An oxygen concentration regulation method based on a hybrid intelligent model

Abstract

The application discloses an oxygen concentration regulation method based on a hybrid intelligent model. The method comprises the following steps: collecting multi-modal environment parameters; using a Kalman filter to optimize and estimate the multi-modal environment parameters, and performing feature extraction on the optimized and estimated multi-modal environment parameters; constructing an oxygen concentration prediction model based on a bidirectional LSTM network and a Transformer model, and training the oxygen concentration prediction model using a proximal policy optimization algorithm and multi-modal feature data; obtaining an oxygen concentration prediction value using real-time multi-modal environment parameters and the trained oxygen concentration prediction model; constructing an adaptive proportional-integral-derivative controller, and generating a ventilation equipment control instruction using the oxygen concentration prediction value and the adaptive proportional-integral-derivative controller to regulate the oxygen concentration. The application can accurately control the oxygen concentration and is suitable for various scenes such as mines, medical treatment and industry, and provides an intelligent solution for high-precision oxygen concentration control.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring and intelligent control technology, specifically to an oxygen concentration regulation method based on a hybrid intelligent model. Background Technology

[0002] Traditional oxygen regulation methods have three major drawbacks:

[0003] (1) Poor model adaptability: It relies on physical models or empirical formulas and is difficult to cope with dynamic disturbances such as air pressure fluctuations and changes in the intensity of human activities;

[0004] (2) Limitations of feature extraction: A single time series model cannot capture the spatial correlation of multiple parameters;

[0005] (3) Control lag problem: The parameters of classical proportional-integral-derivative control are fixed in nonlinear time-varying systems, resulting in response lag and overshoot.

[0006] In summary, existing oxygen regulation methods lack environmental adaptability and online optimization capabilities, and cannot meet the requirements for high-precision regulation. Summary of the Invention

[0007] To address the aforementioned shortcomings in the existing technology, this invention provides an oxygen concentration regulation method based on a hybrid intelligent model.

[0008] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0009] A method for regulating oxygen concentration based on a hybrid intelligent model includes the following steps:

[0010] Multimodal environmental parameters are acquired through a multimodal sensor network;

[0011] The Kalman filter is used to optimize and estimate the multimodal environmental parameters to obtain optimized estimates of the multimodal environmental parameters, and features are extracted from the optimized estimates of the multimodal environmental parameters to obtain multimodal feature data.

[0012] An oxygen concentration prediction model is constructed based on a bidirectional LSTM network and a Transformer model. The oxygen concentration prediction model is then trained using a proximal policy optimization algorithm and multimodal feature data to obtain the trained oxygen concentration prediction model.

[0013] Obtain real-time multimodal environmental parameters, and use the real-time multimodal environmental parameters and the trained oxygen concentration prediction model to obtain the oxygen concentration prediction value;

[0014] An adaptive proportional-integral-derivative (AI-DI) controller is constructed. The predicted oxygen concentration and the AI-DI controller are used to generate control commands for the ventilation equipment. These control commands are then sent to the ventilation equipment for execution to regulate the oxygen concentration.

[0015] Furthermore, multimodal environmental parameters include oxygen concentration, temperature, humidity, and air pressure.

[0016] Furthermore, the multimodal environmental parameters are optimized and estimated using a Kalman filter, including the following steps:

[0017] Differentiated modeling of multimodal environment parameters is performed to configure the state transition model of multimodal environment parameters;

[0018] The state transition model of multimodal environmental parameters is predicted and updated using a Kalman filter to obtain optimized estimates of the multimodal environmental parameters.

[0019] Furthermore, differentiated modeling of multimodal environmental parameters is performed to configure the state transition model for multimodal environmental parameters, specifically as follows:

[0020] A nonlinear state equation is used to model the oxygen concentration to configure the state transition model for the oxygen concentration, which is expressed as:

[0021]

[0022] in: for Optimal estimate of oxygen concentration at time point. It is a nonlinear mapping function. for Optimal estimate of oxygen concentration at time point. for Optimized estimate of temperature at time point. for Optimal estimate of air pressure at any given time. for Time-matter noise;

[0023] A linear observation model is used to model temperature and humidity separately, and state transition models for temperature and humidity are configured separately, as follows:

[0024]

[0025]

[0026] in: The temperature observation matrix, for Optimized estimate of temperature at time point. for The noise of temperature observation at any given time. for Optimized estimate of humidity at any given time. This is the humidity observation matrix. for Optimized estimate of humidity at any given time. for Noise in the observation of humidity at any given time;

[0027] A simplified dynamic model is used to model air pressure to configure the state transition model of air pressure, as follows:

[0028]

[0029] in: for Optimal estimate of air pressure at any given time. This represents the change in atmospheric pressure reference value. for Simplified noise at any given moment.

[0030] Furthermore, the multimodal feature data includes pressure periodicity features, time series features, and meteorological features.

[0031] Furthermore, feature extraction is performed on the optimized and estimated multimodal environmental parameters to obtain multimodal feature data, including the following steps:

[0032] The optimized estimate of air pressure is corrected based on the relationship between air pressure and altitude to obtain the corrected air pressure value, expressed as:

[0033]

[0034] in: This is the corrected air pressure value. This is an optimized estimate of the air pressure. The difference in altitude relative to the reference elevation. It is the atmospheric elevation constant;

[0035] The periodic characteristics of air pressure are extracted based on the corrected air pressure values, including 24-hour amplitude and 1-hour phase components;

[0036] The rate of change of oxygen concentration due to temperature and pressure compensation, calculated based on the corrected air pressure value, is expressed as follows:

[0037]

[0038] in: for Oxygen concentration measurement at time [time]. To adapt the sampling time step according to the dynamic characteristics of the environment, This is an optimized estimate of the temperature. This is the corrected air pressure value. This is the temperature compensation coefficient. This is the air pressure compensation coefficient;

[0039] Based on the oxygen concentration change rate compensated for by temperature and air pressure, statistical features within the time window are extracted, including mean, standard deviation, first difference and autocorrelation coefficient, in order to extract time series features.

[0040] The combined effect of temperature and humidity is calculated to extract meteorological characteristics, and is expressed as follows:

[0041]

[0042] in: As an index of the combined effect of temperature and humidity, This refers to relative humidity.

[0043] Furthermore, the oxygen concentration prediction model includes a bidirectional LSTM network, a Transformer encoder, and a fully connected output layer connected in sequence;

[0044] The bidirectional LSTM network consists of a forward LSTM layer, a backward LSTM layer, and a concatenation layer. The bidirectional LSTM network is used to input multimodal feature data. The forward LSTM layer and the backward LSTM layer are used to process the multimodal feature data to obtain the hidden state vectors of the forward LSTM layer and the backward LSTM layer, respectively. The concatenation layer is used to concatenate the hidden state vectors of the forward LSTM layer and the backward LSTM layer to obtain the temporal feature vector of the multimodal feature data.

[0045] The Transformer encoder comprises an input embedding layer, a multi-head self-attention layer with four attention heads, a feedforward network, and a Transformer layer structure containing three Transformer blocks, connected in sequence. The Transformer encoder takes the temporal feature vectors from multimodal feature data, projects these vectors into a higher-dimensional space using the input embedding layer to obtain multidimensional temporal feature vectors, captures long-term dependencies between these multidimensional temporal feature vectors using the multi-head self-attention layer with four attention heads to obtain multidimensional fused temporal feature vectors, performs a nonlinear transformation on the multidimensional fused temporal feature vectors using the feedforward network to obtain temporal feature interaction vectors, and encodes these temporal feature interaction vectors using the Transformer layer structure containing three Transformer blocks to obtain the encoded feature vectors.

[0046] The fully connected output layer consists of an input layer, a hidden layer, and a linear output layer connected in sequence; the fully connected output layer is used to map the encoded feature vector to control parameters (such as PID control parameters) to generate oxygen concentration prediction values.

[0047] Furthermore, the oxygen concentration prediction model is trained using a proximal policy optimization algorithm and multimodal feature data to obtain the trained oxygen concentration prediction model, including the following steps:

[0048] The advantage function estimate of the oxygen concentration prediction model is calculated and expressed as follows:

[0049]

[0050] in: For oxygen concentration prediction models in environmental conditions Execute control actions The advantage function, This refers to the environmental conditions, specifically the combination of oxygen concentration, temperature, humidity, and air pressure. To control the action, specifically to adjust the opening degree of the ventilation equipment. In environmental conditions Execute control actions The expected cumulative reward, In environmental conditions Benchmark value below;

[0051] The strategy gradient pruning is determined based on the advantage function estimation of the oxygen concentration prediction model, and is expressed as follows:

[0052]

[0053] in: For the clipping function, The probability ratio between the old and new strategies. This is the clipping threshold;

[0054] The policy loss of the oxygen concentration prediction model is calculated based on the advantage function and policy gradient pruning, and is expressed as follows:

[0055]

[0056] in: The strategy loss of the oxygen concentration prediction model. To take the sign of the minimum value;

[0057] The value loss of the oxygen concentration prediction model is calculated and expressed as:

[0058]

[0059] in: The value loss of the current strategy for the oxygen concentration prediction model. For the current strategy in the environment state The following value estimate, For the old strategy in the environment state The target value is calculated from the experience replay cache;

[0060] Based on the strategy loss and value loss of the oxygen concentration prediction model, the total loss of the oxygen concentration prediction model is calculated and expressed as:

[0061]

[0062] in: The total loss of the current strategy for the oxygen concentration prediction model is given. , All are weighted coefficients, determined through grid search combined with cross-validation, in the context of an underground mine. Take 0.5, Take 0.01, This is the policy entropy regularization term;

[0063] The oxygen concentration prediction model is trained based on the total loss of the oxygen concentration prediction model to obtain the trained oxygen concentration prediction model.

[0064] Furthermore, an adaptive proportional-integral-derivative controller is constructed, expressed as:

[0065] ,

[0066] ,

[0067] ,

[0068] ,

[0069]

[0070] in: for The control output at any given time is specifically as follows: The instructions for adjusting the opening degree of the ventilation equipment at all times. ,express Timing control error, for Predicted oxygen concentration at time [time]. To preset the target oxygen concentration, , , These are the proportional gain, integral gain, and derivative gain, respectively, with initial values ​​set empirically. , , These are the proportional gain increment, integral gain increment, and differential gain increment, which are dynamically adjusted through the online learning module, respectively. To control the sensitivity to changes in current error, To suppress historical noise interference.

[0071] The present invention has the following beneficial effects:

[0072] (1) This invention uses a multimodal sensor network to continuously monitor key environmental parameters such as oxygen concentration, air pressure, temperature and humidity, and obtains environmental status in real time. Then, it uses a Kalman filter to preprocess the collected multi-source data, which effectively improves the reliability of the data and filters out noise interference to ensure data accuracy.

[0073] (2) This invention uses a hybrid model that deeply integrates bidirectional LSTM and Transformer to efficiently extract the spatiotemporal features of oxygen concentration, providing solid data support for subsequent regulation. Combined with the near-end strategy optimization algorithm, the control strategy is dynamically optimized, and the oxygen supply parameters are automatically adjusted based on real-time feedback to ensure accurate oxygen supply under different working conditions.

[0074] (3) The present invention constructs an adaptive proportional-integral-derivative controller and dynamically calibrates the control parameters based on the real-time error to ensure the control accuracy. Attached Figure Description

[0075] Figure 1 This is a schematic diagram of a method for regulating oxygen concentration based on a hybrid intelligent model.

[0076] Figure 2 Simulation diagram of oxygen concentration control effect;

[0077] Figure 3 The simulation diagram shows the state estimation performance. Detailed Implementation

[0078] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0079] like Figure 1 As shown, an oxygen concentration regulation method based on a hybrid intelligent model includes steps S1-S5, as detailed below:

[0080] S1. Collect multimodal environmental parameters through a multimodal sensor network.

[0081] In an optional embodiment of the present invention, the multimodal environmental parameters include oxygen concentration, temperature, humidity and air pressure.

[0082] Specifically, a heterogeneous sensor network is first used to collect multimodal environmental parameters in real time, including oxygen concentration, temperature, humidity, air pressure, and atmospheric pressure. The multimodal sensor network is deployed in a cellular topology, with the spacing dynamically calculated based on the target area to ensure data coverage and redundancy. The sampling frequency can be dynamically adjusted according to environmental requirements to adapt to different scenarios. For example, in dynamically changing environments, the sampling frequency can be increased to capture rapidly changing oxygen concentrations; in relatively stable environments, the sampling frequency can be decreased to reduce data redundancy and energy consumption.

[0083] S2. Use a Kalman filter to optimize and estimate the multimodal environmental parameters to obtain optimized estimates of the multimodal environmental parameters, and extract features from the optimized estimates of the multimodal environmental parameters to obtain multimodal feature data.

[0084] In an optional embodiment of the present invention, the present invention utilizes a Kalman filter to optimize and estimate multimodal environmental parameters, including steps A1-A2, as follows:

[0085] A1. Differentiated modeling of multimodal environment parameters is performed to configure the state transition model for multimodal environment parameters, specifically as follows:

[0086] A nonlinear state equation is used to model the oxygen concentration to configure the state transition model for the oxygen concentration, which is expressed as:

[0087]

[0088] in: for Optimal estimate of oxygen concentration at time point. It is a nonlinear mapping function. for Optimal estimate of oxygen concentration at time point. for Optimized estimate of temperature at time point. for Optimal estimate of air pressure at any given time. for Time-matter noise;

[0089] A linear observation model is used to model temperature and humidity separately, and state transition models for temperature and humidity are configured separately, as follows:

[0090]

[0091]

[0092] in: The temperature observation matrix, for Optimized estimate of temperature at time point. for The noise of temperature observation at any given time. for Optimized estimate of humidity at any given time. This is the humidity observation matrix. for Optimized estimate of humidity at any given time. for Noise in the observation of humidity at any given time;

[0093] A simplified dynamic model is used to model air pressure to configure the state transition model of air pressure, as follows:

[0094]

[0095] in: for Optimal estimate of air pressure at any given time. This represents the change in atmospheric pressure reference value. for Simplified noise at any given moment.

[0096] A2. Use a Kalman filter to predict and update the state transition model of multimodal environmental parameters to obtain optimized estimates of the multimodal environmental parameters.

[0097] Step A2 includes the following steps:

[0098] A21. Based on the state transition model of multimodal environment parameters, construct the state vector of modal environment parameters, represented as:

[0099]

[0100] A22. Based on the state vector of modal environment parameters, use a Kalman filter to predict the state transition model of multimodal environment parameters to obtain... and , represented as:

[0101]

[0102] in: Based on Time-state vector pair Predicted oxygen concentration at any given time. for The transition matrix of the state vector at each time step describes a dynamic model of how oxygen concentration changes over time. for The optimized state vector estimate at time step 1;

[0103]

[0104] in: for The covariance matrix of the predicted state vector at time step [time]. for The optimized covariance matrix at time step, It is the transpose symbol. for The process noise covariance matrix at time t represents the noise introduced by dynamic environmental changes.

[0105] A23, based on Calculate the Kalman gain, expressed as:

[0106]

[0107] in: for The weighting coefficients at each time point, balancing the confidence levels of predicted and observed values. for The observation matrix at each time point maps the state to sensor measurements. for The observation noise covariance matrix at time t represents the sensor measurement error.

[0108] A24, based on Using the Kalman gain, the state transition model of the multimodal environment parameters is updated using a Kalman filter to obtain optimized estimates of the multimodal environment parameters, expressed as:

[0109]

[0110] in: for The optimized state vector estimate at time step [time]. for The actual sensor observation at any given time;

[0111]

[0112] in: for The covariance matrix of the state vector after optimization at each time step. It is an identity matrix.

[0113] This invention achieves adaptive fusion of multimodal parameters by configuring independent state transition models and noise covariance matrices for each parameter, and combining the confidence levels of the Kalman gain dynamic equilibrium predictions and observations.

[0114] Multimodal feature data includes pressure periodicity features, time series features, and meteorological features.

[0115] This invention performs feature extraction on the optimized and estimated multimodal environmental parameters to obtain multimodal feature data, including steps B1-B5, as follows:

[0116] B1. Correct the optimized estimate of air pressure based on the relationship between air pressure and altitude to obtain the corrected air pressure value, expressed as:

[0117]

[0118] in: This is the corrected air pressure value. This is an optimized estimate of the air pressure. The difference in altitude relative to the reference elevation. is the atmospheric elevation constant.

[0119] B2. Extract the periodic characteristics of air pressure based on the corrected air pressure value, including the 24-hour amplitude and 1-hour phase components.

[0120] The 24-hour amplitude is used to reflect the range of air pressure changes over 24 hours, while the 1-hour phase component is used to reflect the trend of air pressure changes over 1 hour.

[0121] B3. Calculate the rate of change of oxygen concentration due to temperature and pressure compensation based on the corrected air pressure value, expressed as:

[0122]

[0123] in: for Oxygen concentration measurement at time [time]. To adapt the sampling time step to the dynamic characteristics of the environment, a 1-second time step is used for dynamic environments and a 5-second time step is used for stable environments. This is an optimized estimate of the temperature. This is the corrected air pressure value. This is the temperature compensation coefficient, with a standard condition of 293K. This is the pressure compensation coefficient, with a standard atmospheric pressure of 101.325 kPa.

[0124] B4. Based on the oxygen concentration change rate compensated for by temperature and air pressure, extract statistical features within the time window, including mean, standard deviation, first difference, and autocorrelation coefficient, to extract time series features.

[0125] In this invention, the autocorrelation coefficient is used to measure the correlation of data at different time points in a time series, and its calculation formula is as follows:

[0126]

[0127] in: The autocorrelation coefficient is calculated by taking time points at 30-minute intervals within the time window. For the first time window Rate of change of oxygen concentration at each time point, compensated for by temperature and air pressure This represents the average rate of change in oxygen concentration compensated for by temperature and air pressure within the time window. This represents the standard deviation of the rate of change in oxygen concentration compensated for by temperature and air pressure within the time window.

[0128] B5. Calculate the combined effect index of temperature and humidity to extract meteorological characteristics, expressed as:

[0129]

[0130] in: As an index of the combined effect of temperature and humidity, This refers to relative humidity.

[0131] S3. Construct an oxygen concentration prediction model based on a bidirectional LSTM network and a Transformer model, and train the oxygen concentration prediction model using a near-end policy optimization algorithm and multimodal feature data to obtain the trained oxygen concentration prediction model.

[0132] In an optional embodiment of the present invention, the oxygen concentration prediction model includes a bidirectional LSTM network, a Transformer encoder, and a fully connected output layer connected in sequence.

[0133] The bidirectional LSTM network consists of a forward LSTM layer, a backward LSTM layer, and a concatenation layer. The bidirectional LSTM network is used to input multimodal feature data. It utilizes a forward LSTM layer (with a hidden layer dimension of 128, meaning 128 hidden units per layer) and a backward LSTM layer (with a hidden layer dimension of 128, meaning 128 hidden units per layer) to process the multimodal feature data, obtaining the hidden state vectors of the forward and backward LSTM layers respectively. Finally, the concatenation layer concatenates the hidden state vectors of the forward and backward LSTM layers to obtain the temporal feature vector of the multimodal feature data.

[0134] The Transformer encoder consists of an input embedding layer, a multi-head self-attention layer with four attention heads, a feedforward network, and a Transformer layer structure containing three Transformer blocks, connected in sequence. The Transformer encoder takes over the temporal feature vectors of multimodal feature data. It uses the input embedding layer (with both embedding and model dimensions of 64) to project the temporal feature vectors of multimodal feature data into a higher-dimensional space to obtain multi-dimensional temporal feature vectors. It uses the multi-head self-attention layer with four attention heads to capture the long-term dependencies between multi-dimensional temporal feature vectors to obtain multi-dimensional fused temporal feature vectors. It uses the feedforward network to perform nonlinear transformation on the multi-dimensional fused temporal feature vectors to obtain temporal feature interaction vectors. Finally, it uses the Transformer layer structure containing three Transformer blocks to encode the temporal feature interaction vectors to obtain the encoded feature vectors (64 dimensions).

[0135] Specifically, each Transformer block has residual connections and layer normalization to stabilize the training process.

[0136] The fully connected output layer consists of an input layer, a hidden layer, and a linear output layer connected in sequence; the fully connected output layer is used to map the encoded feature vector to control parameters (such as PID control parameters) to generate oxygen concentration prediction values.

[0137] The fully connected output layer consists of an input layer, a hidden layer, and a linear output layer connected in sequence. The input layer is used to input the encoded feature vector; the hidden layer (128 neurons, using the LeakyReLU activation function) is used to increase the non-linear expressive power of the oxygen concentration prediction model; the linear output layer (output dimension 3) is used to generate the predicted oxygen concentration value, which is then processed by the Softplus function.

[0138] The oxygen concentration prediction model is trained using a proximal policy optimization algorithm and multimodal feature data to obtain the trained oxygen concentration prediction model, including steps C1-C6, as follows:

[0139] C1. Calculate the advantage function estimate of the oxygen concentration prediction model, expressed as:

[0140]

[0141] in: For oxygen concentration prediction models in environmental conditions Execute control actions The advantage function, This refers to the environmental conditions, specifically the combination of oxygen concentration, temperature, humidity, and air pressure. To control the action, specifically to adjust the opening degree of the ventilation equipment. In environmental conditions Execute control actions The expected cumulative reward, In environmental conditions The benchmark value below.

[0142] C2. Determine the strategy gradient pruning based on the advantage function estimation of the oxygen concentration prediction model, expressed as:

[0143]

[0144] in: For the clipping function, The probability ratio between the old and new strategies. This is the clipping threshold.

[0145] C3. Calculate the policy loss of the oxygen concentration prediction model based on the dominance function and policy gradient pruning, expressed as:

[0146]

[0147] in: The strategy loss of the oxygen concentration prediction model. The sign for taking the minimum value.

[0148] C4. Calculate the value loss of the oxygen concentration prediction model, expressed as:

[0149]

[0150] in: The value loss of the current strategy for the oxygen concentration prediction model. For the current strategy in the environment state The following value estimate, For the old strategy in the environment state The target value is calculated from the experience replay cache.

[0151] C5. Based on the strategy loss and value loss of the oxygen concentration prediction model, calculate the total loss of the oxygen concentration prediction model, expressed as:

[0152]

[0153] in: The total loss of the current strategy for the oxygen concentration prediction model is given. , All are weighted coefficients, determined through grid search combined with cross-validation, in the context of an underground mine. Take 0.5, Take 0.01, This is the policy entropy regularization term.

[0154] In this invention, the strategy entropy regularization term is calculated as follows:

[0155]

[0156] in: for Controlling actions at all times For policy networks in environmental state Downstream control action The probability, It is a logarithmic function.

[0157] C6. Train the oxygen concentration prediction model based on the total loss of the oxygen concentration prediction model to obtain the trained oxygen concentration prediction model.

[0158] Specifically, to enable the oxygen concentration prediction model to adapt to long-term environmental changes, this invention deploys an online learning module. An experience playback mechanism and a real-time update algorithm are employed to achieve real-time updates of the oxygen concentration prediction model parameters. When the system encounters new environmental data, it can promptly adjust the oxygen concentration prediction model to adapt to constantly changing environmental conditions. The online learning module continuously optimizes the model parameters, ensuring the long-term stability and effectiveness of oxygen concentration regulation.

[0159] S4. Obtain real-time multimodal environmental parameters, and use the real-time multimodal environmental parameters and the trained oxygen concentration prediction model to obtain the oxygen concentration prediction value.

[0160] In an optional embodiment of the present invention, the present invention acquires real-time multimodal environmental parameters, then processes them according to step S2 to acquire real-time multimodal feature data, and then inputs the real-time multimodal feature data into the trained oxygen concentration prediction model to obtain oxygen concentration prediction values.

[0161] S5. Construct an adaptive proportional-integral-derivative (PID) controller, use the oxygen concentration prediction value and the PID controller to generate control commands for the ventilation equipment, and send the control commands to the ventilation equipment for execution to regulate the oxygen concentration.

[0162] In an optional embodiment of the present invention, the present invention constructs an adaptive proportional-integral-derivative controller, denoted as:

[0163] ,

[0164] ,

[0165] ,

[0166] ,

[0167]

[0168] in: for The control output at any given time is specifically as follows: The instructions for adjusting the opening degree of the ventilation equipment at all times. ,express Timing control error, for Predicted oxygen concentration at time [time]. To preset the target oxygen concentration, , , These are the proportional gain, integral gain, and derivative gain, respectively, with initial values ​​set empirically. , , These are the proportional gain increment, integral gain increment, and differential gain increment, which are dynamically adjusted through the online learning module, respectively. To control the sensitivity to changes in current error, To suppress historical noise interference.

[0169] As shown in Figure 2, the predicted oxygen concentration strongly demonstrates the superior performance of this control method. The rapid response and accurate tracking of the actual value to the target value indicate that, thanks to the significant reduction in the standard deviation of the control error from ±2.5% of the traditional method to ±0.8%, and the effective compression of the system response time to less than 3 seconds, the control method possesses impressively high control accuracy and stability. The control error curve in the lower half of Figure 2 clearly depicts the dynamic trajectory of the error. The error is relatively large in the initial stage, but thanks to the efficiency and adaptive capability of the control strategy, the error quickly and rapidly decreases and stabilizes within a very small range.

[0170] like Figure 3 As shown, the state estimation performance is clearly demonstrated. The comparison between the actual state, estimated value and observed value of oxygen concentration clearly shows that the estimated value accurately tracks the actual state. The root mean square error of the state estimation is significantly reduced by 62%, while the observed value, despite noise interference, still maintains a high degree of consistency with the actual state. This verifies the high accuracy and high reliability of the Kalman filter in this method for state estimation.

[0171] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0172] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0173] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0174] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0175] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for regulating oxygen concentration based on a hybrid intelligent model, characterized in that, The method comprises the following steps: Collecting multi-modal environmental parameters through a multi-modal sensor network; Optimizing and estimating the multi-modal environmental parameters by using a Kalman filter to obtain optimized estimation values of the multi-modal environmental parameters, and performing feature extraction on the optimized estimation values of the multi-modal environmental parameters to obtain multi-modal feature data; Building an oxygen concentration prediction model based on a bidirectional LSTM network and a Transformer model, and training the oxygen concentration prediction model by using a proximal policy optimization algorithm and the multi-modal feature data to obtain a trained oxygen concentration prediction model; Obtaining real-time multi-modal environmental parameters, and using the real-time multi-modal environmental parameters and the trained oxygen concentration prediction model to obtain an oxygen concentration prediction value; Building an adaptive proportional-integral-derivative controller, generating a ventilation equipment control instruction by using the oxygen concentration prediction value and the adaptive proportional-integral-derivative controller, and sending the ventilation equipment control instruction to a ventilation equipment for execution to regulate and control the oxygen concentration; The adaptive proportional-integral-derivative controller is represented as: ， ， ， ， Wherein: is the control output at the moment, specifically the ventilation equipment opening degree instruction at the moment, represents the control error at the moment, is the oxygen concentration prediction value at the moment, is the preset target oxygen concentration, , , respectively, the proportional gain, the integral gain and the differential gain, the initial value is set by experience, , , respectively, the proportional gain increment, the integral gain increment and the differential gain increment dynamically adjusted by the online learning module, is the sensitivity of controlling the current error change, is the historical noise interference suppression.

2. The hybrid intelligent model-based oxygen concentration regulation method according to claim 1, wherein, The multi-modal environmental parameters include oxygen concentration, temperature, humidity, and air pressure.

3. The hybrid intelligent model-based oxygen concentration regulation method of claim 1, wherein, The optimizing and estimating of the multi-modal environmental parameters by using the Kalman filter comprises the following steps: Differentially modeling the multi-modal environmental parameters to configure a state transition model of the multi-modal environmental parameters; Predicting and updating the state transition model of the multi-modal environmental parameters by using the Kalman filter to obtain the optimized estimation values of the multi-modal environmental parameters.

4. The hybrid intelligent model-based oxygen concentration regulation method according to claim 3, wherein, The differentially modeling of the multi-modal environmental parameters to configure the state transition model of the multi-modal environmental parameters is specifically: Modeling the oxygen concentration by using a nonlinear state equation to configure a state transition model of the oxygen concentration, which is represented as: wherein: is the optimized estimate of the oxygen concentration at time instant is a non-linear mapping function, is the optimized estimate of the oxygen concentration at time instant is the optimized estimate of the temperature at time instant is the optimized estimate of the air pressure at time instant is the process noise at time instant Modeling the temperature and the humidity by using linear observation models to respectively configure state transition models of the temperature and the humidity, which are represented as: wherein: is an observation matrix for temperature, is is an optimized estimate of temperature at time, is is observation noise for temperature at time, is is an optimized estimate of humidity at time, is an observation matrix for humidity, is is an optimized estimate of humidity at time, is is observation noise for humidity at time; Modeling the air pressure by using a simplified dynamic model to configure a state transition model of the air pressure, which is represented as: wherein: is an optimized estimate of the air pressure at the moment, is the air pressure reference change amount, is is the simplified noise at the moment.

5. The hybrid intelligent model-based oxygen concentration regulation method of claim 1, wherein, The multi-modal feature data includes air pressure periodicity features, time series features, and meteorological features.

6. The hybrid intelligent model-based oxygen concentration regulation method according to claim 5, wherein, The feature extraction on the multi-modal environmental parameters after the optimizing and estimating comprises the following steps: Correcting the optimized estimation value of the air pressure according to the relationship between the air pressure and the altitude to obtain a corrected air pressure value, which is represented as: wherein: is the corrected air pressure value, is the optimized estimate of air pressure, is the height difference relative to the reference altitude, is the atmospheric elevation constant; Extracting air pressure periodicity features based on the corrected air pressure value, including 24-hour amplitude and 1-hour phase components; Calculating a temperature and air pressure compensated oxygen concentration change rate based on the corrected air pressure value, which is represented as: wherein: is the oxygen concentration measurement value at the moment, is a sampling time step length adapted according to the environmental dynamic characteristics, is an optimized estimation value of the temperature, is a corrected air pressure value, is a temperature compensation coefficient, is an air pressure compensation coefficient; Extracting time series features based on the temperature and air pressure compensated oxygen concentration change rate, including mean value, standard deviation, first-order difference, and autocorrelation coefficient; Calculating a temperature and humidity joint effect index to extract meteorological features, which is represented as: wherein: is the combined effect of temperature and humidity, is the relative humidity.

7. The hybrid intelligent model-based oxygen concentration regulation method of claim 1, wherein, The oxygen concentration prediction model comprises a bidirectional LSTM network, a Transformer encoder, and a fully connected output layer connected in sequence; The bidirectional LSTM network comprises a forward LSTM layer, a backward LSTM layer and a splicing layer; the bidirectional LSTM network is used for inputting the multi-modal feature data, and the multi-modal feature data is processed by the forward LSTM layer and the backward LSTM layer respectively to obtain a hidden state vector of the forward LSTM layer and a hidden state vector of the backward LSTM layer, and the hidden state vector of the forward LSTM layer and the hidden state vector of the backward LSTM layer are spliced by the splicing layer to obtain a time sequence feature vector of the multi-modal feature data; The Transformer encoder comprises an input embedding layer, a multi-head self-attention layer comprising four attention heads, a feedforward network and a Transformer layer structure comprising three Transformer blocks connected in sequence; the Transformer encoder is used for receiving the time sequence feature vector of the multi-modal feature data, projecting the time sequence feature vector of the multi-modal feature data to a higher dimensional space by the input embedding layer to obtain a multi-dimensional time sequence feature vector, capturing long-term dependencies between the multi-dimensional time sequence feature vectors by the multi-head self-attention layer comprising four attention heads to obtain a multi-dimensional fused time sequence feature vector, performing nonlinear transformation on the multi-dimensional fused time sequence feature vector by the feedforward network to obtain a time sequence feature interaction vector, and encoding the time sequence feature interaction vector by the Transformer layer structure comprising three Transformer blocks to obtain an encoded feature vector; The fully connected output layer comprises an input layer, a hidden layer and a linear output layer connected in sequence; the fully connected output layer is used for mapping the encoded feature vector to the control parameter to generate the oxygen concentration prediction value.

8. The hybrid intelligent model-based oxygen concentration regulation method of claim 1, wherein, The oxygen concentration prediction model is trained by using the proximal policy optimization algorithm and the multi-modal feature data to obtain a trained oxygen concentration prediction model, comprising the following steps: The advantage function estimate of the oxygen concentration prediction model is calculated and is represented as: wherein: is an oxygen concentration prediction model at an environment state performs a control action is an advantage function, is an environment state, specifically a combination of oxygen concentration, temperature, humidity, and air pressure, is a control action, specifically adjusting an opening degree of a ventilation device, is an expected cumulative reward of performing a control action at an environment state is a baseline value at an environment state is an environment state, is a baseline value at an environment state; The policy gradient clipping is determined according to the advantage function estimate of the oxygen concentration prediction model and is represented as: wherein: is a clipping function, is a probability ratio of new and old policies, is a clipping threshold; The policy loss of the oxygen concentration prediction model is calculated according to the advantage function of the oxygen concentration prediction model and the policy gradient clipping and is represented as: wherein: is the strategy loss for the oxygen concentration prediction model, is the min function. The value loss of the oxygen concentration prediction model is calculated and is represented as: wherein: is a value loss for the current policy of the oxygen concentration prediction model, is a value estimate for the current policy in the environment state , is a target value for the old policy in the environment state , computed from the experience replay buffer; The total loss of the oxygen concentration prediction model is calculated according to the policy loss of the oxygen concentration prediction model and the value loss of the oxygen concentration prediction model and is represented as: wherein: is the total loss of the current policy of the oxygen concentration prediction model, , are weighting coefficients, the weighting coefficients are determined by grid search combined with cross-validation, in the underground mine scene, takes 0.5, takes 0.01, is the policy entropy regularization term; The oxygen concentration prediction model is trained based on the total loss of the oxygen concentration prediction model to obtain the trained oxygen concentration prediction model.

Images