Method for improving performance of nitrogen-oxygen sensor

Abstract

The invention discloses a method for improving the performance of a nitrogen-oxygen sensor, relates to the field of sensors, and solves the technical problem of poor detection data caused by the fact that the performance of an existing nitrogen-oxygen sensor cannot be adjusted. The method comprises the following steps: S11, acquiring the nitrogen oxide concentration C0 of a reference gas source and the detection concentration C1 detected by a nitrogen-oxygen sensor, and comparing the nitrogen oxide concentration C0 with the detection concentration C1; s12, when the nitrogen oxide concentration C0 is equal to the detection concentration C1, judging that the performance of the nitrogen-oxygen sensor is normal; and when the nitrogen oxide concentration C0 is not equal to the detection concentration C1, correcting the nitrogen-oxygen sensor by calling a correction factor mu. According to the invention, the nitrogen-oxygen sensor is ensured to have high accuracy, the calibration can compensate the aging and fatigue of the sensor material along with the use time, and the service life of the sensor is prolonged.

Description

Technical Field

[0001] The present invention relates to the field of sensors, and more particularly to a method for improving the performance of a nitrogen oxygen sensor. Background Art

[0002] In order to solve the energy consumption and environmental pollution problems brought about by economic development, governments of various countries have formulated strict emission standards, especially in the automobile industry. Automobile engine exhaust is one of the main sources of air pollution. Automobile exhaust includes carbon monoxide (CO), hydrocarbons (HC), non-methane hydrocarbons (NMHC), nitrogen oxides (NO x ) and particulate matter (PM). In order to strictly control automobile engine exhaust pollution, my country has proposed the sixth stage of national motor vehicle pollutant emission standards, which are more stringent than the previous emission standards. x The emission requirements have been greatly improved. Nitrogen oxide sensors are particularly important in exhaust gas treatment. They can detect NO x The concentration of NO x It provides a reference for post-processing and monitoring.

[0003] At present, the emission treatment technologies for nitrogen oxides include NO x The SCR system has become the most promising NOx capture technology due to its high efficiency emission reduction, wide applicability, high efficiency, low energy consumption, stable and reliable operation, etc. x Treatment technology. The technical principle is to control the urea spraying unit to spray urea aqueous solution into the exhaust pipe through the electronic control unit (ECU). Urea is pyrolyzed and hydrolyzed at the exhaust temperature to release ammonia. Ammonia reacts with nitrogen oxides in the exhaust gas in the SCR catalyst to regenerate N2 and H2O, thereby reducing nitrogen oxide emissions. However, this technology requires the nitrogen oxide sensor to have good performance and the detected NOx concentration to be highly accurate. Otherwise, it will cause inaccurate urea injection, resulting in NO x Excessive emissions or high ammonia content in tail gas can pollute the air. However, nitrogen oxide sensors are greatly affected by various parameters, and aging and fatigue can cause performance degradation, resulting in reduced accuracy. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method for improving the performance of a nitrogen oxide sensor in view of the deficiencies in the prior art, thereby solving the technical problem that the performance of the existing nitrogen oxide sensor cannot be adjusted, resulting in poor detection data.

[0005] A method for improving the performance of a nitrogen oxygen sensor according to the present invention comprises:

[0006] S11: obtaining a nitrogen oxide concentration C0 of a reference gas source and a detection concentration C1 detected by a nitrogen oxide sensor, and comparing the nitrogen oxide concentration C0 with the detection concentration C1;

[0007] S12: When the nitrogen oxide concentration C0 is equal to the detection concentration C1, it is determined that the performance of the nitrogen oxide sensor is normal;

[0008] When the nitrogen oxide concentration C0 is not equal to the detection concentration C1, the nitrogen oxide sensor is corrected by calling the correction factor μ.

[0009] As a further improvement, the method for obtaining the correction factor μ is:

[0010] S21: acquiring engine operating condition data, environmental data and data in an engine database, and performing data preprocessing on the engine operating condition data, environmental data, data in the engine database, nitrogen oxide concentration C0 and detection concentration C1;

[0011] S22: Input the pre-processed engine operating data, environmental data, data in the engine database, nitrogen oxide concentration C0 and detection concentration C1 into the neural network model to obtain the correction factor μ.

[0012] Furthermore, the method for generating the neural network model is:

[0013] S31: Building a reinforcement learning model;

[0014] S32: Selecting a reinforcement learning algorithm according to the reinforcement learning model;

[0015] S33: Integrating the reinforcement learning algorithm into the DD-DC neural network model;

[0016] S34: defining the engine operating data, environmental data, data in the engine database, nitrogen oxide concentration C0 and detection concentration C1 in S21 as input layer parameters of the DD-DC neural network model, and the output of the output layer of the DD-DC neural network model is the correction factor μ;

[0017] S35: The engine operating data, environmental data, data in the engine database, nitrogen oxide concentration C0 and detection concentration C1 in S34 are divided into training samples and test samples through preprocessing, and the DD-DC neural network model is trained through the training samples to generate a training model;

[0018] The predicted Q value of the training model and the preset target Q value are used to obtain the predicted variance value, and the predicted variance value is compared with the preset variance standard value.

[0019] When the predicted variance value is equal to the variance standard value, the training is terminated and the neural network model is generated; otherwise, the DD-DC neural network model is continuously trained through the training samples.

[0020] Furthermore, in S32, the reinforcement learning algorithm is a double-delay deep deterministic policy gradient algorithm.

[0021] Furthermore, in S34, the ambient temperature disturbance and the temperature disturbance are input to perform adversarial training on the training model.

[0022] Furthermore, in S35, the neural network model is trained by the test sample to generate a test model, and when the test correction factor of the test model is equal to the preset test standard correction factor, it is verified that the training of the neural network model is completed;

[0023] When the test correction factor of the test model is not equal to the test standard correction factor, the neural network model is trained through S35.

[0024] Furthermore, in S35, the expression for obtaining the predicted Q value of the training model is:

[0025] y′=r+γ(1-done)min(Q1(s′,argmax a Q1(s′,a)+ε), Q2(s′,argmax a Q2(s′,a)+ε));

[0026] Among them, y′ is the predicted Q value, r is the immediate reward, argmax a It represents the selection of the action that maximizes the Q value, done represents whether the termination state is reached, ε is the interference noise, Q1 is the first action value, Q2 is the second action value, Q1(s′, a) is the first action value function, Q2(s′, a) is the second action value function, and γ is the discount factor.

[0027] Furthermore, the expression of the action value function is:

[0028] Q μ (s, a) = Σ s′，r P(s′,r|s,a)[r+γ∑ a′ P(a′|s′)Q μ (s′, a′)];

[0029] Among them, P(s′, r|s, a) is the probability of taking action a from state s to state s′ and obtaining reward r, P(a′|s′) is the probability of taking action a′ in state s′, r is the immediate reward, and Q μ(s′, a′) is the next action value function.

[0030] Furthermore, in S35, the expression for obtaining the predicted variance value is:

[0031] MSE=(y * -y′) 2 ;

[0032] Among them, MSE is the prediction variance value, y′ is the predicted Q value, and y* is the target Q value.

[0033] Furthermore, the measurement voltage of the nitrogen oxide sensor is adjusted by the correction factor μ so that the measurement voltage is stabilized at a preset voltage standard value.

[0034] Beneficial Effects

[0035] The advantages of the present invention are:

[0036] 1. The present invention obtains the nitrogen oxide concentration of the external nitrogen oxide gas source and the detection concentration detected by the nitrogen oxide sensor, and compares the nitrogen oxide concentration with the detection concentration; when the nitrogen oxide concentration is equal to the detection concentration, it is judged that the performance of the nitrogen oxide sensor is normal, and when the nitrogen oxide concentration is not equal to the detection concentration, the nitrogen oxide sensor is corrected by calling the correction factor, thereby eliminating the deviation of the sensor caused by changes in various parameters such as engine operating conditions, operating temperature and environment, ensuring that the nitrogen oxide sensor has high accuracy, and calibration can compensate for aging and fatigue of the sensor material over time, thereby extending the service life of the sensor.

[0037] 2. The present invention combines the reinforcement learning algorithm with the DD-DC neural network model to help the model make better decisions in a dynamic environment. At the same time, through adversarial training, the generalization ability and robustness of the model can be improved, so that the model can work normally even when the environmental conditions change greatly.

[0038] 3. The present invention inputs the pre-processed engine operating data, environmental data, data in the engine database, nitrogen oxide concentration and detection concentration into the neural network model to obtain the correction factor, taking into account the influence of various factors, ensuring the reliability of the data and making the correction result more reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic diagram of the structure of the nitrogen oxygen sensor of the present invention;

[0040] Figure 2 The present invention is a flow chart of the detection and correction of the nitrogen oxygen sensor;

[0041] Figure 3 A flowchart of obtaining a correction factor according to the present invention;

[0042] Figure 4 A flowchart of generating a neural network model of the present invention. DETAILED DESCRIPTION

[0043] The present invention will be further described below in conjunction with the embodiments, but it does not constitute any limitation to the present invention. Any limited number of modifications made by anyone within the scope of the claims of the present invention are still within the scope of the claims of the present invention.

[0044] See also Figure 1-4 , a method for improving the performance of nitrogen oxygen sensor, the method comprising, the present invention is based on dual-chamber NO x The sensor has a structure such as Figure 1 As shown. In the double chamber NO x On the basis of the sensor, the present invention adds a reference gas source and a detection correction module. When the detection correction module is called, the problem of decreased accuracy of the nitrogen oxygen sensor caused by various factors can be automatically corrected, thereby improving the performance of the sensor.

[0045] The basic principle of the nitrogen oxygen sensor is that the main pump located in the first gas chamber pumps out most of the oxygen in the exhaust gas, and the oxygen comes to the second gas chamber through the second diffusion channel. The auxiliary pump located in the second gas chamber converts the remaining small amount of oxygen and NO2 into NO gas and pumps it out. The process is to first preset the Nernst voltage V0' of the main pump and the Nernst voltage V1' of the auxiliary pump, and then compare the actual measured voltage V0 of the main pump and the actual voltage V1 of the auxiliary pump with them, and then adjust the voltage V applied to the main pump. p0 and the voltage V on the auxiliary pump p1 , thereby changing the current I on the main pump p0 and the voltage on the auxiliary pump I p1 The size of the NOx will affect the pumping in or out of oxygen. The measuring pump in the second gas chamber will pump out the oxygen generated by the NOx reaction, and its pump current can reflect the NOx X The present invention uses a neural network algorithm to obtain a correction factor μ, which is used to correct the measured voltage V2', so that the measured voltage V2 is finally stabilized at the preset voltage standard value V2', thereby adjusting the output voltage V p2 The size of the output current I p2 The size of the NOx sensor is adjusted to complete the correction of the NOx sensor. The measurement voltage of the NOx sensor is adjusted by the correction factor μ to stabilize the measurement voltage at the preset voltage standard value.

[0046] Reference gas source detection and correction scheme:

[0047] like Figure 2As shown, S11: obtain the nitrogen oxide concentration C0 of the reference gas source and the detection concentration C1 detected by the nitrogen oxide sensor, and compare the nitrogen oxide concentration C0 with the detection concentration C1. S12: when the nitrogen oxide concentration C0 is equal to the detection concentration C1, it is determined that the performance of the nitrogen oxide sensor is normal.

[0048] When the nitrogen oxide concentration C0 is not equal to the detection concentration C1, the nitrogen oxide sensor is corrected by calling the correction factor μ.

[0049] First, a gas pump is used to pump nitrogen oxides with a concentration of C0 in the reference gas source into the nitrogen oxygen sensor. The reference gas passes through the first gas chamber and the second gas chamber in turn. After detection by the nitrogen oxygen sensor, NO x The concentration C1 is then compared with the detection concentration C1. If the nitrogen oxide concentration C0 is equal to the detection concentration C1, it means that the nitrogen oxide sensor has a good performance and does not need to be corrected, and continues to work normally. If the nitrogen oxide concentration C0 is not equal to the detection concentration C1, the correction module is called to correct the preset Nernst voltage value V2' with the correction factor μ, and then the detection voltage V2 is corrected, thereby changing the output voltage V p2 and output current I p2 The size of the NOx sensor is thus solved. Finally, the display module is called to display the calibration results on the user interface, and the calibration results and various parameters are recorded in the database, which is convenient for users to view the status of the sensor and perform maintenance.

[0050] The method to obtain the correction factor μ is:

[0051] S21: Acquire engine operating condition data, environmental data and data in an engine database, and perform data preprocessing on the engine operating condition data, environmental data, data in an engine database, nitrogen oxide concentration C0 and detection concentration C1.

[0052] S22: Input the pre-processed engine operating data, environmental data, data in the engine database, nitrogen oxide concentration C0 and detection concentration C1 into the neural network model to obtain the correction factor μ. The engine operating data includes engine operating conditions, exhaust temperature and airspeed, and the environmental data includes ambient temperature and ambient humidity.

[0053] like Figure 3As shown, the correction module is called to convert the engine operating conditions, exhaust temperature, airspeed and other data, ambient temperature, ambient humidity, nitrogen oxide concentration C0 and detection concentration C1 into digital signals, and the historical data in the database module is called at the same time, and all the above digital signals are transmitted to the ECU for data preprocessing. The preprocessed data is input into the neural network self-adjusting module, and the correction factor μ can be obtained through the neural network model trained in the neural network self-adjusting module.

[0054] Then the correction factor μ is used to correct the nitrogen oxide sensor. After correction, the reference gas module is used to detect the performance of the nitrogen oxide sensor to determine whether the nitrogen oxide concentration C0 and the detection concentration C1 are equal. If they are equal, the correction is completed. If they are not equal, the neural network self-adjustment module is called again to correct the nitrogen oxide sensor.

[0055] like Figure 4 As shown, the present invention integrates the DD-DC (Direct Data-Driven Control) neural network model and the reinforcement learning algorithm into a neural network model, trains the neural network model, and then completes the correction of the nitrogen oxide sensor.

[0056] The method to generate a neural network model is:

[0057] S31: Build a reinforcement learning model to transform the NOx sensor calibration problem into a Markov decision process (MDP), which generally includes state space (s), action space (a), transfer dynamics (m), reward function (R) and discount factor (γ). Then define the reinforcement learning reward function R: the NOx sensor will be rewarded with a large positive reward if it obtains an accurate correction factor μ, and a large negative reward if it obtains an inaccurate μ. Through a large amount of training, the NOx sensor correction strategy is continuously updated to obtain the maximum cumulative reward G. t , thus proposing the optimal correction strategy μ * .

[0058]

[0059] μ * = argmax E μ [G t ];

[0060] In the above formula, t is the time t, r is the reward value of the reward function (which depends on the strategy μ), γ is the discount factor, and E is the expectation. The expectation of the cumulative reward is generally expressed as the action value function Q μ (s, a) represents.

[0061] Q μ (s, a) = Σ s′，r P(s′,r|s,a)[r+γΣa′ P(a′|s′)Q μ (s′, a′)];

[0062] Where P(s′, r|s, a) is the probability of taking action a from state s to state s′ and obtaining reward r, and P(a′|s′) is the probability of taking action a′ in state s′.

[0063] S32: Select a reinforcement learning algorithm according to the reinforcement learning model.

[0064] In S32, the reinforcement learning algorithm is a double-delayed deep deterministic policy gradient algorithm. Double-delayed deep deterministic policy gradient (TD3) is selected as the reinforcement learning algorithm. TD3 uses the Actor-Critic structure, so the model includes a policy network and a value network. The policy network outputs actions through the state information of the nitrogen oxide sensor, and the value network outputs the Q value to evaluate the quality of the action. At the same time, the algorithm adds a target network to improve training stability. TD3 learns two Q value functions and uses the minimum of these two estimated values ​​during policy updates to reduce the problem of overestimation of the value function. It also enhances the stability and convergence of the algorithm by delaying policy updates, and smoothes the target policy by adding interference noise during training.

[0065] S33: Integrate reinforcement learning algorithm into DD-DC neural network model. The DD-DC neural network model is used as the brain of the intelligent agent. It regards neurons as feedback controllers of the environment, affects the performance of the nitrogen oxide sensor through the correction factors it outputs, and receives feedback from the environment to optimize the control effect. The input parameters are analyzed and processed by reinforcement learning to obtain the best correction strategy and the correction factor μ, so as to correct the nitrogen oxide sensor. The above is the working process of the DD-DC model.

[0066] S34: The engine operating data, environmental data, data in the engine database, nitrogen oxide concentration C0 and detection concentration C1 in S21 are defined as parameters of the input layer of the DD-DC neural network model. The output of the output layer of the DD-DC neural network model is the correction factor μ.

[0067] In S34, the ambient temperature disturbance and the temperature disturbance are input to perform adversarial training on the training model. The hidden layer activation function uses the logsig function, and the output of the output layer is the predicted correction factor μ.

[0068]

[0069] S35: The engine operating data, environmental data, data in the engine database, nitrogen oxide concentration C0 and detection concentration C1 in S34 are divided into training samples and test samples through preprocessing, and the DD-DC neural network model is trained through the training samples to generate a training model.

[0070] The predicted Q value of the training model and the preset target Q value are used to obtain a predicted variance value, and the predicted variance value is compared with the preset variance standard value.

[0071] When the predicted variance value is equal to the variance standard value, the training is terminated and the neural network model is generated; otherwise, the DD-DC neural network model is continuously trained through the training samples.

[0072] The expression for obtaining the predicted Q value of the training model is,

[0073] y′=r+γ(1-done)min(Q1(s′,argmax a Q1(s′,a)+ε), Q2(s′,argmax a Q2(s′,a)+ε));

[0074] Among them, y′ is the predicted Q value, r is the immediate reward, argmax a It represents the selection of the action that maximizes the Q value, done represents whether the termination state is reached, ε is the interference noise, Q1 is the first action value, Q2 is the second action value, Q1(s′, a) is the first action value function, Q2(s′, a) is the second action value function, and γ is the discount factor.

[0075] The derivation process of the above formula is as follows:

[0076] y=r+γ(1-done)min(Q1(s′,argmax a Q1(s′,a)), Q2(s′,argmax a Q2(s′, a));

[0077] y′=r+γ(1-done)min(Q1(s′,argmax a Q1(s′,a)+ε), Q2(s′,argmax a Q2(s′,a)+ε));

[0078] Among them, y′ is the predicted Q value, r is the immediate reward, argmax aIt represents the selection of the action that maximizes the Q value, done represents whether the termination state is reached, ε is the interference noise, Q1 is the first action value, Q2 is the second action value, Q1(s′, a) is the first action value function, Q2(s′, a) is the second action value function, and γ is the discount factor.

[0079] ε~clip(N(0,σ),-c,c);

[0080] Where r is the immediate reward, argmaxa represents the action that maximizes the Q value. done indicates whether the terminal state is reached, ε is the interference noise, and c is the cutoff boundary value of the policy smoothing noise.

[0081] In S35, the neural network model is trained by the test sample to generate a test model, and when the test correction factor of the test model is equal to the target Q value, it is verified that the training of the neural network model is completed;

[0082] When the test correction factor of the test model is not equal to the target Q value, the neural network model is trained through S35.

[0083] The large amount of input data is preprocessed and divided into training samples and test samples. Then the neural network model is trained with the training samples. At the same time, some interference factors such as ambient temperature disturbance and humidity disturbance are input to conduct adversarial training on the model. The robustness of the DD-DC model is enhanced through adversarial training. The model is trained according to the input data to obtain the best correction strategy under different conditions.

[0084] Then, the trained neural network model is tested using test samples to verify its correction effect; finally, based on the test results, the model parameters and algorithms are further optimized.

[0085] The following is the training process of the neural network model:

[0086] First, the DD-DC neural network model receives input data. The DD-DC model has great advantages in collecting environmental data. It can effectively and quickly provide feedback and control the environment. Then the logsig function is used to standardize the input parameters, that is, to map various parameters to the (0, 1) interval. Then the DD-DC neuron selects the appropriate target Q value, i.e., y*, based on the above data. Then the target Q value y* is input into the reinforcement learning module, the sensor correction module is trained using the TD3 algorithm, and the action value function Q is used to train the sensor correction module. μ (s, a) calculate the Q value of the corresponding action, get the Q1 value of the first action and the Q2 value of the second action, use

[0087] y′=r+γ(1-done)min(Q1(s′,argmaxa Q1(s′,a)+ε), Q2(s′,argmax a Q2(s′, a)+ε)); calculate, ε is the interference factor added by adversarial training, which can increase the stability of the model, and calculate a smaller Q value, that is, the predicted Q value y′. By selecting the minimum of the two Q values, the influence of overestimation is reduced. Then, the predicted Q value y′ and the target Q value y* are subjected to variance analysis to obtain the predicted variance value MSE. The expression of the predicted variance value MSE is MsE=(y + -y′) 2 , when the predicted variance MSE is greater than the variance standard value 10 -6 Continue training when the predicted variance MSE is less than the variance standard value 10 -6 When , the training ends. During the training process, each action a will receive a reward r, using Perform cumulative calculation of rewards to obtain the optimal calibration strategy μ * = argmax E μ [G t ], the DD-DC neural model uses the correction factor μ to calibrate the sensor, and then performs another reference gas source module detection. If the nitrogen oxide concentration C0 and the detection concentration C1 are not equal, it means that there is room for improvement in the calibration strategy in this case. The target Q value y* is re-predicted, and then the reinforcement learning model and method are used for training to continuously optimize the calibration strategy. If the nitrogen oxide concentration C0 and the detection concentration C1 are equal, the training is terminated, and the DD-DC neural network model calibrates the nitrogen oxide sensor again according to the correction factor μ.

[0088] The above is only a preferred embodiment of the present invention. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the structure of the present invention, which will not affect the effect of the implementation of the present invention and the practicality of the patent.

Claims

1. A method for improving the performance of a nitrogen oxide sensor, characterized in that: The method includes, S11: obtaining a nitrogen oxide concentration C0 of a reference gas source and a detection concentration C1 detected by a nitrogen oxide sensor, and comparing the nitrogen oxide concentration C0 with the detection concentration C1; S12: When the nitrogen oxide concentration C0 is equal to the detection concentration C1, it is determined that the performance of the nitrogen oxide sensor is normal; When the nitrogen oxide concentration C0 is not equal to the detection concentration C1, the nitrogen oxide sensor is corrected by calling the correction factor μ.

2. A method for improving the performance of a nitrogen oxygen sensor according to claim 1, characterized in that: The method for obtaining the correction factor μ is: S21: acquiring engine operating condition data, environmental data and data in an engine database, and performing data preprocessing on the engine operating condition data, environmental data, data in the engine database, nitrogen oxide concentration C0 and detection concentration C1; S22: Input the pre-processed engine operating data, environmental data, data in the engine database, nitrogen oxide concentration C0 and detection concentration C1 into the neural network model to obtain the correction factor μ.

3. A method for improving the performance of a nitrogen oxygen sensor according to claim 2, characterized in that: The method for generating the neural network model is: S31: Building a reinforcement learning model; S32: Selecting a reinforcement learning algorithm according to the reinforcement learning model; S33: Integrating the reinforcement learning algorithm into the DD-DC neural network model; S34: defining the engine operating data, environmental data, data in the engine database, nitrogen oxide concentration C0 and detection concentration C1 in S21 as input layer parameters of the DD-DC neural network model, and the output of the output layer of the DD-DC neural network model is the correction factor μ; S35: The engine operating data, environmental data, data in the engine database, nitrogen oxide concentration C0 and detection concentration C1 in S34 are divided into training samples and test samples through preprocessing, and the DD-DC neural network model is trained through the training samples to generate a training model; The predicted Q value of the training model and the preset target Q value are used to obtain the predicted variance value, and the predicted variance value is compared with the preset variance standard value. When the predicted variance value is equal to the variance standard value, the training is terminated and the neural network model is generated; otherwise, the DD-DC neural network model is continuously trained through the training samples.

4. A method for improving the performance of a nitrogen oxygen sensor according to claim 3, characterized in that: In S32, the reinforcement learning algorithm is a double-delay deep deterministic policy gradient algorithm.

5. The method for improving the performance of a nitrogen oxygen sensor according to claim 3, characterized in that: In S34, the ambient temperature disturbance and the temperature disturbance are input to perform adversarial training on the training model.

6. The method for improving the performance of a nitrogen oxygen sensor according to claim 3, characterized in that: In S35, the neural network model is trained by the test sample to generate a test model, and when the test correction factor of the test model is equal to the preset test standard correction factor, it is verified that the training of the neural network model is completed; When the test correction factor of the test model is not equal to the test standard correction factor, the neural network model is trained through S35.

7. The method for improving the performance of a nitrogen oxygen sensor according to claim 3, characterized in that: In S35, the expression for obtaining the predicted Q value of the training model is: y′=r+γ(1-done)min(Q1(s′,argmax a Q1(s′,a)+ε),Q2(s′,argmax a Q2(s′,a)+ε)); Among them, y′ is the predicted Q value, r is the immediate reward, argmax a It represents the selection of the action that maximizes the Q value, done represents whether the termination state is reached, ε is the interference noise, Q1 is the first action value, Q2 is the second action value, Q1(s′, a) is the first action value function, Q2(s′, a) is the second action value function, and γ is the discount factor.

8. The method for improving the performance of a nitrogen oxygen sensor according to claim 7, characterized in that: The expression of the action value function is: Q μ (s,a)=∑ s′，r P(s′,r|s,a)[r+γ∑ a′ P(a′|s′))Q μ ((s′,a′)]; Among them, P(s′, r|s, a) is the probability of taking action a from state s to state s′ and obtaining reward r, P(a′|s′) is the probability of taking action a′ in state s′, r is the immediate reward, and Q μ (s′, a′) is the next action value function.

9. The method for improving the performance of a nitrogen oxygen sensor according to claim 7, characterized in that: In S35, the expression for obtaining the predicted variance value is: MSE=(y * -y′) 2 ; Among them, MSE is the prediction variance value, y′ is the predicted Q value, and y* is the target Q value.

10. The method for improving the performance of a nitrogen oxygen sensor according to claim 1, characterized in that: The measurement voltage of the nitrogen oxide sensor is adjusted by the correction factor μ so that the measurement voltage is stabilized at a preset voltage standard value.

Images