A subway air-conditioning environment mode detection system and detection method

Abstract

A subway air-conditioning environment mode detection system and detection method. The system includes an air-conditioning system arranged in the public area environment of the subway station, and also includes a multi-type sensor acquisition system, an air-conditioning system host computer, a neural network environment mode detection model, a K210 module and a display module; the detection method includes the following steps: selecting key variables; collecting environmental data; data preprocessing; constructing a simple model; determining the number of neurons in the hidden layer of the model; algorithm optimization to establish the best model. The present invention enhances the stability of the backpropagation neural network by optimizing the BP neural network model and using the Adam optimization algorithm. The model processing result is fed back to the air-conditioning system host computer, and the air-conditioning system host computer issues the current mode instruction to control the combined linkage among various fans, dampers and air-conditioning units, realizes the adjustment of the in-station environment, and effectively improves the timeliness, intelligence and detection efficiency of the subway air-conditioning system's environment mode detection.

Description

Technical Field

[0001] The present invention relates to a subway air - conditioning environment mode detection system and a detection method, belonging to the technical field of detection. Background Art

[0002] With the continuous development of the city, the subway has become one of the important components of urban transportation. Since the subway is constructed underground and the passenger flow in the station is dense, as a guarantee equipment for ventilation, refrigeration, and fire smoke exhaust, the air - conditioning system in the public area of the subway station plays an extremely important role.

[0003] At present, the operation of the fans, air valves, and air - conditioning rooms in the public - area air - conditioning system of the subway station adopts combined linkage control in a mode. The air - conditioning system works in different environment modes, and different working modes correspond to different combined working conditions of the air - conditioning system. However, the detection and switching of the air - conditioning system environment mode require manual judgment and participation in the selection and adjustment of the working mode according to the operation environment. Since the staff cannot adjust the air - conditioning environment mode in real - time and quickly according to the on - site environment in the subway station, there are problems of untimely, low - efficiency, and non - intelligent detection of the environment mode and adjustment of the working mode of the in - station air - conditioning system. Summary of the Invention

[0004] Aiming at the problems existing in the above - mentioned prior art, the present invention provides a subway air - conditioning environment mode detection system and a detection method. The detection system and method can effectively improve the timeliness, intelligence, and detection efficiency of the environment mode detection of the subway air - conditioning system, and then realize the real - time and high - efficiency adjustment of the working mode of the air - conditioning system, and improve the subway operation environment.

[0005] To achieve the above object, the present invention provides a subway air - conditioning environment mode detection system, which includes an air - conditioning system arranged in the public - area environment of the subway station, and also includes a multi - type sensor acquisition system, an air - conditioning system host computer, a neural - network environment mode detection model, a K210 module, and a display module;

[0006] The multi - type sensor acquisition system includes a plurality of distributed smoke sensors, carbon dioxide sensors, and temperature sensors. The smoke sensors collect the smoke concentration data in the station in real - time and send the collected smoke concentration data signals to the air - conditioning system host computer in real - time. The carbon dioxide sensors collect the carbon dioxide concentration data in the station in real - time and send the collected carbon dioxide concentration data signals to the air - conditioning system host computer in real - time. The temperature sensors collect the temperature data in the station in real - time and send the collected temperature data signals to the air - conditioning system host computer in real - time;

[0007] The neural - network environment mode detection model is deployed on the K210 module, and the K210 module communicates with the air - conditioning system host computer through a serial port;

[0008] The air conditioning system includes a fan, a damper, and an air handling unit, which are connected to the output end of the upper computer of the air conditioning system;

[0009] The display module is connected to the output end of the K210 module;

[0010] The upper computer of the air conditioning system is used to receive in real time the smoke concentration data signal, carbon dioxide concentration data signal, and temperature data signal transmitted by the multi-class sensor acquisition system, send the received data signal to the K210 module in real time, and process it by the neural network environment mode detection model. The processed signal is fed back to the upper computer of the air conditioning system. The upper computer of the air conditioning system conveys the environmental mode control instruction to the air conditioning system according to the received feedback signal, and starts the combined linkage of the fan, damper, and air handling unit under the corresponding environmental mode control to adjust the smoke concentration, carbon dioxide concentration, and temperature in the station; At the same time, the K210 module outputs the environmental mode detection result to the display module for display.

[0011] Further, the K210 module is an embedded AI development module, with a 64-bit dual-core processor and a general neural network accelerator KPU built into the chip. It has 8M 64-bit SRAM, a dedicated external FLASH interface, and a TF card to increase its own storage space. At the same time, it is equipped with an FPIOA field programmable IO array, supporting mainstream deep learning frameworks such as TensorFlow, Keras, and Caffe based on the Python language.

[0012] Further, it also includes a voice module, which is connected to the output end of the K210 module and is used to output the environmental mode detection result for voice broadcast.

[0013] A method for detecting the environmental mode of a subway air conditioner includes the following steps:

[0014] 1) Select key variables: According to the operating conditions of the subway air conditioning mode, select temperature, carbon dioxide concentration, and smoke concentration as the key environmental data for detection, that is, key variables;

[0015] 2) Collect environmental data: Collect the key environmental data in the subway station and the ultra-low and ultra-high environmental data that cannot be obtained in the subway station in the laboratory as experimental data for training and testing the model;

[0016] 3) Data preprocessing: Classify the collected experimental data into standby mode, ventilation mode, refrigeration mode, ventilation and refrigeration mode, smoke exhaust mode, and fire mode, and label them as 0, 1, 2, 3, 4, 5 respectively by using category labels, and then normalize the classified and labeled data;

[0017] 4) Construct a simple model: Determine the input layer and output layer of the subway environment mode neural network according to the collected environmental data categories and air-conditioning environment mode categories;

[0018] 5) Determine the number of neurons in the hidden layer of the model: Determine the range of the number of neurons in the hidden layer according to the formula where h is the number of neurons in the hidden layer, m and n are the number of neurons in the input layer and output layer respectively, and a is an adjustment constant between 1 and 10;

[0019] a. Train the model on the entire data set using the backpropagation and gradient descent algorithms. The formula is:

[0020] a [1](i) = g(W [1] X (i) + b [1] );

[0021]

[0022] In the formula, a [1](i) represents the output of the i-th sample in the hidden layer, W [1] represents the network weight between the input layer and the hidden layer, X (i) represents the i-th input sample, and b [1] represents the bias of the hidden layer neurons;

[0023] a [2](i) represents the predicted output of the i-th sample in the output layer, and is represented by the predicted value , W [2] represents the weight between the hidden layer and the output layer, and b [2] represents the bias of the output layer neurons;

[0024] The above parameters all represent matrix vectors;

[0025] b. To make the BP neural network approximate any non-linear function, select Relu as the activation function for the hidden layer. To make the neural network perform multi-task classification, select Softmax as the activation function for the output layer. The formulas are respectively:

[0026] g(x) = max(0, x);

[0027]

[0028] In the formula, g(x) represents the activation function of the hidden layer, and f(x) represents the activation function of the output layer; C = n represents the number of neurons in the output layer, and e x represents the linear exponential output of each neuron, and x is a vector of [n, 1];

[0029] c. The cost function between the actual output and the expected output is:

[0030]

[0031] m represents the total number of samples, and n represents the number of neurons in the output layer. represents the predicted value of the j-th class under the i-th sample. represents the true value of the j-th class under the i-th sample.

[0032] d. Use the preprocessed dataset to train neural networks with different numbers of neurons in the hidden layer. Set the number of epochs for the loop, and observe the change in the cost error of the training set samples through experiments; determine the number of neurons in the hidden layer corresponding to the curve with the minimum cost error as the number of neurons in the hidden layer of the model.

[0033] 6) Algorithm optimization to establish the best model: Adopt the Adam optimization algorithm, which dynamically adjusts the learning rate of each parameter with the first-order moment estimate and second-order moment estimate of the gradient, introduces momentum and adaptive learning rate to enable the algorithm to accelerate model learning, improve network recognition accuracy, and reduce oscillations during convergence; finally, use the optimized model as the best detection model for the subway air-conditioning environment mode.

[0034] Further, in step 3), Min-Max deviation normalization is used to normalize the data, so that the values of each collected data result are mapped between [0, 1]. The Min-Max normalization transformation function is defined as follows:

[0035]

[0036] In the formula, Max is the maximum value of the sample data, and Min is the minimum value of the sample data.

[0037] Further, in step 4), a 3-layer BP neural network is selected for modeling. The ambient temperature, carbon dioxide concentration, and smoke concentration are selected as the key variables for subway environment mode detection, that is, the category of the collected environmental data is 3, and the number of neurons in the input layer of the model is determined to be 3; the subway air-conditioning environment mode categories are ventilation mode, refrigeration mode, ventilation and refrigeration mode, smoke exhaust mode, and fire mode, and the number of neurons in the output layer of the model is determined to be 6.

[0038] Further, in step 6), taking the t-th iteration of gradient descent as an example, the network weight update process of the Adam optimization algorithm is as follows:

[0039] t = t + 1;

[0040]

[0041]

[0042] In the formula, m tand v t are the first-order momentum term and the second-order momentum term respectively. β1 and β2 are hyperparameters with default values of 0.9 and 0.999, and m t and v t 's bias correction values are and Their calculation formulas are respectively:

[0043]

[0044]

[0045] Update the weights and biases. The weight W [l] t and the bias at the t-th iteration are:

[0046]

[0047]

[0048] In the above formula, W [l] t-1 represents the model weight at the (t - 1)-th iteration, and b [l] t-1 represents the model bias at the (t - 1)-th iteration. α is the learning rate with a default value of 0.001, and ε is taken as 10 -8 .

[0049] By optimizing the BP neural network model and using the Adam optimization algorithm, the present invention enhances the stability of the backpropagation neural network. The model processing result is fed back to the upper computer of the air-conditioning system, and the upper computer of the air-conditioning system issues the current mode instruction to control the combined linkage among various fans, dampers, and air-conditioning units, realizing the adjustment of the station environment, effectively improving the timeliness, intelligence, and detection efficiency of the environmental mode detection of the subway air-conditioning system, and further realizing the real-time and high efficiency of the working mode adjustment of the air-conditioning system, thus improving the subway operation environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is the schematic diagram of the subway air-conditioning environment mode detection and control of the present invention;

[0051] Figure 2 is the flow chart of the subway air-conditioning environment mode detection of the present invention;

[0052] Figure 3 is the subway environment mode classification label diagram of the present invention;

[0053] Figure 4 is the experimental relationship curve diagram of the number of hidden layer neurons and the cost error of the present invention;

[0054] Figure 5 It is a flowchart for determining the best model for detecting the subway air-conditioning environment mode of the present invention;

[0055] Figure 6 The model performance graph before optimization in the embodiment of the present invention;

[0056] Figure 7 It is the model performance graph after optimization in the embodiment of the present invention;

[0057] Figure 8 It is the structure diagram of the best subway air-conditioning environment mode detection model in the embodiment of the present invention. Detailed implementation manners

[0058] The present invention will be further described below with reference to the accompanying drawings.

[0059] As Figure 1 shown, a subway air-conditioning environment mode detection system includes an air-conditioning system arranged in the public area environment of the subway station, and also includes a multi-class sensor acquisition system, an air-conditioning system host computer, a neural network environment mode detection model, a K210 module, and a display module;

[0060] The multi-class sensor acquisition system includes a plurality of distributed smoke sensors, carbon dioxide sensors, and temperature sensors. The smoke sensors collect the smoke concentration data in the station in real time and send the collected smoke concentration data signals to the air-conditioning system host computer in real time. The carbon dioxide sensors collect the carbon dioxide concentration data in the station in real time and send the collected carbon dioxide concentration data signals to the air-conditioning system host computer in real time. The temperature sensors collect the temperature data in the station in real time and send the collected temperature data signals to the air-conditioning system host computer in real time;

[0061] The neural network environment mode detection model is deployed on the K210 module, and the K210 module communicates with the air-conditioning system host computer through a serial port;

[0062] The air-conditioning system includes a fan, a damper, and an air-conditioning unit, which are connected to the output end of the air-conditioning system host computer;

[0063] The display module is connected to the output end of the K210 module;

[0064] The host computer of the air conditioning system is used to receive in real time the smoke concentration data signal, carbon dioxide concentration data signal, and temperature data signal transmitted by the multi-class sensor acquisition system, and send the received data signal to the K210 module in real time for processing by the neural network environmental mode detection model. The processed signal is fed back to the host computer of the air conditioning system, and the host computer of the air conditioning system conveys the environmental mode control instruction to the air conditioning system according to the received feedback signal, and starts the combined linkage of the fan, air valve, and air conditioning unit under the corresponding environmental mode control to adjust the smoke concentration, carbon dioxide concentration, and temperature in the station; at the same time, the K210 module outputs the environmental mode detection result to the display module for display.

[0065] Further, the K210 module is an embedded AI development module, with a 64-bit dual-core processor and a general neural network accelerator KPU built into the chip, having 8M 64-bit SRAM, a dedicated external FLASH interface, and a TF card to increase its own storage space. At the same time, it is equipped with an FPIOA field programmable IO array, supporting mainstream deep learning frameworks such as TensorFlow, Keras, and Caffe based on the Python language.

[0066] Further, it also includes a voice module, and the voice module is connected to the output end of the K210 module for outputting the environmental mode detection result for voice broadcast.

[0067] As Figure 2 shown, a method for detecting the environmental mode of a subway air conditioner includes the following steps:

[0068] 1) Select key variables: According to the operating conditions of the subway air conditioning mode, select temperature, carbon dioxide concentration, and smoke concentration as the key environmental data for detection, that is, key variables;

[0069] 2) Collect environmental data: Collect the key environmental data in the subway station and the ultra-low and ultra-high environmental data that cannot be obtained in the subway station in the laboratory as experimental data for training and testing the model;

[0070] 3) Data preprocessing: Classify the collected experimental data into standby mode, ventilation mode, refrigeration mode, ventilation and refrigeration mode, smoke exhaust mode, and fire mode, and label them as 0, 1, 2, 3, 4, 5 respectively through the use of category labels, and then perform normalization processing on the classified and labeled data, as Figure 3 shown;

[0071] 4) Build a simple model: Determine the input layer and output layer of the subway environmental mode neural network according to the collected environmental data categories and air conditioning environmental mode categories;

[0072] 5) As Figure 4As shown, determine the number of neurons in the hidden layer of the model: According to the formula Determine the range of the number of neurons in the hidden layer. In the formula, h is the number of neurons in the hidden layer, m and n are the number of neurons in the input layer and the output layer respectively, and a is a regulation constant between 1 and 10;

[0073] a. Use the backpropagation and gradient descent algorithms to train the model on the entire dataset. The formula is:

[0074] a [1](i) = g(W [1] X (i) + b [1] );

[0075]

[0076] In the formula, a [1](i) represents the output of the i-th sample in the hidden layer, W [1] represents the network weight between the input layer and the hidden layer, X (i) represents the i-th input sample, and b [1] represents the bias of the hidden layer neurons;

[0077] a [2](i) represents the predicted output of the i-th sample in the output layer, and is represented by the predicted value W [2] represents the weight between the hidden layer and the output layer, and b [2] represents the bias of the output layer neurons;

[0078] All of the above parameters represent matrix vectors;

[0079] b. To make the BP neural network approximate any non-linear function, select Relu as the activation function of the hidden layer, and to make the neural network perform multi-task classification, select Softmax as the activation function of the output layer. The formulas are respectively:

[0080] g(x) = max(0, x);

[0081]

[0082] In the formula, g(x) represents the activation function of the hidden layer, and f(x) represents the activation function of the output layer; C = n represents the number of neurons in the output layer, and e x represents the linear exponential output of each neuron, and x is a vector of [n, 1];

[0083] c. The cost function between the actual output and the expected output is:

[0084]

[0085] m represents the total number of samples, and n represents the number of neurons in the output layer. represents the predicted value of the j-th class under the i-th sample. represents the true value of the j-th class under the i-th sample;

[0086] d. Use the preprocessed dataset to train neural networks with different numbers of neurons in the hidden layer. Set the number of epochs for the loop, and observe the change in the cost error of the training set samples through experiments; determine the number of neurons in the hidden layer corresponding to the curve with the minimum cost error as the number of neurons in the hidden layer of the model;

[0087] 6) As Figure 5 shown, optimize the algorithm and establish the best model: Use the Adam optimization algorithm to dynamically adjust the learning rate of each parameter with the first-order moment estimate and second-order moment estimate of the gradient, introduce momentum and adaptive learning rate to make the algorithm accelerate model learning, improve network recognition accuracy, and reduce oscillations when converging; finally, use the optimized model as the best detection model for the subway air-conditioning environment mode.

[0088] Furthermore, in step 3), use Min-Max deviation normalization to normalize the data, so that the collected data result values are mapped between [0, 1]. The Min-Max normalization transformation function is defined as follows:

[0089]

[0090] In the formula, Max is the maximum value of the sample data, and Min is the minimum value of the sample data.

[0091] Furthermore, in step 4), select a 3-layer BP neural network for modeling. Select ambient temperature, carbon dioxide concentration, and smoke concentration as the key variables for subway environment mode detection, that is, the category of the collected environmental data is 3, and determine the number of neurons in the input layer of the model to be 3; the subway air-conditioning environment mode categories are ventilation mode, refrigeration mode, ventilation and refrigeration mode, smoke exhaust mode, and fire mode, and determine the number of neurons in the output layer of the model to be 6.

[0092] Furthermore, in step 6), taking the t-th iteration of gradient descent as an example, the network weight update process of the Adam optimization algorithm is as follows:

[0093] t = t + 1;

[0094]

[0095]

[0096] In the formula, m t and v tThey are the first-order momentum term and the second-order momentum term respectively. β1 and β2 are hyperparameters with default values of 0.9 and 0.999, and m t and v t The bias correction values of are and Their calculation formulas are respectively:

[0097]

[0098]

[0099] Update the weights and biases. The weight W at the t-th iteration [l] t and the bias are:

[0100]

[0101]

[0102] In the above formula, W [l] t-1 represents the model weight at the (t - 1)-th iteration, and b [l] t-1 represents the model bias at the (t - 1)-th iteration. α is the learning rate with a default value of 0.001, and ε is taken as 10 -8 .

[0103] Example:

[0104] Using the Python language and the TensorFlow learning framework, build a neural network model on the Pycharm software. First, determine that the number of input layer neurons of the neural network is 3 according to the input environmental parameter variables of temperature, carbon dioxide concentration, and smoke concentration data. According to the 6 types of modes of the subway environment air conditioner, determine that the number of output layer neurons is 6. The number of neurons in the middle hidden layer is determined according to the formula to determine that the range of the number of hidden layer neurons is [4, 13]. Then, build neural network models with different network structures such as 3 - 4 - 6, 3 - 5 - 6, 3 - 6 - 6... 3 - 13 - 6 through experiments. Set the same number of loops to 500 rounds in the experiment, use the gradient descent method for error convergence, calculate the error loss of each network structure through the cost function and display it in the form of a curve, record the experimental data and compare them. As Figure 4 shown: When the number of hidden layers is 11, the corresponding loss of the network model is the smallest, which is 0.023, and determine the model structure as 3 - 11 - 6;

[0105] Optimize this model using an optimization algorithm, such as Figure 6 and Figure 7As shown in the figure, experimental comparative analysis was carried out on the two models before and after optimization. Table 1 is the experimental design table. A total of 4 groups of experiments were designed, and the training data and test data were divided into 2:1, specifically as follows:

[0106] Table 1

[0107]

[0108] The comparison of the experimental results of the models before and after optimization is shown in Table 2 as follows:

[0109] Table 2

[0110]

[0111] It can be seen from Table 2 that the number of prediction errors (M), accuracy rate (ACC), and MacroF1 macro-average of the two models before and after experimental calculation optimization were compared. As the number of training sets and test sets increased, the accuracy rates of both models increased. The average test accuracy rate of the traditional BP neural network was 96.89%, and the average accuracy rate of the optimized BP neural network model was 97.99%. The accuracy rate of the model increased by 1.1%; the MacroF1 average value of the optimized BP neural network increased from 96.42% to 97.63%, and the model accuracy and stability were improved; the average number of prediction errors of the optimized model decreased by 45.61% compared with that before optimization. The optimized BP neural network model with the same experimental number was superior to the traditional BP neural network model in terms of the number of prediction errors, accuracy rate, and stability;

[0112] As shown in Table 3 is the comparison of the experimental performance of the multi-classification model:

[0113] Table 3

[0114]

[0115] The model designed in the present invention is AD-BP in the table. The dataset with the experiment number 4 in Table 1 is selected as the experimental data. Through experiments, a multi-faceted comparative analysis is carried out on the traditional machine learning multi-classification model and the BP neural network models before optimization (BP) and after optimization (AD-BP). As shown in the experimental results in Table 3, the running detection times of the Naive Bayes model and the Logistic Regression model are relatively small, which are 1.2 ms and 4.9 ms respectively, but their accuracies are relatively low, which are 96.91% and 94.57% respectively; the accuracy of the Random Forest model is the highest at 99.94%, and its stability is the best, but the running detection time is relatively long at 35.9 ms; the accuracy of the optimized AD-BP neural network model is relatively high at 99.88%, second only to 99.94% of the Random Forest model, and the running test time is relatively small at 12 ms. It combines the advantages of high precision and short running time, and generally has better performance. As the training set data increases, the accuracy of the AD-BP neural network model will be further improved, and the running time of the model remains unchanged, and it has better effect on processing large datasets. As Figure 8 shown, as shown in the structure diagram of the best subway air-conditioning environment mode detection model provided by the present invention, through experimental comparison, finally the optimized model (3-11-6) is used as the best subway air-conditioning environment mode detection model.

Claims

1. A method for detecting the environmental mode of a subway air conditioner, characterized in that, The steps are as follows: 1) Select key variables: According to the operating conditions of the subway air-conditioning mode, select temperature, carbon dioxide concentration, and smoke concentration as the key environmental data for detection, that is, key variables; 2) Collect environmental data: Collect the key environmental data in the subway station and collect the ultra-low and ultra-high environmental data that cannot be obtained in the subway station in the laboratory as experimental data for training and testing the model; 3) Data preprocessing: Classify the collected experimental data into modes, including standby mode, ventilation mode, refrigeration mode, ventilation and refrigeration mode, smoke exhaust mode, and fire mode, and label them as 0, 1, 2, 3, 4, 5 respectively using category labels. Then perform normalization processing on the classified and labeled data; 4) Build a simple model: Determine the input layer and output layer of the subway environmental mode neural network according to the collected environmental data categories and air-conditioning environmental mode categories; 5) Determine the number of neurons in the hidden layer of the model: According to the formula , determine the range of the number of neurons in the hidden layer. In the formula, is the number of neurons in the hidden layer, and are the numbers of neurons in the input layer and the output layer respectively, is a regulation constant between 1 and 10; a. The model is trained on the entire dataset using backpropagation and the gradient descent algorithm, and the formula is: ; ; Wherein, represents the output of the i-th sample in the hidden layer, represents the network weight between the input layer and the hidden layer, represents the i-th input sample, represents the bias of the hidden layer neurons; represents the predicted output of the i-th sample in the output layer, and is represented by the predicted value ; represents the weight between the hidden layer and the output layer, represents the bias of the output layer neurons; the above parameters all represent matrix vectors; b. To make the BP neural network approximate any nonlinear function, select Relu as the activation function for the hidden layer. To make the neural network perform multi-task classification, select Softmax as the activation function for the output layer. The formulas are as follows: ; ; Wherein, represents the activation function of the hidden layer, represents the activation function of the output layer; C = n represents the number of neurons in the output layer, represents the linear exponential output of each neuron, is a vector of [n, 1]; c. The cost function between the actual output and the expected output is as follows: ; m represents the total number of samples, n represents the number of neurons in the output layer, represents the predicted value of the j-th class under the i-th sample, represents the true value of the j-th class under the i-th sample; d. Use the preprocessed data set to train neural networks with different numbers of hidden layer neurons, set the number of epochs, and observe the change of the cost error of the training set samples through experiments; determine the number of hidden layer neurons corresponding to the curve with the minimum cost error as the number of hidden layer neurons of the model; 6) Algorithm optimization to establish the best model: Adopt the Adam optimization algorithm, dynamically adjust the learning rate of each parameter with the first-order moment estimate and second-order moment estimate of the gradient, introduce momentum and adaptive learning rate to make the algorithm accelerate model learning, improve network recognition accuracy, and reduce oscillations when converging; finally, use the optimized model as the best detection model for the subway air-conditioning environmental mode; among them, taking the t-th iteration of gradient descent as an example, the network weight update process of the Adam optimization algorithm is: ; ; ; wherein, and are the first-order momentum term and the second-order momentum term respectively, and are hyperparameters with default values of 0.9 and 0.999, and have bias correction values of and , and their calculation formulas are respectively: ; ; Update the weights and biases. The weights and biases at the t-th iteration are: ; ; where represents the model weights at the (t - 1)-th iteration, represents the model biases at the (t - 1)-th iteration, is the learning rate, defaulting to 0.001, is taken as .

2. The subway air-conditioning environment mode detection method according to claim 1, wherein In step 3), Min-Max deviation normalization is used to normalize the data, so that the collected data result values are mapped between [0, 1]. The Min-Max normalization conversion function is defined as follows: ; where Max is the maximum value of the sample data and Min is the minimum value of the sample data.

3. A subway air-conditioning environment mode detection method according to claim 1, characterized in that In step 4), select a 3-layer BP neural network for modeling, select environmental temperature, carbon dioxide concentration, and smoke concentration as the key variables for subway environmental mode detection, that is, the category of the collected environmental data is 3, and determine the number of neurons in the input layer of the model to be 3; the categories of subway air-conditioning environmental modes are ventilation mode, refrigeration mode, ventilation and refrigeration mode, smoke exhaust mode, and fire mode, and determine the number of neurons in the output layer of the model to be 6.

4. The subway air-conditioning environment mode detection method according to claim 1, wherein The detection system it is based on includes an air-conditioning system installed in the public area environment of the subway station, and also includes a multi-class sensor collection system, an upper computer of the air-conditioning system, a neural network environmental mode detection model, a K210 module, and a display module; The multi-class sensor acquisition system described above includes multiple distributed smoke sensors, carbon dioxide sensors, and temperature sensors. The smoke sensors collect the smoke concentration data in the station in real time and send the collected smoke concentration data signals to the upper computer of the air conditioning system in real time. The carbon dioxide sensors collect the carbon dioxide concentration data in the station in real time and send the collected carbon dioxide concentration data signals to the upper computer of the air conditioning system in real time. The temperature sensors collect the temperature data in the station in real time and send the collected temperature data signals to the upper computer of the air conditioning system in real time; The neural network environment mode detection model is deployed on the K210 module, and serial communication is carried out between the K210 module and the upper computer of the air conditioning system; The air conditioning system includes a fan, a damper, and an air conditioning unit, which are connected to the output end of the upper computer of the air conditioning system; The display module is connected to the output end of the K210 module; The upper computer of the air conditioning system is used to receive the smoke concentration data signal, carbon dioxide concentration data signal, and temperature data signal transmitted by the multi-class sensor acquisition system in real time, send the received data signals to the K210 module in real time, and process them by the neural network environment mode detection model. The processed signal is fed back to the upper computer of the air conditioning system. The upper computer of the air conditioning system conveys the environmental mode control instruction to the air conditioning system according to the received feedback signal, and starts the combined linkage of the fan, damper, and air conditioning unit under the corresponding environmental mode control to adjust the smoke concentration, carbon dioxide concentration, and temperature in the station; At the same time, the K210 module outputs the environmental mode detection result to the display module for display.

5. A subway air-conditioning environment mode detection method according to claim 4, characterized in that, The K210 module described above is an embedded AI development module. The chip is built with a 64-bit dual-core processor and a general neural network accelerator KPU, has 8M 64bit SRAM, a dedicated external FLASH interface, and a TF card to increase its own storage space. At the same time, it is equipped with an FPIOA field programmable IO array, supporting mainstream deep learning frameworks such as TensorFlow, Keras, and Caffe based on the Python language.

6. A subway air-conditioning environment mode detection method according to claim 4 or 5, characterized in that It also includes a voice module, which is connected to the output end of the K210 module and is used to output the environmental mode detection result for voice broadcast.

Images