Intelligent control method and system of biological safety cabinet

Abstract

The invention provides an intelligent control method and system for a biological safety cabinet, and relates to the technical field of safety cabinet control, and the method comprises the steps: obtaining the real-time data of an air pressure difference, a wind speed and a particulate matter concentration ratio collected by an edge calculation node, and monitoring a closed device through a sensor array; performing feature extraction on the real-time data by adopting a first long-short term memory network algorithm to form a dynamic feature set; if the signal-to-noise ratio of the voice signal after noise reduction is higher than a preset threshold value, instruction analysis is executed, and if the signal-to-noise ratio is lower than the threshold value, user control intention data are obtained; according to the user control intention data, performing biological feature authentication on the operator by adopting a face recognition algorithm, and if the matching degree is higher than a preset threshold value, generating an equipment control authority; and generating a configuration updating instruction of the closed equipment according to the real-time state mapping data and the optimized operation parameters, and executing equipment linkage control. According to the invention, intelligent monitoring and adaptive adjustment of the biosafety cabinet are realized, and the safety and efficiency of equipment operation are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of safety cabinet control, and in particular to an intelligent control method and system for a biological safety cabinet. Background Art

[0002] As a core piece of laboratory biosafety equipment, biosafety cabinets (BSCs) play a crucial role in preventing pathogen leakage and ensuring operator safety. They isolate hazardous substances through negative pressure airflow and high-efficiency filtration systems and are widely used in fields such as microbiology experiments and pathogen research. However, existing BSC control methods have significant limitations, making them incapable of meeting the demands of modern laboratories for intelligent and remote management.

[0003] Traditional systems rely on manual operation of UV lamps, fans, and other equipment, resulting in operational lags and potential safety hazards. This is particularly true in unattended or emergency situations, making timely response and risk control difficult. Furthermore, existing equipment lacks sophisticated environmental perception and multimodal interaction capabilities, making it unable to adapt to dynamic laboratory changes, resulting in low safety and efficiency. Key challenges in this area lie in the real-time nature of remote control, the robustness of multimodal interaction, and the intelligent integration of equipment. First, the real-time nature of remote status monitoring and control is limited by the efficiency of data collection and transmission, making it difficult to dynamically map the operating status of the biosafety cabinet and rapidly respond. Second, voice interaction and biometric authentication are susceptible to noise interference in complex laboratory environments, requiring improvements in noise immunity and authentication accuracy. Finally, the coordinated control logic between UV lamps, fans, and other equipment is relatively simple, making it difficult to intelligently adjust operating modes based on environmental changes. These technical challenges result in insufficient safety assurance for biosafety cabinets in unattended scenarios, limiting their application in high-risk experiments.

[0004] Therefore, how to achieve real-time monitoring and remote control of the safe's operating status through edge computing and cloud synchronization, while combining noise-resistant multimodal interaction and intelligent linkage strategies to improve the accuracy and safety of equipment operation, has become a key issue that needs to be urgently addressed in this study. Summary of the Invention

[0005] The present invention provides an intelligent control method for a biosafety cabinet, which mainly includes: Acquire real-time data on air pressure difference, wind speed, and particle concentration ratio collected by edge computing nodes, and use sensor arrays to monitor negative pressure airflow parameters in enclosed equipment; Using a first long short-term memory network algorithm to extract features from the real-time data, generating an abnormality alarm if the extracted feature value deviates from a preset threshold, and forming a dynamic feature set; Based on the dynamic feature set, the data is transmitted to a remote server through a cloud synchronization mechanism, and a second long short-term memory network algorithm is used to perform time series trend prediction to determine the stability trend data of the airflow and filtration system; Extracting real-time state mapping data from the stability trend data, combining it with a voice interaction interface to receive user instructions, and using an adaptive noise suppression algorithm to perform noise reduction on the instructions; If the signal-to-noise ratio of the denoised voice signal is higher than a preset threshold, command parsing is performed; if it is lower than the preset threshold, a repeat prompt is generated to obtain user control intention data; Based on the user control intention data, a face recognition algorithm is used to perform biometric authentication on the operator, and if the matching degree is higher than a preset threshold, device control permission is generated; Based on the device control authority, a reinforcement learning algorithm is used to construct a reward function with the goals of minimizing energy consumption and maximizing the particle concentration ratio, and the operating parameters of the environmental control device are dynamically adjusted; According to the real-time status mapping data and the optimized operating parameters, a configuration update instruction of the sealed equipment is generated and equipment linkage control is executed.

[0006] The technical solution provided by the embodiment of the present invention may have the following beneficial effects: The present invention discloses an intelligent biosafety cabinet control system, which collects real-time data such as air pressure difference, wind speed, and particle concentration through edge computing nodes, and uses a sensor array to monitor the negative pressure airflow parameters of the closed equipment. A long short-term memory network algorithm is used for feature extraction and time series trend prediction, and voice interaction and face recognition are combined to achieve user identity authentication and command parsing. A reward function is constructed based on a reinforcement learning algorithm, and the operating parameters of the biosafety cabinet are dynamically optimized with the goals of minimizing energy consumption and maximizing the particle concentration ratio. The present invention realizes intelligent monitoring and adaptive adjustment of the biosafety cabinet, improves the safety and efficiency of equipment operation, and provides a more reliable environmental control solution for biological laboratories. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] For better understanding and implementation, the technical solution of the present application is described in detail below with reference to the accompanying drawings.

[0008] Figure 1 This is a flow chart of an intelligent control method for a biological safety cabinet of the present invention; Figure 2 A schematic diagram of an intelligent control method for a biological safety cabinet according to the present invention; Figure 3 is another schematic diagram of an intelligent control method for a biological safety cabinet of the present invention; Figure 4 This is a structural diagram of an intelligent control system for a biological safety cabinet according to the present invention. DETAILED DESCRIPTION

[0009] To further illustrate the technical means and effectiveness of the present invention in achieving its intended purpose, exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. In the following description, when referring to the drawings, identical numerals in different figures represent identical or similar elements, unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all possible embodiments consistent with the present application. Rather, they are merely examples of methods and systems consistent with certain aspects of the present application, as detailed in the appended claims.

[0010] The terms used in this application are for the purpose of describing specific embodiments only and are not intended to limit this application. The singular forms "a," "an," "the," and "the" used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to any or all possible combinations of one or more of the associated listed items.

[0011] The following describes in detail the specific implementation methods, features and effects of the present invention in conjunction with the accompanying drawings and preferred embodiments.

[0012] Example 1 like Figure 1-3 The intelligent control method of a biological safety cabinet in this embodiment may specifically include: Step S101: obtain real-time data of air pressure difference, wind speed and particle concentration ratio collected by the edge computing node, and use a sensor array to monitor the negative pressure airflow parameters of the closed equipment.

[0013] The air pressure difference, wind speed value, and particle concentration data of key monitoring points of the closed equipment are obtained to form a multidimensional parameter set; the multidimensional parameter set is preprocessed, and the sensor noise is eliminated by the Kalman filter algorithm to obtain smoothed airflow parameters; the air pressure difference is calculated based on the smoothed airflow parameters, and the air pressure difference is compared with the preset threshold to determine whether the negative pressure value of the equipment is within the safe operating range; if the air pressure difference is lower than the preset threshold, an abnormal event is recorded; the wind speed values collected at the monitoring points are weighted averaged to obtain fused wind speed data; the airflow direction diagram inside the equipment is drawn based on the fused wind speed data to identify airflow blind spots; the particle concentration data of the key monitoring points are time-series analyzed and the LSTM (Long Short-Term A prediction model is established using a neural network with long-short-term memory (LSTM) technology. Historical particle concentration data is used to train the prediction model to obtain short-term particle concentration trends. The airflow parameters are transmitted to the cloud in a hierarchical manner according to real-time requirements, with high-priority abnormal data being transmitted first. Correlation analysis is performed between the airflow parameters stored in the cloud and equipment status indicators, and a mapping relationship between parameter anomalies and equipment failures is established using a decision tree algorithm. Feature selection methods are used to determine key parameters and identify potential risk points.

[0014] It should be noted that the above-mentioned enclosed equipment refers to the enclosed equipment inside the biosafety cabinet.

[0015] For example, in the airflow monitoring and analysis system in closed equipment, the acquisition of multi-dimensional parameter sets is a basic link.

[0016] For example, in a biosafety cabinet, differential pressure sensors, hot wire anemometers, and particle counters can be placed at key points such as the work area, exhaust vents, and air inlets to collect data on air pressure difference, wind speed, and particle concentration.

[0017] Specifically, the raw data collected by sensors often contains noise, and the Kalman filtering algorithm can effectively eliminate these interferences.

[0018] For example, the data measured by the wind speed sensor may fluctuate at 0.45±0.05m / s. After Kalman filtering, a smooth stable value of 0.45m / s can be obtained, improving data reliability.

[0019] In one embodiment, the system compares the processed air pressure difference with a preset safety threshold. If the measured value is lower than the threshold, it is recorded as an abnormal event of insufficient negative pressure, providing a timely warning of possible safety risks.

[0020] Preferably, weighted averaging is performed on the multi-point wind speed data, and different weights can be assigned according to the importance of the monitoring points.

[0021] For example, the weight of the measuring point in the working area is 0.5, and that in other areas is 0.25. The obtained fused wind speed data can better reflect the overall airflow state. The airflow direction diagram drawn based on this can intuitively display the blind spot area, such as the area in the lower right corner where the wind speed is less than 0.2m / s.

[0022] It should be noted that particle concentration prediction uses an LSTM network, which effectively captures time series characteristics. By inputting concentration data from the past 24 hours, the model can predict concentration trends over the next four hours, allowing for early detection of abnormal increases.

[0023] In one possible implementation, airflow parameters are transmitted in real-time hierarchical order, such as abnormal negative pressure data (priority 1), abnormal particle concentration (priority 2), and normal parameters (priority 3), ensuring that critical information is processed first.

[0024] It is understandable that the cloud analyzes historical data through a decision tree algorithm to establish a mapping relationship between "negative pressure below -15Pa and particle concentration exceeding 200 / m³" and "filter blockage", identify risk points, and implement preventive maintenance.

[0025] In step S102, a first long short-term memory network algorithm is used to extract features from the real-time data. If the extracted feature value deviates from a preset threshold, an abnormality alarm is generated to form a dynamic feature set.

[0026] Real-time time series data is acquired, and the real-time time series data is segmented using a sliding window method, where the window size of the sliding window is a preset value, and the sliding step size of the sliding window is a preset value, to obtain a time series segment set; for the time series segment set, a deep learning network is used to extract features to obtain a feature value set, where the deep learning network is a recurrent neural network, and the dimension of the feature value set is a preset dimension; a threshold judgment is performed on each feature value in the feature value set, and if the feature value exceeds the preset threshold, an abnormal alarm is generated to obtain an abnormal alarm record set; based on the abnormal alarm record set, a clustering algorithm is used to group the abnormal data, where the number of clusters of the clustering algorithm is a preset value, to obtain an abnormal data classification set; high-frequency abnormal patterns are extracted from the abnormal data classification set, the frequency of occurrence of the high-frequency abnormal patterns is counted, and the time distribution of the abnormal occurrence is determined; based on the time distribution of the abnormal occurrence, a time series prediction model is used to predict the abnormal trend of a preset time period in the future to obtain an abnormal prediction result; based on the abnormal prediction result, a dynamic feature set is generated, and the dynamic feature set is transmitted to cloud storage to obtain a cloud feature data set. In the closed environment airflow monitoring system, the processing of real-time time series data is a key link.

[0027] For example, the system can collect negative pressure values once per second, and use a sliding window method with a 60-second window size and a 10-second sliding step to segment the continuous data stream into multiple overlapping segments for subsequent analysis.

[0028] Specifically, recurrent neural networks are particularly suitable for processing such time series data.

[0029] In one embodiment, the GRU network can compress the original 60-dimensional negative pressure data into an 8-dimensional feature vector, capturing the data fluctuation pattern instead of focusing only on the absolute value, thereby improving the sensitivity of anomaly detection.

[0030] It should be noted that feature value threshold judgment is the core of anomaly detection.

[0031] For example, when the component representing the fluctuation amplitude in the extracted eigenvalue exceeds 0.75 (normal range 0.1-0.5), the system will generate an "abnormal negative pressure fluctuation" alarm and record the time of occurrence and duration.

[0032] Preferably, the K-means clustering algorithm can classify the collected abnormal alarms into three preset categories: short-term fluctuation, sustained excursion, and periodic oscillation. In biosafety cabinet monitoring, sustained excursion anomalies are often associated with filter blockage, while periodic oscillations may indicate a ventilation system failure.

[0033] In one possible implementation, the system extracts high-frequency patterns such as "frequent occurrence of short-term fluctuation anomalies between 10:00 and 11:30 on weekdays" and uses the ARIMA model to predict the probability of anomalies occurring within the next 24 hours, such as predicting "the probability of anomalies tomorrow morning is 85%."

[0034] It is understandable that the generation of dynamic feature sets integrates historical anomaly patterns and prediction results to form a triple data structure of "anomaly type-frequency-prediction probability". After uploading to the cloud, it can be used for cross-device analysis, identifying common problems, and optimizing maintenance strategies.

[0035] Step S103 : Based on the dynamic feature set, the data is transmitted to a remote server through a cloud synchronization mechanism, and a second long short-term memory network algorithm is used to perform time series trend prediction to determine the stability trend data of the airflow and filtration system.

[0036] Airflow and filtration system operating status data collected by the sensor array is acquired to generate a dynamic feature set, which includes multidimensional parameters such as airflow velocity, pressure differential, and particulate matter concentration. This dynamic feature set is encrypted and packaged via a cloud-based synchronization mechanism, and a secure transmission channel is established for transmission to a remote server, where a timestamp and device identification information are recorded. The encrypted data received by the remote server is decrypted, and the dynamic feature set is cleaned using a data preprocessing module to remove outliers and noise, resulting in a normalized feature matrix. A time series data stream is constructed based on this normalized feature matrix, with training and validation sets divided according to time windows. A data dependency graph is generated using the Graphviz tool. A long-short-term memory network is applied to this time series data stream, with memory units and forget gate thresholds set to generate a trained prediction model. This prediction model is used to perform trend predictions for future time windows, generating airflow and filtration system stability trend data. If this stability trend data falls below a preset threshold, an intelligent adjustment mechanism is triggered to automatically adjust filtration system parameters and send an alert to the management platform.

[0037] For example, a sensor array collects data on airflow and filtration system operation, generating a dynamic feature set containing multi-dimensional parameter data such as airflow velocity, pressure differential, and particulate matter concentration. This dynamic feature set is encrypted and packaged via a cloud-based synchronization mechanism, and a secure transmission channel is established for transmission to a remote server data center, where timestamps and device identification information are recorded. After receiving the encrypted data, the remote server uses a data preprocessing module to clean the dynamic feature set, remove outliers and noise, and normalize it to produce a normalized feature matrix. Based on the normalized feature matrix, a time series data stream is constructed, with training and validation sets divided into time windows. A data dependency graph is constructed using the Graphviz tool. A long short-term memory (LSTM) network is applied to the time series data stream, with memory cells and forget gate thresholds configured to capture long-term dependencies between airflow and filtration system parameters. The trained LSTM model performs trend predictions for future time windows, outputting airflow and filtration system stability trend data to determine the direction of system operational status changes. If the stability trend data falls below the warning threshold, an intelligent adjustment mechanism is triggered to automatically adjust the filtration system parameters and simultaneously send a warning message to the management platform. In a closed environment airflow monitoring system, the data collected by the sensor array includes multi-dimensional parameters such as airflow velocity, pressure difference, and particle concentration, which constitute a dynamic feature set.

[0038] For example, airflow velocity can be measured once per second using a thermal wind speed sensor, ranging from 0.2 to 1.5 m / s. Pressure differential is recorded by a differential pressure sensor, typically ranging from 50 to 200 Pa. Particle concentration is measured using a laser particle counter, in μg / m³. These parameters collectively reflect the operating status of the biosafety cabinet, and the dynamic feature set provides the foundation for subsequent analysis.

[0039] Specifically, the cloud synchronization mechanism encrypts and packages the dynamic feature set through the AES-256 encryption algorithm, establishes a secure transmission channel of the TLS protocol, transmits the data to the remote server, and records the timestamp and device identification.

[0040] For example, a biosafety cabinet with the device identifier BSC-001 is uploaded at 2025-04-18 10:00:00. The timestamp ensures accurate data timing, and the device identifier facilitates traceability. Encrypted packaging and secure channels ensure the confidentiality and integrity of data during transmission.

[0041] In one possible implementation, after receiving the encrypted data, the remote server decrypts it using the corresponding decryption key and cleans the data through a data preprocessing module. This cleaning process includes removing outliers, such as erroneous data showing a sudden change in air velocity to 10 m / s, and eliminating noise interference through median filtering.

[0042] Preferably, the normalized feature matrix scales the data to the range of 0-1 to facilitate subsequent modeling.

[0043] For example, airflow velocity, pressure difference, and particle concentration are mapped to standardized intervals to form a unified feature matrix.

[0044] As can be understood, based on the normalized feature matrix, the time series data stream is divided into a training set and a validation set using 60-second time windows, with the training set accounting for 80% and the validation set accounting for 20%. Graphviz was used to generate a data dependency graph, showing the temporal dependency between airflow velocity and pressure difference.

[0045] For example, an increase in pressure differential can lead to a decrease in airflow velocity. This relationship is intuitively presented in the graph, assisting in model design. For time series data streams, long-short-term memory networks capture long-term dependencies by setting memory cells and forget gate thresholds.

[0046] In one embodiment, the network includes 128 hidden units, with a forget gate threshold set to 0.7. Data from the past seven days is used for training, and trends are predicted for the next 24 hours. The trained prediction model outputs data on airflow and filtration system stability trends. For example, it predicts that the pressure differential may drop to 45 Pa (lower than the normal range) over the next six hours.

[0047] For example, if stability trend data falls below a preset threshold, such as a pressure differential below 50 Pa, the system triggers an intelligent adjustment mechanism, automatically adjusting the filtration system fan speed to 1200 rpm and simultaneously sending an alert to the management platform, stating "Insufficient pressure, filter inspection required." This alert includes a timestamp and anomaly type, enabling management to quickly respond. The intelligent adjustment mechanism optimizes system performance through real-time feedback.

[0048] It's important to note that the above process, from data collection to prediction and adjustment, forms a closed-loop management system. The standardized, encrypted transmission of dynamic feature sets and model predictions ensure efficient data processing and system stability. Early warning and adjustment mechanisms further enhance the timeliness of equipment maintenance and extend its service life.

[0049] Step S104 , extracting real-time state mapping data from the stability trend data, combining with a voice interaction interface to receive user instructions, and performing noise reduction processing on the instructions using an adaptive noise suppression algorithm.

[0050] The original time series data set is obtained from the stability trend database, and the data set includes timestamps and numerical values. Pandas is used to preprocess the time series data set, fill missing values and align timestamps to obtain preprocessed time series data. For the preprocessed time series data, the Pelt algorithm is used to detect change points, divide the time series into segments, and obtain segment labels and the start and end times of each segment. Based on the segment labels and start and end times, the mean and variance in each window are calculated with a preset time window length to obtain a window statistics table. The user voice signal is collected through PyAudio, and the Mel spectrum features of the voice signal are extracted using librosa to obtain a feature vector. Spectral subtraction is applied to the feature vector to remove environmental noise and obtain a noise reduction feature matrix. The noise reduction feature matrix is input into the pre-trained Kaldi speech recognition model to obtain text instructions and confidence scores. Based on the window statistics table and the text instructions, an association matrix is constructed, where the rows of the association matrix correspond to the window number, the columns correspond to the instruction type number, and the matrix element value is the product of the window state change rate and the instruction matching degree. Traverse the association matrix, obtain the operation corresponding to the maximum value, update the status field of the database through SQL statements, and write the operation time, the text instruction, and the status value into the database log table.

[0051] It's important to note that PyAudio and librosa are two commonly used Python libraries for speech processing, each with distinct functions and uses. PyAudio is used for audio acquisition, playback, and processing, and is a tool for implementing audio stream input and output. librosa is used for audio signal analysis and feature extraction, particularly for calculating audio features such as Mel-spectrograms for further speech analysis and processing. Kaldi is a powerful and flexible open-source speech recognition toolkit that provides a complete pipeline from audio feature extraction to final recognition results. It supports a variety of acoustic models and decoding methods, and is capable of handling large-scale speech recognition tasks.

[0052] For example, a raw dataset is read from a stability trend database. Pandas is used to fill missing values and align timestamps, outputting preprocessed time series data. The Pelt algorithm is used to detect change points in the time series data, segmenting the time series into segments and outputting segment labels and start and end times. Based on the segmentation results, the mean and variance within 30-second windows are calculated to generate a window statistics table. User voice commands are collected using PyAudio, and librosa is called to extract mel-spectrogram features, outputting an 80-dimensional feature vector. Spectral subtraction is applied to the feature vector to remove ambient noise, generating a denoised feature matrix. The feature matrix is then fed into a pretrained Kaldi speech recognition model, which outputs the text command and confidence score. Based on the window statistics table and the text command, an m-row, n-column association matrix is constructed, where m is the window number, n is the command type number, and the matrix elements are the product of the state change rate and the command matching degree. The association matrix is traversed to select the operation corresponding to the maximum value, and the database status field is updated using SQL statements. The operation time, command text, and status value are written to the database log table.

[0053] It should be noted that pandas is a very powerful data processing and analysis tool. When performing tasks such as missing value filling, timestamp alignment, and data resampling, Pandas provides concise and efficient methods that can greatly improve data processing efficiency.

[0054] For example, in a closed environment airflow monitoring system, the stability trend database stores biosafety cabinet operating data, including timestamps and numerical values such as airflow velocity and pressure differential. The timestamp is recorded as 2025-04-18 10:00:00, and the airflow velocity is 0.8 m / s. When using pandas preprocessing, missing values may be caused by temporary sensor failure. For example, if a timestamp lacks a pressure differential, this can be filled using linear interpolation. Assuming the previous and next data are 100 Pa and 102 Pa, the interpolated value is 101 Pa. Aligning timestamps ensures that the data is separated by 1 second, generating uniform preprocessed time series data and ensuring temporal consistency for subsequent analysis.

[0055] Specifically, the Pelt algorithm is used to detect change points in time series and identify sudden changes in air velocity or pressure difference.

[0056] For example, a sudden drop in airflow velocity from 0.8 m / s to 0.3 m / s could be caused by a clogged filter, triggering a change point. The Pelt algorithm uses statistical characteristics to segment data, generating segment labels such as "normal," "abnormal," and start and end times. For example, the segment from 2025-04-18 10:05:00 to 10:10:00 is considered "abnormal." This segmentation provides a clear time series for subsequent analysis.

[0057] In a possible implementation, 60 seconds is used as a time window, and the mean and variance of each window are calculated to form a window statistics table.

[0058] For example, the mean air velocity in a window is 0.7 m / s and the variance is 0.01, reflecting operational stability. The table records the window number, mean, and variance to facilitate linking to subsequent instructions.

[0059] Preferably, PyAudio captures user voice signals, such as when an operator says "check filters." Librosa extracts Mel-spectrogram features to form a feature vector that captures the frequency distribution of the voice. Spectral subtraction removes ambient noise, such as the sound of a running fan, generating a noise reduction feature matrix to ensure clear voice signals.

[0060] For example, after noise reduction, the feature matrix retains the clear spectrum of speech and improves recognition accuracy.

[0061] In one embodiment, the noise reduction feature matrix is input into a Kaldi speech recognition model, which outputs the text command "Check filter" with a confidence score of 0.95. This high confidence score ensures the command is reliable. The model is pre-trained to accommodate laboratory ambient noise and is suitable for biosafety cabinet scenarios.

[0062] As you can understand, the correlation matrix is constructed based on the window statistics table and the text instructions. Rows represent window numbers, and columns represent instruction types, such as "check" and "adjust." Matrix elements are the product of the window state change rate (e.g., airflow velocity decrease rate 0.1 m / s) and the instruction matching degree (e.g., 0.9).

[0063] For example, a window with a high change rate and high instruction matching degree and a large element value reflects an urgent need for operation.

[0064] For example, traverse the association matrix to obtain the operation corresponding to the maximum value, such as "adjust fan speed." Use a SQL statement to update the database status field to "adjusting," recording the operation time 2025-04-18 10:15:00, the command "adjust fan," and the status value "1200 rpm." This information is written to a log table for easy traceability and auditing.

[0065] It's important to note that the above process forms a closed-loop management system through data processing, change point detection, voice recognition, and correlation analysis. The introduction of voice commands improves operational convenience, while SQL logging ensures operational traceability, collectively optimizing the maintenance efficiency of biosafety cabinets.

[0066] Step S105: If the signal-to-noise ratio of the voice signal after noise reduction is higher than a preset threshold, instruction analysis is performed; if it is lower than the preset threshold, a repeat prompt is generated to obtain user control intention data.

[0067] The original voice data is collected by a microphone array, and the original voice data contains environmental noise. The LMS algorithm is used to perform noise reduction processing on the original voice data to obtain a voice signal after noise reduction. The signal-to-noise ratio of the voice signal after noise reduction is calculated, and the signal-to-noise ratio value is determined by the formula SNR=10*log10(Ps / Pn), where Ps represents the energy of the voice signal after noise reduction and Pn represents the noise energy. The preset signal-to-noise ratio threshold is read from the XML configuration file, and the preset signal-to-noise ratio threshold is determined according to the application scenario. If the signal-to-noise ratio value is higher than the preset signal-to-noise ratio threshold, the voice signal after noise reduction is semantically parsed, and the HTK tool is used to identify the content of the voice signal after noise reduction to obtain a user control instruction. If the signal-to-noise ratio value is lower than the preset signal-to-noise ratio threshold, a repeated prompt process is triggered. A control intention data packet is generated according to the user control instruction, and the control intention data packet contains the instruction type, parameters and execution priority, and is passed to the control system to perform the corresponding operation.

[0068] For example, in a closed environment airflow monitoring system, collecting raw voice data using a microphone array is the key to achieving voice interaction.

[0069] For example, the microphone array consists of four high-sensitivity microphones evenly distributed around the biosafety cabinet's operating panel, capturing operator voice commands, such as "Reduce fan speed." Raw voice data often contains ambient noise, such as the low-frequency hum of a running fan or laboratory background noise. The microphone array uses spatial filtering to enhance the signal in the direction of the target speech and initially attenuate noise in non-target directions, providing higher-quality input for subsequent processing.

[0070] Specifically, the LMS algorithm (Least Mean Squares) is used to reduce noise in raw speech data. The LMS algorithm uses an adaptive filter to iteratively adjust weights to separate speech signals from noise.

[0071] In one embodiment, when the operator says "increase pressure", the LMS algorithm estimates the ambient noise in real time based on a preset reference noise model and subtracts it from the original signal.

[0072] For example, fan noise is concentrated at 200Hz. The LMS algorithm filters out this frequency range, preserving the clarity of the voice signal. The de-noised voice signal is more suitable for subsequent analysis.

[0073] Preferably, the signal-to-noise ratio is calculated to evaluate the noise reduction effect. The signal-to-noise ratio is obtained by comparing the energy of the speech signal after noise reduction with the energy of the residual noise.

[0074] For example, after noise reduction, the speech signal energy is 100 units, the noise energy is 5 units, and the signal-to-noise ratio is approximately 13dB. The XML configuration file stores a preset signal-to-noise ratio threshold, such as 10dB, suitable for laboratory environments. The configuration file can also define thresholds for different scenarios, such as 8dB for quiet environments and 12dB for noisy environments. If the signal-to-noise ratio is above 10dB, the signal quality is sufficient for semantic parsing. If it is below the threshold, the system triggers a re-prompt process, prompting the operator to re-enter the command.

[0075] In one possible implementation, the HTK tool is used for semantic parsing and command recognition. HTK uses a hidden Markov model to analyze the de-noised speech signal, extract speech features, and match them to pre-trained command templates.

[0076] For example, if an operator says "check airflow," HTK recognizes the command and outputs the text "check airflow status." The HTK model is pre-trained using a dataset of commonly used laboratory commands to ensure high accuracy. The recognition result also includes a confidence score, such as 0.92, to reflect the credibility of the command.

[0077] It is understandable that the control intent data packet is generated based on the identified instruction and includes the instruction type, parameters and execution priority.

[0078] For example, a command requesting "Reduce fan speed" generates a data packet with the command type set to "Adjust," the parameter set to "Wind speed -0.1 m / s," and the priority set to "High." This data packet is transmitted to the control system in JSON format, and the system adjusts the biosafety cabinet fan speed to 0.6 m / s accordingly. High-priority commands are executed first, ensuring timely response to critical operations.

[0079] It should be noted that the repeated prompt process is initiated when the signal-to-noise ratio is insufficient.

[0080] For example, if the signal-to-noise ratio is 7dB, below the 10dB threshold, the system plays a "Please repeat the command" prompt through the speaker and adjusts the microphone array gain to improve the next capture quality. This process improves the system's robustness in noisy environments and ensures accurate command delivery.

[0081] For example, if an operator triggers a repeated prompt twice in a row, the system can temporarily switch to a backup microphone array, avoiding the potentially disrupted microphone and further optimizing the acquisition. This expansion solution enriches the system's adaptability and ensures the stability of voice interaction.

[0082] Step S106 , based on the user control intention data, a face recognition algorithm is used to perform biometric authentication on the operator, and if the matching degree is higher than a preset threshold, device control authority is generated.

[0083] Obtain an operator's facial image, perform denoising and normalization on the facial image using OpenCV, and obtain a first feature set. Obtain the operator's facial features from a pre-stored database and generate a second feature set. Use OpenCV to calculate the degree of match between the first feature set and the second feature set to obtain a matching result. If the matching result is higher than a preset threshold, generate an authentication pass flag and determine that the authentication is successful. Generate device control permission data based on the authentication pass flag and determine permission allocation data. Activate the REST API interface using the permission allocation data to obtain the device operation status. Use Log4j to record the device operation status and authentication results to generate an operation log.

[0084] It's understood that, in addition to the aforementioned facial recognition, this application also features personalized settings. Specifically, each person can set their own command preferences and operating modes, such as command sentence preferences, UV exposure duration, and lighting intensity. This personalized setting combined with facial recognition can further enhance safety and make safe cabinet operation more intelligent.

[0085] The purpose of adding personalized settings to the voice interaction system of the biological safety cabinet is to enhance the user's work experience and safety. Based on the personalized configuration of the face recognition module, each user can adjust some functions according to their personal preferences, making the system more tailored to their needs. The specific contents are as follows: (1) Combination of facial recognition and personalized settings Facial Recognition Module: Every time a worker approaches a biosafety cabinet, the system automatically recognizes and confirms their identity, ensuring only authorized personnel can operate the system. Through facial recognition, the system automatically loads a user's personalized settings upon entering the work area.

[0086] Personalized Configuration: Once the user's identity is confirmed, the system automatically adjusts settings based on their pre-set preferences, such as voice command preferences, UV exposure duration and intensity, etc. This not only improves operational convenience but also enhances the comfort of the working environment.

[0087] (2) Personalized settings for voice commands Sentence Preferences: Users can set up frequently used voice commands based on their personal preferences. For example, some users prefer short, concise instructions, while others prefer more detailed descriptions. Through the voice recognition system, users can select and adjust the structure of sentences in the settings to make interactions smoother and more personalized.

[0088] Adjusting command feedback: In addition to personalized voice commands, voice systems can also adjust feedback methods based on user needs. For example, some users may prefer concise voice prompts, while others may prefer detailed operating instructions or warning messages.

[0089] (3) Personalized setting of UV exposure duration and intensity UV exposure duration: UV radiation is a key feature of biosafety cabinets for disinfection. Users can set the duration of UV exposure based on experimental requirements. For example, some experiments may require longer UV exposure times, while others may only require a shorter disinfection process. The system allows users to customize the duration of UV exposure based on their work needs.

[0090] UV intensity adjustment: In addition to the duration of the exposure, the intensity of the UV light can also be customized to meet user needs. This is crucial for biosafety requirements in different experiments. Users can adjust the intensity of the UV light to ensure effective disinfection without compromising the accuracy of experimental operations.

[0091] (4) Light intensity adjustment Ambient Light Adjustment: The light intensity within a biosafety cabinet significantly impacts the user's work environment. Personalized settings allow users to adjust the light intensity based on their comfort level. For example, some users may prefer softer lighting, while others prefer brighter lighting to better observe operational details.

[0092] Automatic lighting adjustment in different working modes: The system can automatically adjust the lighting brightness according to different working modes (such as sample processing mode and cleaning mode). For example, in cleaning mode, the system can automatically increase the lighting intensity to ensure that users can see all details; while in sample processing mode, the brightness may be lowered to reduce light interference with experiments.

[0093] (5) System self-regulation and learning Intelligent Adaptation: The system can record user preferences and operating habits and optimize based on long-term usage data. If a user repeatedly uses a specific setting in a certain work environment, the system can automatically recommend that setting or automatically load it the next time the user uses it. This significantly saves users time while ensuring operational consistency.

[0094] Optimizing operating modes: In addition to basic settings, biosafety cabinets can also intelligently adjust the device's status according to the needs of different operating modes. For example, in high-risk experimental environments, the system may automatically extend the UV exposure time or increase the air filtration intensity to ensure the safety of the experimental process.

[0095] (6) Security and privacy protection Data protection: Because the system involves user personalization and facial recognition technology, it must ensure the security and privacy of user data. All personalization and authentication data must be encrypted and stored to prevent leakage. Furthermore, the system should undergo regular security checks and updates to prevent any security vulnerabilities.

[0096] Permission management: Different permissions can be set for different users. For example, certain operations can only be modified by experimental personnel or managers at a certain level, thus ensuring security and data accuracy.

[0097] In summary, the biosafety cabinet's voice interaction system, by combining facial recognition modules and personalized settings, not only provides a convenient operating experience but also significantly improves work efficiency and safety. Users can customize voice commands, UV exposure duration, light intensity, and other functions according to their needs. The system also automatically learns user preferences to further optimize work modes. Through intelligent and personalized settings, the biosafety cabinet can provide more efficient and safer laboratory operations.

[0098] For example, in the voice interaction system of a biosafety cabinet, facial recognition is used to ensure the operator's identity security and prevent unauthorized personnel from operating the equipment.

[0099] Specifically, the system captures an image of the operator's face via a camera. The image resolution is 1280x720 pixels, and the format is RGB. OpenCV is used for image denoising, using a Gaussian blur method to smooth the image and reduce interference from lighting changes or camera noise. Normalization is performed to scale pixel values to a range of 0-1, generating a first feature set. For example, this includes 128-dimensional facial landmark data, such as the locations of the corners of the eyes and the tip of the nose.

[0100] In one possible implementation, a pre-existing database stores the operator's facial features. 68 key points are extracted using the dlib library to generate a second feature set. The database is saved in JSON format, including information such as the operator ID, feature vector, and registration time. OpenCV uses Euclidean distance to calculate the degree of match between the first and second feature sets, generating a matching score.

[0101] Preferably, the preset threshold is set to 0.85. If the matching score is 0.9, which is higher than the threshold, an authentication pass mark is generated.

[0102] For example, after authentication, the system generates device control permission data and determines permission allocation data in JSON format, including the operator ID, permission level (e.g., "senior"), and validity period (e.g., 30 minutes). Based on the permission allocation data, the REST API is activated, which sends a command (e.g., "Turn on UV light") via a POST request. The device then returns the operating status, such as "UV light on" or "Wind speed 0.5 m / s."

[0103] It should be noted that Log4j records operation logs, and the log files are stored on the local server in the format of "time-operator ID-operation status-authentication result".

[0104] In one embodiment, the log records that operator Zhang executed "Increase Wind Speed" at 10:00 on April 18, 2025, with an authentication score of 0.92 and a status of "Success". The log supports querying by time or ID for easy tracing.

[0105] Understandably, in the extended solution, the system adjusts camera exposure parameters to improve image quality in low-light conditions. If the matching score is below 0.85 but above 0.7, the system triggers secondary authentication, prompting the operator to adjust their posture. This process improves authentication robustness, ensures accurate permission allocation, and secure device operation.

[0106] In step S107 , based on the device control authority, a reinforcement learning algorithm is used to construct a reward function with the goals of minimizing energy consumption and maximizing the particle concentration ratio, and the operating parameters of the environmental control device are dynamically adjusted.

[0107] Real-time data from environmental control devices is acquired at a fixed frequency via the Modbus TCP protocol. This data includes current, PM2.5 concentration, temperature, relative humidity, and fan speed. A 17-dimensional state vector is generated, containing the difference between indoor and outdoor PM2.5 concentrations, the temperature deviation setpoint, the humidity deviation setpoint, and the current fan speed. A three-level permission matrix is constructed based on the RBAC model, defining wind speed adjustment ranges for administrators, maintenance personnel, and users as full, restricted, and narrow, respectively. A permission range table is generated. The action space is defined as a 3D vector consisting of wind speed increment, filter level, and cooling mode switch, generating an action vector. A normalized reward function R = 0.6 × (1-E / E_max) + 0.4 × (C_out / C_in) is constructed, where E is the device power consumption, E_max is the maximum device power consumption, C_out is the sliding average of the outdoor PM2.5 concentration, and C_in is the sliding average of the indoor PM2.5 concentration. The instantaneous reward value is calculated. A three-layer fully connected neural network is used to implement a deep Q network. The neural network includes a 17-node input layer, a 64-node hidden layer, and a 3-node output layer. The parameters are updated using the Adam optimizer, and data is sampled from the experience replay buffer for training to generate a control strategy. Before issuing a control instruction, the wind speed value is checked to see if it exceeds the authority range table and the equipment safety limit. If it does, it is truncated to the nearest legal value to generate an adjusted wind speed set value. The wind speed set value is issued to the inverter via the OPCUA protocol to obtain the deviation between the actual speed and the target speed and generate a deviation record. The 17-dimensional state vector, action vector, and instant reward value are stored in a time series database to generate a record containing a timestamp.

[0108] For example, in a real-time data acquisition scenario of an environmental control device, data interaction is achieved through the ModbusTCP protocol, and the acquisition frequency is once per second.

[0109] Specifically, the device's sensors monitor real-time current values (e.g., 10A), PM2.5 concentrations (e.g., 30μg / m³ indoors and 100μg / m³ outdoors), temperature (e.g., 25°C), relative humidity (e.g., 60%), and fan speed (e.g., 1200rpm). After protocol parsing, the collected data generates a 17-dimensional state vector, including the indoor and outdoor PM2.5 concentration difference (e.g., 70μg / m³), temperature deviation (e.g., 2°C), humidity deviation (e.g., 5%), and current fan speed. This data collection is performed via an industrial gateway that supports ModbusTCP client mode, ensuring stable data transmission.

[0110] In one possible implementation, a three-level permission matrix is constructed based on the RBAC model to clearly define the wind speed adjustment permissions for different roles. Administrators can adjust the full range, such as 0-2000 rpm; maintenance personnel can adjust the speed within a limited range, such as 500-1500 rpm; and users can adjust the speed within a narrow range, such as 800-1200 rpm. The permission range table is stored in JSON format and contains role IDs and corresponding ranges.

[0111] It should be noted that permission verification is performed before the control command is issued. If the user attempts to set 1500rpm, the system will truncate it to 1200rpm to ensure compliance with the operation.

[0112] Specifically, the action space is defined as a 3D vector, which includes wind speed increments such as +100 rpm, filtration levels such as high-efficiency filtration, and cooling mode switches such as on.

[0113] For example, if indoor PM2.5 levels exceed the standard, the system generates an action vector, setting the wind speed increment to +200 rpm and switching to high-efficiency filtration mode. This action generation is based on a deep Q-network, which takes as input a 17-dimensional state vector, processes it through a 64-node hidden layer, and outputs a 3-dimensional action vector.

[0114] Preferably, the experience replay buffer stores 1000 historical data, and the Adam optimizer updates parameters at a learning rate of 0.001 to improve strategy stability.

[0115] In one embodiment, the normalized reward function combines power consumption and PM2.5 concentration ratio.

[0116] For example, if the device power consumption is 500W, the maximum power consumption is 1000W, the outdoor PM2.5 sliding average is 90μg / m³, and the indoor PM2.5 sliding average is 20μg / m³, the calculated reward value reflects the energy conservation and air quality optimization effect. The reward value guides network training and optimizes the control strategy.

[0117] It is understandable that before the control command is issued, the system checks whether the wind speed exceeds the authority or safety limit such as 1800rpm. If it exceeds the limit, it is adjusted to a legal value such as 1700rpm.

[0118] In one possible implementation, the adjusted wind speed setpoint is sent to the inverter via the OPC UA protocol. The inverter then reports the actual speed, such as 1690 rpm, which deviates by 10 rpm from the target. This deviation is then recorded and stored in a log. The 17-dimensional state vector, action vector, and reward value are stored in the InfluxDB database with a timestamp of 2025-04-18 10:00:00 to support subsequent analysis.

[0119] For example, operation and maintenance personnel can query the status vector of a certain period of time and analyze the equipment operation trend.

[0120] It should be noted that in the extended solution, if network delays interrupt data collection, the system activates local caching and retransmits data after the connection is restored to ensure record integrity. This solution uses multi-level verification and feedback to ensure control accuracy and data traceability, enhancing the intelligence level of environmental control devices.

[0121] Step S108: generating a configuration update instruction for the sealed equipment according to the real-time status mapping data and the optimized operating parameters and executing equipment linkage control.

[0122] Acquire state sensor data of the sealed device, which includes timestamp, temperature, pressure, and vibration intensity, and generate an original state data set. Use Pandas to perform a dropna operation on the original state data set to delete missing values and obtain a first state data set. Use a query operation to filter out abnormal values in the temperature, pressure, and vibration intensity in the first state data set that exceed 3 times the standard deviation to obtain a cleaning state data set. Use Scikit-learn's StandardScaler to perform Z-score standardization on the cleaning state data set to obtain a standardized state data set with a mean of 0 and a variance of 1. If the temperature or pressure in the standardized state data set exceeds the preset threshold table, the target temperature and target pressure are calculated by a PID controller to generate an optimized parameter set. Based on the optimized parameter set and the preset device instruction mapping table, a configuration update instruction containing a set value and execution time is generated. Before executing the configuration update instruction, determine whether the target temperature is within the preset temperature range and whether the target pressure is within the preset pressure range. If the verification passes, send the configuration update instruction to the device actuator. Obtain the status sensor data of the updated device. If the temperature fluctuation is less than the preset fluctuation threshold and persists for a preset time, generate a linkage control signal containing the current operating mode and transmit it to the associated device via the Modbus protocol. Receive status messages containing the actual temperature and pressure from the associated device, calculate the deviation between the actual value and the target value, generate control effect data containing the deviation value, and write it to the original status data set.

[0123] Scikit-learn is a powerful Python library that provides a variety of machine learning algorithms and tools widely used in data processing and modeling. StandardScaler is a class in Scikit-learn that performs Z-score normalization on data, converting it to a mean of 0 and a standard deviation of 1. This ensures that different features have the same scale, optimizing the performance of machine learning algorithms.

[0124] For example, the state sensor data collection of the closed equipment is centered on timestamps to ensure data traceability.

[0125] For example, a sensor collects data once a minute, recording timestamps such as 2025-04-18 10:00:00, temperature such as 28°C, pressure such as 101kPa, and vibration intensity such as 0.5mm / s, generating a raw state data set. The data collection device uses industrial-grade sensors to ensure accuracy, and the data is transmitted to the gateway via an RS485 interface to ensure real-time performance.

[0126] It is understandable that the original state dataset may contain missing values due to sensor failure or transmission interruption, affecting subsequent analysis.

[0127] In one possible implementation, Pandas' dropna operation is used to handle missing values.

[0128] For example, if the pressure value of a record is empty, dropna will delete it and generate the first-state dataset. This operation ensures data integrity and provides a reliable foundation for subsequent cleaning.

[0129] It should be noted that missing value processing needs to be combined with device operation logs to analyze the causes of missing values, such as sensor disconnection or power fluctuations, in order to optimize the collection process.

[0130] Specifically, the query operation filters outliers based on the 3 times standard deviation rule.

[0131] For example, if the mean temperature is 27°C, the standard deviation is 1°C, and the threshold is 24-30°C, then if a temperature record is 32°C, the query will exclude it. Similarly, outliers in pressure and vibration intensity are also filtered out to generate a cleaned data set. This step uses statistical methods to eliminate extreme values and improve data quality.

[0132] Preferably, StandardScaler performs Z-score standardization to convert the temperature, pressure, and vibration intensity of the cleaning status data set to have a mean of 0 and a variance of 1.

[0133] For example, a temperature of 28°C might be normalized to 0.5, and a pressure of 101 kPa might be normalized to 0.3. This process unifies the dimensions and facilitates subsequent control algorithm processing.

[0134] It is understood that normalization preserves the data distribution characteristics and ensures that different physical quantities can be compared.

[0135] In one embodiment, if the temperature or pressure in the normalized data exceeds a preset threshold value table, such as a temperature threshold value of 1.5 or a pressure threshold value of 1.2, the PID controller calculates the optimized parameters.

[0136] For example, when the temperature exceeds the target, the PID controller adjusts the cooling power based on the deviation to generate a target temperature of 26°C. When the pressure exceeds the target, the valve opening is adjusted to generate a target pressure of 100 kPa. The optimized parameter set reflects the real-time needs of the equipment and ensures operational stability.

[0137] For example, based on a device command mapping table, an optimized parameter set generates a configuration update command. The mapping table defines a temperature of 26°C as corresponding to 60% cooling power and a pressure of 100 kPa as corresponding to 50% valve opening, along with an execution time, such as 10:00:05. After the command is generated, the system verifies that the target temperature is between 24°C and 30°C, and the pressure is between 95 and 105 kPa. If verification passes, the command is sent to the actuator via the Modbus protocol. This verification mechanism prevents invalid commands and improves control reliability.

[0138] In one embodiment, after the update, the sensor collects a temperature of, for example, 26.2°C, with a fluctuation less than a 0.5°C threshold value for 5 minutes, and generates a linkage control signal.

[0139] For example, a signal instructing an associated device to enable auxiliary cooling is transmitted via Modbus. The associated device reports an actual temperature of 26.1°C and a pressure of 100.2 kPa. Deviations, such as 0.1°C for temperature and 0.2 kPa for pressure, are calculated to generate control effect data. This deviation data is written to the original state dataset, enabling operations and maintenance personnel to analyze control effectiveness.

[0140] It can be understood that the deviation data is stored in the original data set to form a closed loop.

[0141] For example, operations personnel can query data from 10:00:00 to 10:10:00 on April 18, 2025, to analyze temperature fluctuation trends and optimize PID parameters. This solution ensures equipment operation accuracy and traceability through data cleaning, standardization, control, and feedback.

[0142] Example 2 See also Figure 4 The present invention provides an intelligent control system for a biological safety cabinet, which uses the above-mentioned method, including: Data acquisition and monitoring module: used to obtain real-time data on air pressure difference, wind speed, and particulate matter concentration ratio collected by edge computing nodes, and use sensor arrays to monitor negative pressure airflow parameters of closed equipment; Feature extraction and alarm module: used to extract features from the real-time data using a first long short-term memory network algorithm, and generate an abnormality alarm if the extracted feature value deviates from a preset threshold, thereby forming a dynamic feature set; Data transmission and prediction module: used to transmit the dynamic feature set to a remote server through a cloud synchronization mechanism, use the second long short-term memory network algorithm to perform time series trend prediction, and determine the stability trend data of the airflow and filtration system; Data extraction and voice module: used to extract real-time status mapping data from the stability trend data, receive user instructions in conjunction with the voice interaction interface, and use an adaptive noise suppression algorithm to reduce noise on the instructions; Signal-to-noise ratio judgment module: If the signal-to-noise ratio of the voice signal after noise reduction is higher than a preset threshold, the command analysis is performed; if it is lower than the threshold, a repeat prompt is generated to obtain user control intention data; Biometric authentication module: used to perform biometric authentication on the operator based on the user control intention data using a face recognition algorithm, and generate device control permissions if the matching degree is higher than a preset threshold; Device operation dynamic adjustment module: Based on the device control authority, a reinforcement learning algorithm is used to construct a reward function with the goals of minimizing energy consumption and maximizing the particle concentration ratio, and the operating parameters of the environmental control device are dynamically adjusted; Instruction generation and linkage control module: used to generate configuration update instructions for the closed equipment and perform equipment linkage control based on the real-time status mapping data and optimized operating parameters.

[0143] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0144] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0145] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0146] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to equivalent embodiments using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. An intelligent control method for a biological safety cabinet, characterized by: The steps include: Acquire real-time data on air pressure difference, wind speed, and particle concentration ratio collected by edge computing nodes, and use sensor arrays to monitor negative pressure airflow parameters in enclosed equipment; Using a first long short-term memory network algorithm to extract features from the real-time data, generating an abnormality alarm if the extracted feature value deviates from a preset threshold, and forming a dynamic feature set; Based on the dynamic feature set, the data is transmitted to a remote server through a cloud synchronization mechanism, and a second long short-term memory network algorithm is used to perform time series trend prediction to determine the stability trend data of the airflow and filtration system; Extracting real-time state mapping data from the stability trend data, combining it with a voice interaction interface to receive user instructions, and using an adaptive noise suppression algorithm to perform noise reduction on the instructions; If the signal-to-noise ratio of the denoised voice signal is higher than a preset threshold, command parsing is performed; if it is lower than the preset threshold, a repeat prompt is generated to obtain user control intention data; Based on the user control intention data, a face recognition algorithm is used to perform biometric authentication on the operator, and if the matching degree is higher than a preset threshold, device control permission is generated; Based on the device control authority, a reinforcement learning algorithm is used to construct a reward function with the goals of minimizing energy consumption and maximizing the particle concentration ratio, and the operating parameters of the environmental control device are dynamically adjusted; According to the real-time status mapping data and the optimized operating parameters, a configuration update instruction of the sealed equipment is generated and equipment linkage control is executed.

2. The intelligent control method for a biological safety cabinet according to claim 1, characterized in that: The acquisition of real-time data on air pressure difference, wind speed, and particle concentration ratio collected by the edge computing node and the use of a sensor array to monitor the negative pressure airflow parameters of the sealed equipment include: Obtain the air pressure difference, wind speed and particle concentration data of key monitoring points of closed equipment to form a multi-dimensional parameter set; Preprocessing the multidimensional parameter set, eliminating sensor noise through a Kalman filter algorithm, and obtaining smoothed airflow parameters; Calculating an air pressure difference based on the smoothed airflow parameters, and comparing the air pressure difference with a preset threshold to determine whether the negative pressure value of the device is within a safe operating range; If the air pressure difference is lower than the preset threshold, an abnormal event is recorded; Performing weighted averaging processing on the wind speed values collected at the monitoring points to obtain fused wind speed data; Draw an airflow direction diagram inside the device based on the fused wind speed data to identify airflow blind spots; Perform time series analysis on the particle concentration data of the key monitoring points and establish a prediction model using an LSTM neural network; The prediction model is trained using historical particle concentration data to obtain a short-term trend in particle concentration; The airflow parameters are transmitted to the cloud in a hierarchical manner according to real-time requirements, with high-priority abnormal data being transmitted first; Perform correlation analysis on airflow parameters and equipment status indicators stored in the cloud, and establish a mapping relationship between parameter anomalies and equipment failures through a decision tree algorithm; Feature selection methods are used to determine key parameters and identify potential risk points.

3. The intelligent control method for a biological safety cabinet according to claim 1, characterized in that: The first long short-term memory network algorithm is used to extract features from the real-time data, and an abnormality alarm is generated if the extracted feature value deviates from a preset threshold, thereby forming a dynamic feature set, including: Acquire real-time time series data, and segment the real-time time series data using a sliding window method, wherein the window size of the sliding window is a preset value, and the sliding step length of the sliding window is a preset value, to obtain a time series segment set; For the time series segment set, a deep learning network is used to perform feature extraction to obtain a feature value set, where the deep learning network is a recurrent neural network, and the dimension of the feature value set is a preset dimension; Performing a threshold judgment on each eigenvalue in the eigenvalue set, and generating an abnormality alarm if the eigenvalue exceeds a preset threshold, thereby obtaining an abnormality alarm record set; According to the abnormal alarm record set, the abnormal data is grouped using a clustering algorithm, the number of clusters of the clustering algorithm is a preset value, and an abnormal data classification set is obtained; Extracting high-frequency abnormal patterns from the abnormal data classification set, counting the occurrence frequency of the high-frequency abnormal patterns, and determining the time distribution of abnormal occurrence; Based on the time distribution of the anomaly occurrence, a time series prediction model is used to predict the anomaly trend in a preset time period in the future to obtain an anomaly prediction result; A dynamic feature set is generated based on the abnormality prediction result, and the dynamic feature set is transmitted to a cloud storage to obtain a cloud feature data set.

4. The intelligent control method for a biological safety cabinet according to claim 1, characterized in that: The dynamic feature set is transmitted to a remote server via a cloud synchronization mechanism, and a second long short-term memory network algorithm is used to perform time series trend prediction to determine the stability trend data of the airflow and filtration system, including: Acquiring airflow and filtration system operating status data collected by the sensor array to obtain a dynamic feature set, wherein the dynamic feature set includes multi-dimensional parameters of airflow velocity, pressure difference, and particulate matter concentration; Encrypt and package the dynamic feature set through a cloud synchronization mechanism, establish a secure transmission channel, transmit it to a remote server, and record the timestamp and device identification information; Decrypting the encrypted data received from the remote server, using a data preprocessing module to clean the dynamic feature set, remove outliers and noise interference, and obtain a normalized feature matrix; Constructing a time series data stream based on the normalized feature matrix, dividing the training set and the validation set according to the time window, and using the Graphviz tool to generate a data dependency graph; For the time series data stream, a long short-term memory network is applied, memory units and forget gate thresholds are set, and a trained prediction model is obtained; Perform trend prediction on the future time window by using the prediction model to obtain the stability trend data of the airflow and filtration system; If the stability trend data is lower than a preset threshold, the intelligent adjustment mechanism is triggered to automatically adjust the filtering system parameters and send an early warning message to the management platform.

5. The intelligent control method for a biological safety cabinet according to claim 1, characterized in that: The extracting of real-time state mapping data from the stability trend data, receiving user instructions in conjunction with a voice interaction interface, and performing noise reduction processing on the instructions using an adaptive noise suppression algorithm include: Acquire an original time series data set from a stability trend database, wherein the data set includes a timestamp and a value; Preprocessing the time series data set, filling missing values and aligning timestamps to obtain preprocessed time series data; For the preprocessed time series data, the Pelt algorithm is used to detect change points, divide the time series into segments, and obtain segment labels and the start and end times of each segment; According to the segmentation labels and the start and end times, the mean and variance in each window are calculated with a preset time window length to obtain a window statistics table; Collecting a user's voice signal and extracting Mel-spectrogram features of the voice signal to obtain a feature vector; Applying spectral subtraction to the eigenvector to remove environmental noise to obtain a noise reduction feature matrix; Inputting the noise reduction feature matrix into a pre-trained speech recognition model to obtain a text instruction and a confidence score; Constructing an association matrix based on the window statistics table and the text instructions, wherein the rows of the association matrix correspond to the window numbers, the columns correspond to the instruction type numbers, and the matrix element values are the products of the window state change rate and the instruction matching degree; Traverse the association matrix, obtain the operation corresponding to the maximum value, and update the status field of the database through SQL statements; The operation time, the text instruction and the status value are written into the database log table.

6. The intelligent control method for a biological safety cabinet according to claim 1, characterized in that: If the signal-to-noise ratio of the voice signal after noise reduction is higher than a preset threshold, instruction parsing is performed; if it is lower than the threshold, a repeat prompt is generated to obtain user control intention data, including: Collecting original speech data through a microphone array, wherein the original speech data contains environmental noise; The LMS algorithm is used to perform noise reduction processing on the original speech data to obtain a noise-reduced speech signal; Calculating the signal-to-noise ratio of the noise-reduced speech signal and determining a signal-to-noise ratio value; Reading a preset signal-to-noise ratio threshold from an XML configuration file, where the preset signal-to-noise ratio threshold is determined according to an application scenario; If the signal-to-noise ratio value is higher than the preset signal-to-noise ratio threshold, semantically analyzing the noise-reduced voice signal, identifying the content of the noise-reduced voice signal, and obtaining a user control instruction; If the signal-to-noise ratio value is lower than the preset signal-to-noise ratio threshold, a repeat prompt process is triggered; A control intention data packet is generated according to the user control instruction, and the control intention data packet includes the instruction type, parameters and execution priority, and is transmitted to the control system to execute the corresponding operation.

7. The intelligent control method for a biological safety cabinet according to claim 1, characterized in that: The method of performing biometric authentication on the operator using a face recognition algorithm based on the user control intention data and generating device control permissions if the matching degree is higher than a preset threshold includes: Obtaining an operator's facial image, and performing denoising and normalization processing on the facial image using OpenCV to obtain a first feature set; Obtaining operator facial features from a pre-stored database to generate a second feature set; Using OpenCV to calculate the matching degree between the first feature set and the second feature set to obtain a matching result; If the matching result is higher than the preset threshold, an authentication pass mark is generated to determine that the authentication is successful; Generating device control authority data according to the authentication pass mark and determining authority allocation data; Activate the REST API interface through the authority allocation data to obtain the device operation status; Record the device operation status and authentication results, and generate an operation log.

8. The intelligent control method for a biological safety cabinet according to claim 1, characterized in that: Based on the device control authority, a reinforcement learning algorithm is used to construct a reward function with the goal of minimizing energy consumption and maximizing the particle concentration ratio, and the operating parameters of the environmental control device are dynamically adjusted, including: Acquire real-time data from the environmental control device at a fixed frequency using the Modbus TCP protocol. The real-time data includes current, PM2.5 concentration, temperature, relative humidity, and fan speed. Generate a 17-dimensional state vector containing the difference between indoor and outdoor PM2.5 concentrations, a temperature deviation setpoint, a humidity deviation setpoint, and the current fan speed. A three-level permission matrix is constructed based on the RBAC model, defining the wind speed adjustment ranges for administrators, maintenance personnel, and users as full range, restricted range, and narrow range, and generating a permission range table. Defining the action space as a 3D vector, wherein the 3D vector includes the wind speed increment, the filter level, and the cooling mode switch, and generating an action vector; Construct a normalized reward function R = 0.6 × (1-E / E_max) + 0.4 × (C_out / C_in), where E is the device power consumption, E_max is the maximum device power consumption, C_out is the sliding average of outdoor PM2.5 concentration, and C_in is the sliding average of indoor PM2.5 concentration. Calculate the instantaneous reward value. A three-layer fully connected neural network was used to implement the deep Q network, which consisted of a 17-node input layer, a 64-node hidden layer, and a 3-node output layer. The parameters were updated using the Adam optimizer, and data was sampled from the experience replay buffer for training to generate the control strategy. Before issuing a control command, check whether the wind speed value exceeds the authority range table and the equipment safety limit. If it exceeds, truncate it to the nearest legal value and generate an adjusted wind speed set value; Send the wind speed set value to the inverter through the OPCUA protocol, obtain the deviation between the actual speed and the target speed, and generate a deviation record; The 17-dimensional state vector, action vector, and immediate reward value are stored in a time series database to generate a record containing a timestamp.

9. The intelligent control method for a biological safety cabinet according to claim 1, characterized in that: Generating configuration update instructions for the sealed equipment and executing equipment linkage control based on the real-time status mapping data and the optimized operating parameters includes: Acquire state sensor data of the sealed device, wherein the state sensor data includes a timestamp, temperature, pressure, and vibration intensity, and generate an original state data set; Using Pandas to perform a dropna operation on the original state data set to obtain a first state data set; Filtering outliers in the first state data set through a query operation to obtain a cleaned state data set; performing Z-score normalization on the cleaning state data set to obtain a normalized state data set; If the temperature or pressure in the standardized state data set exceeds a preset threshold value table, the target temperature and target pressure are calculated by a PID controller to generate an optimized parameter set; Generate a configuration update instruction including a set value and an execution time according to the optimization parameter set and a preset device instruction mapping table; Before executing the configuration update instruction, determine whether the target temperature is within the preset temperature range and whether the target pressure is within the preset pressure range. If the verification passes, send the configuration update instruction to the device actuator. Obtain the status sensor data of the updated device. If the temperature fluctuation is less than the preset fluctuation threshold and lasts for a preset time, a linkage control signal containing the current operating mode is generated and transmitted to the associated device via the Modbus protocol; A status message containing actual temperature and actual pressure is received from an associated device, a deviation between the actual value and the target value is calculated, control effect data containing the deviation value is generated, and the data is written into the original status data set.

10. An intelligent control system for a biological safety cabinet, applying the method according to any one of claims 1 to 9, characterized in that: include: Data acquisition and monitoring module: used to obtain real-time data on air pressure difference, wind speed, and particulate matter concentration ratio collected by edge computing nodes, and use sensor arrays to monitor negative pressure airflow parameters of closed equipment; Feature extraction and alarm module: used to extract features from the real-time data using a first long short-term memory network algorithm, and generate an abnormality alarm if the extracted feature value deviates from a preset threshold, thereby forming a dynamic feature set; Data transmission and prediction module: used to transmit the dynamic feature set to a remote server through a cloud synchronization mechanism, use the second long short-term memory network algorithm to perform time series trend prediction, and determine the stability trend data of the airflow and filtration system; Data extraction and voice module: used to extract real-time status mapping data from the stability trend data, receive user instructions in conjunction with the voice interaction interface, and use an adaptive noise suppression algorithm to reduce noise on the instructions; Signal-to-noise ratio judgment module: If the signal-to-noise ratio of the voice signal after noise reduction is higher than a preset threshold, the command analysis is performed; if it is lower than the threshold, a repeat prompt is generated to obtain user control intention data; Biometric authentication module: used to perform biometric authentication on the operator based on the user control intention data using a face recognition algorithm, and generate device control permissions if the matching degree is higher than a preset threshold; Device operation dynamic adjustment module: Based on the device control authority, a reinforcement learning algorithm is used to construct a reward function with the goals of minimizing energy consumption and maximizing the particle concentration ratio, and the operating parameters of the environmental control device are dynamically adjusted; Instruction generation and linkage control module: used to generate configuration update instructions for the closed equipment and perform equipment linkage control based on the real-time status mapping data and optimized operating parameters.

Images