after obtaining the parameter suggestion values of all three expert models and their corresponding uncertainty indicators, the system begins to perform the final fusion decision.

[0060] In a specific fusion strategy, the system adopts a dynamic weighting method based on the confidence of each model. The core of this method is to evaluate the quantitative uncertainty index of each heterogeneous strategy model when outputting its recommended value in real time. For models with high uncertainty index values (i.e., insufficient confidence in the current recommendation), the system dynamically assigns them a lower fusion weight. Conversely, models with lower uncertainty index values and more sufficient confidence will obtain a relatively higher weight, thus having more power in the final decision. The final collaborative operation mode parameter is obtained by weighting and summing the parameter recommendations output by each model and the dynamically calculated weights.

[0061] In a specific embodiment, the structure parameters of the plurality of heterogeneous strategy models are as follows: the temperature field optimization model adopts a convolutional neural network (CNN) architecture containing three layers of three-dimensional convolution (Conv3D); the input three-dimensional temperature field data is processed into a 32x32x8 tensor, the first convolutional layer uses 16 5x5x3 convolutional kernels with a step size of 2; the second and third layers both use 32 3x3x3 convolutional kernels with a step size of 1; a ReLU activation function and a batch normalization layer are connected after all convolutional layers; the particle settling model adopts a stacked double-layer long short-term memory network (LSTM) with a hidden state dimension of 128 to effectively capture the time series dynamics of the suspended particle concentration changes; the energy consumption optimization model adopts a relatively simple four-layer fully connected network with neuron counts of 256, 128, 64, and 32.

[0062] Further, before integrating the parameter values output by the plurality of heterogeneous strategy models, the intelligent decision module also performs: providing the parameter adjustment recommendation output by one of the heterogeneous strategy models as input to all other heterogeneous strategy models to predict the estimated impact of implementing the parameter adjustment recommendation on other optimization objectives; and according to whether any index in the estimated impact exceeds a preset catastrophe threshold, vetoing the parameter adjustment recommendation before integration.

[0063] Before the parameter values output by multiple heterogeneous strategy models are integrated, the application introduces a suggestion-impact pre-evaluation mechanism based on inter-model cross-validation to prevent a single model from outputting potentially destructive control suggestions in extreme cases; specifically, when a temperature field optimization model outputs a large-amplitude control parameter adjustment suggestion, the parameter adjustment suggestion is not immediately weighted and fused; instead, it is first stored by the system and used as input to query the energy consumption optimization model and the particle deposition model in reverse to predict "what impact will this adjustment have on the total energy consumption of the system and the concentration of suspended particles if it is executed?", if any other model predicts that its core indicators will deteriorate beyond a preset disaster threshold, the system will directly reject the initial adjustment suggestion and trigger a re-planning process based on the current state; this mechanism establishes a rigid balance between models, adding a "rational review" link to the entire collaborative decision-making process, ensuring that the final integrated output takes into account the bottom-line safety of all targets.

[0064] Through the integrated decision engine composed of multiple specialized models and the dynamic fusion strategy based on model self-uncertainty perception, the collaborative optimization of multiple conflicting targets is realized, and the mechanism can automatically adjust the weights of each professional target according to the specific situation, ensuring the robustness of the decision while keeping the control focus consistent with the current most important task of the laboratory.

[0065] Further, the process of processing the vibration signal and the pressure signal to extract the vibration spectrum feature and the pressure fluctuation mode data includes: performing fast Fourier transform on the vibration signal to obtain a vibration spectrum, extracting the amplitude of the fundamental frequency and each harmonic related to the rotating component speed of the heating, ventilation and air conditioning system and the energy value at the preset bearing fault characteristic frequency from the vibration spectrum, and filling them into a fixed-length feature vector according to a pre-set dimension order as the vibration spectrum feature; calculating the root mean square value, kurtosis and skewness of the pressure signal in a sliding time window as statistical characteristics, performing continuous wavelet transform to extract the time-frequency energy distribution of pressure transient events, and splicing the data obtained by vectorizing the statistical characteristics and the time-frequency energy distribution to form the pressure fluctuation mode data representing the pressure fluctuation mode.

[0066] Specifically, the original, high-frequency time series signals collected from the multi-modal sensing module are converted into structured, fixed-length feature vectors for subsequent AI health baseline model analysis. The internal processing flow of this unit is divided into two parallel parts for vibration signals and pressure signals; for vibration signal processing, the system first obtains high-frequency sampled original time series signals from the acceleration sensor installed on the key rotating components (such as the main circulating water pump bearing seat), the signal is segmented into data segments with fixed length, and after applying a window function (such as the Hanning window) to reduce spectral leakage, it is converted from the time domain to the frequency domain by performing a Fast Fourier Transform (FFT) to obtain a high-resolution vibration spectrum, then the feature extraction step accurately extracts two types of key information from the spectrum: one is the amplitude of the fundamental frequency and its harmonics corresponding to the device speed; the other is the energy value at a plurality of specific fault characteristic frequencies calculated in advance according to the device bearing model. Finally, these extracted harmonic amplitudes and fault energy values are filled into a fixed-length feature vector according to a pre-set dimension order, which is the vibration spectrum feature that can fully represent the device health status; for pressure signal processing, the system uses a double analysis path to capture its complex dynamics, first, in a sliding time window, the original signal from the dynamic pressure sensor in the main pipeline is statistically analyzed to calculate its root mean square, kurtosis, and skewness, etc. indicators to represent the macroscopic form of pressure fluctuations; to capture the pressure transients caused by valve switching and other non-stationary events, 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 transient events in time and frequency dimensions, finally, the 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 can fully describe the pressure dynamics in the pipeline as the final pressure fluctuation pattern.

[0067] By combining frequency domain and time-frequency domain analysis, structured and highly sensitive feature vectors for early weak faults can be extracted from high-noise original sensor signals; not only providing high-quality, high-information-density inputs for subsequent AI model accurate diagnosis, but also significantly improving the entire system's ability to identify potential physical risks from massive dynamic data.

[0068] Further, the AI health baseline model includes:

[0069] The input embedding layer converts 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 determination layer is responsible for quantifying the degree of abnormality of the system by calculating the root mean square error between the actual input feature vector and the health prediction vector output by the Transformer encoding layer, based on the output of the encoding layer, to generate a real-time anomaly score. At the same time, the system will calculate a dynamically changing health state threshold based on the statistical distribution (e.g., mean and three standard deviations) of all anomaly scores confirmed to be in a healthy state within a certain period of time (e.g., 24 hours). When the real-time calculated anomaly score significantly exceeds this dynamic threshold, the system determines that the HVAC system has an abnormality.

[0076] When the anomaly determination layer issues an alarm, the root cause diagnosis layer is activated, aiming to achieve explainable AI and accurately locate the problem source. In this embodiment, a SHAP algorithm is used to analyze and calculate the contribution of each dimension (i.e., each frequency bin, each statistical feature) in the input feature vector to the output result that leads to a large prediction deviation (high anomaly score).

[0077] After the calculation is completed, the system obtains a contribution score vector with the same dimension as the input vector, and identifies the key features with the highest contribution scores (e.g., the 3rd harmonic amplitude of the fundamental frequency and the kurtosis of the pressure signal). Finally, the system matches this key feature combination with the patterns pre-set in the physical fault feature signature library (e.g., the combination pattern is highly consistent with the fault signature of a slightly unbalanced fan blade), and outputs the final diagnosis information.

[0078] The physical fault feature signature library is preferably implemented as a JSON format rule database that can be configured by operation and maintenance personnel. Each entry in the library corresponds to a known physical fault, and the entry includes a unique fault ID, a fault name description, and one or more "feature-condition-threshold" triplets. When the calculated key contribution features simultaneously satisfy all the matching rules under a certain fault entry, the system determines that the specific fault has occurred. For example, the fault signature entry of "slightly unbalanced fan blade" can be defined as: {“fault_id”:“F001”,“fault_name”:“slightly unbalanced fan blade”,“rules”:[{“feature”:“vibration frequency spectrum-3rd harmonic amplitude of fundamental frequency”,“condition”:“>”,“threshold”:0.5}, {“feature”:“pressure signal-kurtosis”,“condition”:“in_range”,“threshold”:[3.5,4.5]}]} This structured definition makes the basis of fault diagnosis clear and easy to extend and maintain in the future.

[0079] Through self-supervised learning of the Transformer, the health paradigm of the equipment can be deeply mastered without a large number of fault samples; the dynamic threshold can effectively adapt to the working condition changes and reduce false positives; the SHAP algorithm is used to transparentize the AI decision-making process, which can quantify the fault and trace it back to the specific physical feature source, providing a clear and reliable decision basis for predictive maintenance, and realizing the upgrade from passive alarm to active diagnosis.

[0080] Further, the process of driving the actuator of the heating and air conditioning system includes: fusing the compensation control sequence, the cooperative operation mode parameter and the diagnosis information to form a control intention, the compensation control sequence serving as a feedforward regulation reference, the cooperative operation mode parameter serving as a feedback optimization target, and the diagnosis information serving as a dynamic constraint condition; when the diagnosis information indicates that a component has a health risk, dynamically adjusting the fusion strategy to reduce the operation load of the component; querying a control instruction mapping table to analyze the control intention into a specific physical control instruction sequence for each actuator in the heating and air conditioning system; performing item-by-item safety checking on the amplitude, change rate and combination logic of the physical control instruction sequence according to preset device physical limits, process safety requirements and energy consumption limitation rules; for instructions that exceed the safety boundary, adjusting the instruction parameters that exceed the single-item numerical limit to the safety boundary value, and rechecking the logical safety of the entire instruction combination; if the checking is passed, issuing the corrected instruction, and if the checking is not passed, intercepting the issuance of the instruction and triggering an alarm.

[0081] Specifically, the node receives three types of data packets from the cloud platform in each control cycle, including the compensation control sequence as a feedforward regulation reference, the cooperative operation mode parameter as a feedback optimization target, and the diagnosis information reporting the system health status; in the initial health state, the node takes the compensation control sequence as the basis of the open-loop control plan, and runs a local PID controller to perform closed-loop fine adjustment with the cooperative operation mode parameter as the real-time set point, so as to execute long-term planning while responding to real-time minor changes.

[0082] The core of this process is that it can dynamically adjust the fusion strategy according to the diagnosis information, for example, when the node receives the diagnosis information: {component ID: VAV-03, fault mode: wind valve actuator response hysteresis, risk level: medium, suggestion: avoid frequent or small-range adjustment}, the fusion strategy will be dynamically adjusted, the system will reduce the weight of the fine adjustment of the VAV-03 air volume in the cooperative operation mode parameter, 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 sacrifice local optimization and energy efficiency to ensure the overall environmental stability and operation safety.

[0083] The fused "control intent" (e.g. {AHU supply air temperature: 18.5℃, VAV-03 damper opening: 45%,...}) needs to be parsed into physical commands understandable by the actuators, which is done by querying a control command mapping table, a pre-configured database containing the conversion relationship from physical targets to specific control signals, e.g. "AHU supply air temperature 18.5℃" is parsed into "chilled water regulating valve opening signal: 5.2V"; "VAV-03 damper opening 45%" is parsed into "damper actuator pulse signal: send 90 opening pulses".

[0084] At the last moment before the commands are issued, the system will perform a rigorous item-by-item safety check. The first layer of checks will check whether the values and rates of change of each command in the sequence exceed the physical limits of the equipment. After all individual values are corrected to be within the safe range, the second layer of checks will check whether there is a logical conflict in the entire command combination, such as the error command of maximum cooling and heating at the same time. Only after both layers of checks pass, the corrected command sequence will be finally issued to the actuators. If the second layer of logical checks fails, the command will be intercepted and a high-level alarm will be triggered to ensure the absolute safety of the system.

[0085] The driving actuators in the HVAC system further comprises: an actuator physical model based on the physical response time, action energy consumption curve and system thermal inertia parameters of each actuator; receiving a physical control command sequence and considering the physical control command sequence as a multi-target state in a future time window; solving the optimal execution path problem with the optimization goal of reaching the target state in the shortest time and the lowest total energy consumption, and performing timing rearrangement and issue time optimization on each command in the physical control command sequence to generate an optimized execution queue; driving the actuators in the HVAC system according to the order and time stamp defined in the execution queue.

[0086] Specifically, in order to solve the difference between the ideal control instruction and the dynamic characteristics (such as delay, inertia) of the physical world execution mechanism, and further reduce the instantaneous energy consumption and mechanical wear and tear of the system in the adjustment process, the execution control module integrates a control instruction timing optimization engine based on a physical model before finally driving the execution mechanism; The engine has a physical model containing each valve, air valve and other execution mechanisms, which accurately defines its complete action response time (for example, VAV-01 air valve needs 45 seconds from 0% to 100%), the action energy consumption curve at different opening degrees, and the system thermal inertia caused by its action (for example, after the cold water valve reaches the target opening, it takes 90 seconds for the supply air temperature to stabilize). When the engine receives a set of target physical control instructions generated by the upper decision-making module (for example, "in 10 seconds, set VAV-01 opening to 80% and main fan frequency to 45Hz"), it does not immediately issue. Instead, by solving the constrained programming problem with the optimization goal of minimizing the total energy consumption of execution, the timing rearrangement and path optimization of the instruction set are performed. For example, the engine calculates and finds that if "first increase the main fan frequency to 45Hz, wait for 2 seconds for the pipeline static pressure to stabilize, and then drive the VAV-01 air valve to open", the overall instantaneous power consumption impact and mechanical stress will be much smaller than executing the two actions simultaneously; Therefore, the engine generates an optimized execution queue with precise timestamps, and strictly follows the timing defined by the queue to issue instructions one by one and smoothly to the physical execution mechanism.

[0087] By optimizing the physical execution timing of the instructions, the instantaneous power consumption impact and equipment mechanical wear and tear can be significantly reduced. It avoids the action conflict between actuators, makes the system adjustment process smoother and more stable, thereby effectively prolonging the equipment life and improving the overall operation efficiency of the system.

[0088] Further, the safety boundary rule is a configurable rule library, which dynamically loads the safety rule configuration file in JSON format corresponding to the current laboratory working situation according to the laboratory current working situation obtained from the laboratory information management system, and adjusts the specific limit value and logic of the verification.

[0089] The safety boundary rule is not a fixed set of global rules, but a dynamically associated and configurable rule library with the current working situation of the laboratory, specifically, the system obtains the current state of the laboratory in real time through the interface with the laboratory information management system (LIMS), such as "Standby_Mode" and "High_Sensitivity_Optical_Experiment_Mode", each mode corresponds to an independent safety rule configuration file encoded in JSON (JavaScript Object Notation) format; when the system enters the "High_Sensitivity_Optical_Experiment_Mode" mode, the corresponding JSON configuration file will be loaded, which clearly defines more stringent safety 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 safety boundary makes the safety protection capability of the application no longer fixed, but can be accurately and coded self-adaptive adjustment according to the actual needs of the experimental process, realizing fine protection of core scientific research tasks.

[0090] By fusing feedforward, feedback and diagnostic information, a resilient control mode with forward-looking planning and real-time fine-tuning capability is constructed, which can dynamically adjust the strategy when the component is abnormal to ensure the stability of the system, and the final output to the actuator is ensured to have safety and logical rationality by the multi-level verification mechanism.

[0091] By fusing prediction, optimization and diagnosis, a control strategy is provided for high-precision laboratory environment control, which provides significant comprehensive improvement. It introduces a feedforward compensation control based on model prediction, which can actively predict and intervene the substantial impact of disturbances on the environment, thereby achieving higher environmental stability than traditional feedback control. At the same time, the system makes collaborative decisions through multiple heterogeneous models that focus on different targets such as energy consumption, temperature field, cleanliness, etc., and uses an uncertainty-aware dynamic fusion strategy to make more robust and reasonable control decisions, improving the operation economy. In addition, the system also has weak fault diagnosis capability based on AI, which can identify early performance degradation of equipment online, and guide predictive maintenance through accurate root cause positioning, effectively improving the long-term operation reliability of the system.

[0092] Embodiment two:

[0093] In order to improve the automatic air conditioning capability in the constant temperature and humidity intelligent physical laboratory, an intelligent laboratory air quality control method is introduced, and the specific process is as shown in Figure 4 .

[0094] First, the three-dimensional temperature and humidity field data is collected in the laboratory grid, and the local micro-environment data is collected around the key equipment such as the core optical platform and the laser through high-precision composite sensors; at the same time, the structure vibration, air suspended particle concentration, and vibration and pressure signals of the fan and water pump of the main air handling unit during operation are collected; through the program interface, the method actively obtains the scheduled experiment plan from the laboratory information management system (LIMS), and parses the operation such as "turn on the high-power laser" into future disturbance source event data carrying specific time and power parameters.

[0095] Next, after receiving the disturbance event data from LIMS, the method combines the data with the current laboratory state data and inputs it into a time series prediction model based on the Transformer architecture; the model will prospectively deduce the specific influence curve of the disturbance event on indoor temperature and humidity in the future; then, the method generates a time-sequenced compensation control sequence through a pre-trained reverse dynamic neural network model according to the prediction curve; the sequence can instruct the heating, ventilation and air conditioning system to intervene in advance and gently adjust before the disturbance actually occurs, thereby actively offsetting the upcoming environmental fluctuations.

[0096] At the same time, the real-time collected laboratory state data forms an input vector, which is simultaneously input into multiple heterogeneous strategy models trained with different optimization objectives (such as minimum energy consumption, most stable temperature and humidity, and highest cleanliness); through a dynamic weighting strategy based on uncertainty perception, the method integrates the recommended values output by multiple models to generate a set of collaborative operation parameters; the parameters can automatically make the most reasonable dynamic trade-off between energy saving and environmental stability and other conflicting goals according to whether the laboratory is in standby or experimental state.

[0097] Then, the vibration and pressure signals collected from the heating, ventilation and air conditioning equipment are continuously processed, and through fast Fourier transform, continuous wavelet transform and other means, a feature vector that can represent the weak performance changes of the equipment is extracted; then, the method analyzes the feature vector through an AI health baseline model; the model has mastered the norm of the equipment in a healthy state through self-supervised learning, so it can generate an anomaly score by calculating the deviation of the current features from the healthy norm, and output diagnostic information containing specific physical root cause positioning when an anomaly occurs, such as "early fouling of a certain transducer of the humidifier".

[0098] Finally, the aforementioned generated compensation control sequence (feedforward), cooperative operation parameters (feedback), and diagnostic information (constraint) are fused to form the final control intent; for example, the diagnostic information will make the method actively reduce the operating load of the faulty component during control; then, the method parses the control intent into physical control instructions for each actuator by querying the mapping table; before the instructions are issued, a safety boundary rule module will strictly check the amplitude, rate of change, and combination logic of the instructions, and only after ensuring absolute safety will the final instructions be issued to drive the actuators in the HVAC to complete accurate air quality regulation.

[0099] While embodiments of the present application have been shown and described, it is to be understood that the embodiments described are merely divergences of the principles and spirit of the present application and that numerous modifications, changes, substitutions, and alterations can be made thereto without departing from the spirit and scope of the present application as defined by the appended claims and their equivalents.

Claims

1. An intelligent laboratory air quality control system, characterized in that, The application relates to a physical laboratory intelligent control system. The data acquisition module acquires state data of the physical laboratory, including three-dimensional temperature field data, suspended particle data and structural vibration data; acquires disturbance source event data; acquires vibration signals and pressure signals of the HVAC; The intelligent decision-making module inputs the state data and the disturbance source event data into a control prediction model for prediction, generates a compensation control sequence; forms an input vector from the state data; simultaneously inputs the input vector into multiple heterogeneous strategy models for forward reasoning, integrates parameter values output by the multiple heterogeneous strategy models, and generates cooperative operation parameters; The fault diagnosis module processes the vibration signals and the pressure signals, extracts vibration frequency spectrum features and pressure fluctuation mode data; analyzes the vibration frequency spectrum features and the pressure fluctuation mode data through an AI health degree baseline model, and outputs diagnosis information; The execution control module fuses the compensation control sequence, the cooperative operation parameters and the diagnosis information, generates physical control instructions, checks the physical control instructions through a safety boundary rule, and drives an execution mechanism in the HVAC.

2. The intelligent laboratory air quality control system of claim 1, wherein, The data acquisition module acquires data in the following process: a temperature sensor array is deployed in a three-dimensional space coordinate system in the laboratory to acquire three-dimensional temperature field data; laser particle counters are deployed in experimental equipment areas, air supply outlets and air return outlets to acquire suspended particle data; acceleration sensors are installed on building bearing structures, instrument platforms and HVAC unit bases to acquire structural vibration data; disturbance source event data such as door opening and closing, personnel walking and equipment starting and stopping are identified and recorded; vibration sensors and pressure sensors are installed on the HVAC to acquire vibration signals and pressure signals.

3. The intelligent laboratory air quality control system of claim 1, wherein, The control prediction model includes: the disturbance source event data and the state data are fused and converted into an initial state vector through an encoder network; the initial state vector is input into a time series prediction model based on a Transformer architecture, and iterative time series deduction is performed inside the model according to dynamically determined time steps to generate a specific future environmental state change curve; according to the future environmental state change curve, a target compensation curve opposite to the target compensation is determined, and the target compensation curve is decomposed into discrete target state points; for each discrete target state point, the target state point and the current environmental state are taken as inputs to query a pre-off-line trained neural network model for directly mapping state changes to control instructions, so as to directly output specific control instruction values required to achieve the target state point in a non-iterative forward calculation mode; all solved specific control instruction values are combined in chronological order to generate the compensation control sequence.

4. The intelligent laboratory air quality control system of claim 1, wherein, The process of generating the cooperative operation parameter comprises: constructing a state input vector from the three-dimensional temperature field data, the suspended particle data and the structure vibration data; the plurality of heterogeneous strategy models comprise energy consumption optimization models, temperature field optimization models and particle deposition models for three different optimization objectives; the plurality of 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 plurality of heterogeneous strategy models for parallel forward inference, and a parameter suggestion value and a quantitative uncertainty index for representing the confidence of the model to the current suggestion value are output during the forward inference; the plurality of parameter suggestion values output by the plurality of heterogeneous strategy models are integrated through a dynamic weighted fusion strategy based on the inverse of the uncertainty, 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 plurality of parameter suggestion values are weighted and summed using the fusion weight to generate the cooperative operation parameter.

5. The intelligent laboratory air quality control system of claim 1, wherein, The process of processing the vibration signal and the pressure signal to extract the vibration spectrum feature and the pressure fluctuation mode data comprises: performing fast Fourier transform on the vibration signal to obtain a vibration spectrum, extracting the amplitude of a fundamental frequency and each harmonic frequency related to the rotating component speed of the heating ventilation air conditioner and the energy value at a preset bearing fault characteristic frequency from the vibration spectrum, and filling the extracted data into a fixed-length feature vector in a preset dimension order as the vibration spectrum feature; calculating the root mean square value, kurtosis and skewness of the pressure signal in a sliding time window as statistical characteristics, performing continuous wavelet transform to extract the time-frequency energy distribution of the pressure transient event, and splicing the data obtained by vectorizing the statistical characteristics and the time-frequency energy distribution to form the pressure fluctuation mode data representing the pressure fluctuation mode.

6. The intelligent laboratory air quality control system of claim 1, wherein, The AI health degree baseline model comprises: an input embedding layer for converting the extracted vibration spectrum feature and pressure fluctuation mode data into a high-dimensional feature vector; a Transformer encoding layer for receiving the high-dimensional feature vector, modeling the internal correlation between each feature dimension through an internal multi-head self-attention mechanism, and outputting a health prediction vector for the current state; an anomaly determination layer for generating an anomaly score by calculating the deviation between the health prediction vector and the actually input feature vector, and determining that the system is abnormal when the anomaly score exceeds a dynamic threshold; a root cause diagnosis layer activated when an anomaly is determined, calculating the contribution score of each input feature to the prediction deviation by analyzing the attention weight and model gradient information in the Transformer encoding layer, identifying the key feature with the highest contribution score, and matching the key feature with a preset physical fault characteristic signature library to output diagnosis information containing specific physical root cause positioning.

7. The intelligent laboratory air quality control system of claim 1, wherein, The process of driving the actuators of the HVAC system comprises: fusing the compensation control sequence, the cooperative operation mode parameters and the diagnostic information to form a control intention, the compensation control sequence serving as a feedforward regulation reference, the cooperative operation mode parameters serving as a feedback optimization target, and the diagnostic information serving as a dynamic constraint condition; when the diagnostic information indicates that a component has a health risk, dynamically adjusting the fusion strategy to reduce the operation load of the component; querying a control instruction mapping table to analyze the control intention into a specific physical control instruction sequence for each actuator in the HVAC system; performing item-by-item safety checking on the amplitude, the change rate and the combination logic of the physical control instruction sequence according to preset device physical limits, process safety requirements and energy consumption limitation rules; for an instruction exceeding the safety boundary, adjusting the instruction parameter exceeding the single-item numerical limit to the safety boundary value and rechecking the logical safety of the entire instruction combination; if the checking is passed, issuing the corrected instruction; if the checking is not passed, intercepting the issuance of the instruction and triggering an alarm.

8. An intelligent laboratory air quality control method, characterized in that: state data of a physical laboratory are collected, the state data comprising three-dimensional temperature field data, suspended particle data and structural vibration data; disturbance source event data are obtained; vibration signals and pressure signals of the HVAC system are obtained; the state data and the disturbance source event data are input into a control prediction model for prediction to generate a compensation control sequence; the state data are formed into an input vector; the input vector is simultaneously input into a plurality of heterogeneous strategy models for forward reasoning, and the parameter values output by the plurality of heterogeneous strategy models are integrated to generate cooperative operation parameters; the vibration signals and the pressure signals are processed to extract vibration frequency spectrum features and pressure fluctuation mode data; the vibration frequency spectrum features and the pressure fluctuation mode data are analyzed by an AI health degree baseline model to output diagnostic information; the compensation control sequence, the cooperative operation parameters and the diagnostic information are fused to generate physical control instructions, and the physical control instructions are checked by a safety boundary rule to drive the actuators in the HVAC system.

Images