Correlation entropy-based event-triggered recursive fuzzy neural network multivariable control method

Abstract

The invention provides an event-triggered recursive fuzzy neural network multivariable control method based on correlation entropy. The method comprises the following steps: acquiring sewage treatment process data; constructing an initial fuzzy neural network through the sewage treatment process data; constructing an event trigger structure adjustment mechanism, and performing structure adjustment on the initial fuzzy neural network through the event trigger structure adjustment mechanism to obtain an optimized fuzzy neural network; performing parameter updating on the optimized fuzzy neural network through the correlation entropy to obtain an updated fuzzy neural network; verifying the updated fuzzy neural network through a Lyapunov function to obtain a recursive fuzzy neural network; and performing variable control on the sewage treatment process through the recursive fuzzy neural network. According to the method, the sewage treatment process is controlled through the recursive fuzzy neural network, the problem that dissolved oxygen and nitrate nitrogen are difficult to control is solved, and the control accuracy and efficiency are improved.

Description

Technical Field

[0001] The invention relates to the technical field of sewage treatment, and in particular to an event-triggered recursive fuzzy neural network multivariable control method based on correlation entropy. Background Art

[0002] With the development of industry and the consumption of domestic water, water pollution has received widespread attention from all countries. Sewage treatment technology can effectively remove pollutants in wastewater and plays an indispensable role in environmental protection. However, due to the complex physical and biochemical reactions and external interference in the sewage treatment process, there are problems such as uncertainty, nonlinearity, and dynamic time-varying in the sewage treatment process. Therefore, ensuring the normal operation of the sewage treatment process faces major challenges.

[0003] At present, in sewage treatment plants, the activated sludge process is the most commonly used sewage treatment method. The activated sludge process consists of three treatment zones: anaerobic zone, anoxic zone, and aerobic zone. The main process of denitrification reaction is carried out in the anoxic zone, and the denitrification process converts nitrate nitrogen (NO3-N) into harmless nitrogen gas. At the same time, the main nitrification reaction occurs in the aerobic zone, and the nitrification reaction oxidizes ammonia nitrogen (NH4-N) to NO3-N, which enters the anaerobic zone again through internal reflux NO3-N. In the activated sludge process, the dissolved oxygen (DO) concentration affects the nitrification and denitrification processes, while the NO3-N concentration is an important factor affecting the denitrification reaction, and there is a coupling relationship between the two. However, it is difficult for traditional controllers to achieve stable control of DO and NO3-N concentrations. Therefore, it is very necessary to design an event-triggered recursive fuzzy neural network multivariable control method based on correlation entropy. Summary of the invention

[0004] The purpose of the present invention is to provide an event-triggered recursive fuzzy neural network multivariable control method based on correlation entropy, so as to control the sewage treatment process through the constructed recursive fuzzy neural network, solve the problem of difficulty in controlling dissolved oxygen and nitrate nitrogen, and improve the accuracy and efficiency of control.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] An event-triggered recursive fuzzy neural network multivariable control method based on correlation entropy comprises the following steps:

[0007] Obtain sewage treatment process data; sewage treatment process data includes dissolved oxygen concentration and nitrate nitrogen concentration;

[0008] An initial fuzzy neural network is constructed through the sewage treatment process data; the initial fuzzy neural network includes: an input layer, a radial basis function layer, a rule layer, a normalization layer and an output layer; the input layer, the radial basis function layer, the rule layer, the normalization layer and the output layer are connected in sequence;

[0009] Construct an event-triggered structural adjustment mechanism, and adjust the structure of the initial fuzzy neural network through the event-triggered structural adjustment mechanism to obtain an optimized fuzzy neural network;

[0010] The parameters of the optimized fuzzy neural network are updated through the correlation entropy to obtain an updated fuzzy neural network;

[0011] The updated fuzzy neural network is verified by Lyapunov function, and the recursive fuzzy neural network is obtained;

[0012] Variable control of sewage treatment process by recursive fuzzy neural network.

[0013] Optionally, the input variables of the input layer include: dissolved oxygen concentration error, dissolved oxygen concentration error change rate, nitrate nitrogen concentration error and nitrate nitrogen concentration error change rate;

[0014] The calculation formula of the membership of the radial basis function layer is: Among them, a ij (t) is the membership of the i-th input variable to the j-th radial basis function, c ij (t) is the membership function center of the i-th input variable to the j-th radial basis function, σ ij (t) is the membership function width of the i-th input variable to the j-th radial basis function, o i (t) is the output vector of the i-th neuron in the input layer;

[0015] The expression of the rule layer is: Among them, φ j (t) is the output of the jth rule layer node at the current moment, f j (t) is the sigmoid function, n is the number of neurons in the input layer;

[0016] The expression of the normalization layer is: in, is the output of the jth normalization layer neuron, and P is the total number of radial basis function layer neurons;

[0017] The expression of the output layer is: Among them, y q (t) is the output of the qth output layer node at time t, is the output weight, and Q is the number of output nodes.

[0018] Optionally, an event-triggered structural adjustment mechanism is constructed, and the initial fuzzy neural network is structurally adjusted by the event-triggered structural adjustment mechanism to obtain an optimized fuzzy neural network, including:

[0019] Construct the maximum relevant entropy criterion;

[0020] The loss function is obtained by the maximum relevant entropy criterion;

[0021] The event definition is obtained through the loss function;

[0022] The structure of the initial fuzzy neural network is adjusted according to the event definition to obtain the optimized fuzzy neural network.

[0023] Optionally, the calculation formula of the maximum correlation entropy criterion is: Among them, H(t) is the maximum correlation entropy, y k is the expected output, is the actual output, σ g is the kernel width, e k is the sample error, and K is the total number of samples.

[0024] Optionally, the loss function is calculated as: Among them, J(t) is the loss function, e(t) is the composite error, σ g is the kernel width.

[0025] Optionally, the event definition expression is: Among them, Eve1 is the first event, Eve2 is the second event, Eve3 is the third event, J(t) is the loss function, ξ1 and ξ2 are both event triggering thresholds, δ is the time period, γ δ (t) is the change of loss function within δ, t is the current time, t E is the last triggering moment, T E is the maximum safety limit threshold.

[0026] Optionally, the initial fuzzy neural network is structurally adjusted according to the event definition to obtain an optimized fuzzy neural network, including:

[0027] When the first event is triggered, the number of neurons in the regular layer is increased;

[0028] When the second event is triggered, the normalization layer node is removed;

[0029] When the third event is triggered, the structure of the initial fuzzy neural network remains unchanged.

[0030] Optionally, the parameters of the optimized fuzzy neural network are updated by the correlation entropy to obtain an updated fuzzy neural network, specifically: the parameters of the optimized fuzzy neural network are updated by the gradient descent method; the calculation formula of the gradient descent method is: Among them, Θ(t+1) is the parameter matrix of the next moment, Θ(t) is the parameter matrix of the current moment, β(t) is the positive learning rate, J k (t) is the cost function.

[0031] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: the event-triggered recursive fuzzy neural network multivariable control method based on correlation entropy provided by the present invention comprises: obtaining sewage treatment process data; constructing an initial fuzzy neural network through the sewage treatment process data; constructing an event-triggered structural adjustment mechanism, and adjusting the structure of the initial fuzzy neural network through the event-triggered structural adjustment mechanism to obtain an optimized fuzzy neural network; updating the parameters of the optimized fuzzy neural network through correlation entropy to obtain an updated fuzzy neural network; verifying the updated fuzzy neural network through the Lyapunov function to obtain a recursive fuzzy neural network; and controlling the variables of the sewage treatment process through the recursive fuzzy neural network. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0033] Figure 1 is a flow chart of the multivariable control method of the present invention;

[0034] Figure 2 The flowchart of adjusting the initial fuzzy neural network structure of the present invention;

[0035] Figure 3 It is a neural network verification flow chart of the present invention;

[0036] Figure 4 A comparison diagram of fixed dissolved oxygen concentration set value trajectories according to an embodiment of the present invention;

[0037] Figure 5 A comparison diagram of fixed dissolved oxygen concentration tracking control error trajectories in an embodiment of the present invention;

[0038] Figure 6 A comparison diagram of fixed nitrate nitrogen concentration set value trajectories according to an embodiment of the present invention;

[0039] Figure 7 A comparison diagram of the fixed nitrate nitrogen concentration tracking control error trajectory of an embodiment of the present invention;

[0040] Figure 8 A graph showing the number of fixed recursive fuzzy neural rules according to an embodiment of the present invention;

[0041] Fig. 9 A comparison diagram of dynamic dissolved oxygen concentration set value trajectories of an embodiment of the present invention;

[0042] Fig.10A comparison diagram of dynamic dissolved oxygen concentration tracking control error trajectories according to an embodiment of the present invention;

[0043] Fig.11 A comparison diagram of dynamic nitrate nitrogen concentration set value trajectories according to an embodiment of the present invention;

[0044] Fig.12 A comparison diagram of dynamic nitrate nitrogen concentration tracking control error trajectories according to an embodiment of the present invention;

[0045] Fig.13 This is a graph showing the number of dynamic recursive fuzzy neural rules according to an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and understandable, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] like Figure 1 As shown, the present invention provides an event-triggered recursive fuzzy neural network multivariable control method based on correlation entropy, comprising the following steps:

[0049] Step 100: Obtaining sewage treatment process data; the sewage treatment process data includes dissolved oxygen concentration and nitrate nitrogen concentration;

[0050] Step 200: constructing an initial fuzzy neural network through sewage treatment process data; the initial fuzzy neural network includes: an input layer, a radial basis function (RBF) layer, a rule layer, a normalization layer and an output layer; the input layer, the radial basis function (RBF) layer, the rule layer, the normalization layer and the output layer are connected in sequence;

[0051] Specifically, in this embodiment, the number of neurons in the input layer is n=4, the number of neurons in the RBF layer is P=1, and the number of neurons in the output layer is Q=2. The nodes in the input layer are connected by input variables x(t)=[x1,x2,x3,x4] T The activated input variables include: dissolved oxygen concentration error x1, dissolved oxygen concentration error change rate x2, nitrate nitrogen concentration error x3 and nitrate nitrogen concentration error change rate x4; the expression of the input layer is:

[0052] o i (t) = x i(t),i=1,2,3,...,n;

[0053] x1(t)=e O (t) = S O,5 (t)-S O,set (t);

[0054] x2=Δe O (t);

[0055] x3(t)=e NO (t) = S NO,2 (t)-S NO,set (t);

[0056] x4=Δe NO (t);

[0057] Among them, x i (t) is the input value of the ith neuron in the input layer, o i (t) is the output vector of the i-th neuron in the input layer, S O,5 (t) is the actual concentration of dissolved oxygen, S O,set (t) is the set value of dissolved oxygen concentration, S NO,2 (t) is the actual concentration of nitrate nitrogen, S NO,set (t) is the set value of nitrate nitrogen concentration, e O (t) and e NO (t) represent the tracking control errors of dissolved oxygen concentration and nitrate nitrogen concentration, Δe O (t) and Δe NO (t) represents the tracking control error change of dissolved oxygen concentration and nitrate nitrogen concentration respectively.

[0058] The calculation formula of the membership of the RBF layer is:

[0059]

[0060] Among them, a ij (t) is the membership of the i-th input variable to the j-th radial basis function, c ij (t) is the membership function center of the i-th input variable to the j-th radial basis function, σ ij (t) is the membership function width of the i-th input variable to the j-th radial basis function, o i (t) is the output vector of the i-th neuron in the input layer.

[0061] The expression of the rule layer is:

[0062]

[0063] Among them, φ j(t) is the output of the jth rule layer node at the current moment, φ j (t-1) is the output of the jth rule layer node at the previous moment, f j (t) is the sigmoid function, λ j (t) is the connection weight of internal feedback.

[0064] The expression of the normalization layer is:

[0065]

[0066] in, is the output of the jth normalized layer neuron;

[0067] The expression of the output layer is:

[0068]

[0069] Among them, y q (t) is the output of the qth output layer node at time t, is the output weight.

[0070] Step 300: construct an event-triggered structural adjustment mechanism, and adjust the structure of the initial fuzzy neural network through the event-triggered structural adjustment mechanism to obtain an optimized fuzzy neural network; the specific steps are as follows: Figure 2 As shown, including:

[0071] Step 301: constructing a maximum correlation entropy criterion;

[0072] Specifically, the calculation formula of the maximum correlation entropy criterion is:

[0073]

[0074] Among them, H(t) is the maximum correlation entropy, y k is the expected output, is the actual output, σ g is the kernel width, e k is the sample error, and K is the total number of samples.

[0075] Step 302: Obtaining a loss function through the maximum relevant entropy criterion;

[0076] Specifically, the loss function is obtained by rewriting the maximum relevant entropy criterion to minimize the difference between the expected output and the actual output. The calculation formula of the loss function is:

[0077]

[0078] Among them, J(t) is the loss function, e(t) is the composite error, σ gis the kernel width. The composite error is expressed as:

[0079] e(t)=α(t)e O (t)+(1-α(t))e NO (t);

[0080] α(t)=|e O (t)|(|e O (t)|+|e NO (t) |);

[0081] Among them, α(t) is the error coefficient at time t. And through γ δ (t) represents the change of the loss function in the time period δ, γ δ The calculation formula of (t) is: δ (t) = J(t) - J(t-δ); where J(t) is the loss function at the current moment, and J(t-δ) is the loss function at the previous δ moments.

[0082] Step 303: Obtaining event definition through loss function;

[0083] Specifically, the expression of the event definition is:

[0084]

[0085] Among them, Eve1 is the first event, Eve2 is the second event, Eve3 is the third event, J(t) is the loss function, ξ1 and ξ2 are both event triggering thresholds, δ is the time period, t is the current time, t E is the last triggering moment, T E is the maximum safety limit threshold.

[0086] It should be noted that Eve1 indicates that the correlation entropy is greater than the maximum allowed threshold ξ1, γ δ (t)>0 means that the correlation entropy has an increasing trend; Eve2 means that the correlation entropy is less than the lower threshold ξ2, γ δ (t) < 0 indicates that the correlation entropy is decreasing; Eve3 indicates that the time interval from the last event to the triggering exceeds the maximum safety limit threshold T E .

[0087] Step 304: Adjust the structure of the initial fuzzy neural network according to the event definition to obtain an optimized fuzzy neural network.

[0088] Specifically, when a specific event defined by the event is triggered, the structure and parameters of the initial fuzzy neural network are also updated; if no specific event is triggered, the state of the initial fuzzy neural network should be maintained until the next event occurs. In the process of adjusting the structure of the initial fuzzy neural network, the maximum and minimum outputs of the normalization layer nodes must be determined first, which are expressed as:

[0089]

[0090] in, is the maximum output of the normalization layer node, is the minimum output of the normalization layer node.

[0091] More specifically, when Eve1 appears, it means that the similarity between the output and the set value is insufficient, that is, the error is large, and rules need to be added, that is, high-precision control is achieved by increasing the number of neurons in the rule layer. And the parameters and weights of the newly added neuron nodes are initialized. The expression after initialization is as follows:

[0092]

[0093] Among them, c new (t) and σ new (t) are the center vector and width vector of the new membership function node, λ new (t) is the new internal connection weight, is the new output weight, c g (t) and σ g (t) are the center vector and width vector of the g-th membership function node, λ gr (t) represents the connection weight of the g-th internal feedback.

[0094] When Eve2 appears, the dth redundant normalization layer node is removed, and the expressions of the corresponding parameters and output weights of this node are:

[0095]

[0096] Among them, c d (t) and σ d (t) are the center vector and width vector of the d-th membership function node, λ d (t) and are the internal connection weight and output weight of the dth normalized layer node respectively. And select the sth node with the smallest Euclidean distance to the dth node, and set its parameters as follows:

[0097]

[0098] Among them, c' s (t),σ's (t), λ′ s (t) and are the membership function center, membership function width, internal feedback connection weight and output weight of the sth node after parameter adjustment, c s (t),σ s (t), λ s (t) and are the membership function center, membership function width, internal feedback connection weight and output weight of the sth node before parameter adjustment, and They are the normalized outputs of the d-th node and the s-th node respectively.

[0099] When Eve3 appears, the structure of the initial fuzzy neural network remains unchanged.

[0100] It should be noted that the event-triggered structural adjustment mechanism improves the generalization ability of the initial fuzzy neural network and suppresses the interference of noise and outliers.

[0101] Step 400: updating the parameters of the optimized fuzzy neural network by using the correlation entropy to obtain an updated fuzzy neural network;

[0102] Specifically, the gradient descent method is used to optimize the parameters of the fuzzy neural network for online updating; the relevant entropy criterion in the online updating process ensures that the network output is close to the expected output by maximizing the error probability density function.

[0103] More specifically, the cost function in the online update process is calculated as: Among them, J k (t) is the cost function. Based on the cost function, the calculation formula using the gradient descent method is: Among them, Θ(t+1) is the parameter matrix of the next moment, Θ(t) is the parameter matrix of the current moment, and β(t) is the positive learning rate. The update process is completed when the positive learning rate meets the range condition, and the range condition is:

[0104] Step 500: Verify the updated fuzzy neural network through the Lyapunov function to obtain a recursive fuzzy neural network; the specific steps are as follows: Figure 3 As shown, including:

[0105] Step 501: verifying the stability of the fixed structure of the updated fuzzy neural network;

[0106] Specifically, the normalization layer in the fixed structure of the updated fuzzy neural network has P nodes. The expression of the Lyapunov function is set as: Then calculate the deviation of V1(t), the calculation formula is:

[0107]

[0108] ζ(t)=e 2 (t+1)-e 2 (t) = 2e(t)Δe(t) + Δe 2 (t);

[0109] Then perform e(t+1) Tate expansion on , and the expansion process formula is:

[0110]

[0111] in And combined with the above calculation process, we can get: ζ(t) = e 2 (t)[(1-η(t)J(t)e(t)Ψ(t)) 2 -1].

[0112] Assume ψ max (t) = max{ψ c (t),ψ σ (t),ψ c (t),ψ σ (t)}, we can get Ψ(t)≤4ψ max (t), and then we get In addition, 0<J(t)≤1, and based on the above conditions, we can get ζ(t)<0 and ΔV1(t)<0; and because V1(t)>0, it can be known that when the kernel width satisfies the condition When , the fixed structure of the updated fuzzy neural network is stable.

[0113] Step 502: verifying the stability of the structure growth phase of the updated fuzzy neural network;

[0114] Specifically, the number of normalized layer nodes in the structure growth phase of the updated fuzzy neural network increases from P to (P+1). The expression of the Lyapunov function is set as:

[0115]

[0116] in, is the output error of the qth output node at time t after a new normalization layer node is added, and its calculation formula is:

[0117]

[0118] in, is the output value of the qth output node of the network with (P+1) normalization layer nodes, is the output value of the qth output node of a network with P normalized layer nodes. It can be seen that the derivative of V2(t) is:

[0119]

[0120] in It can be obtained by calculation, and the calculation process is as follows:

[0121]

[0122] From the above calculation formula, we can get Through the stability verification of the fixed structure, it is known that therefore At the same time, because V2(t) is greater than 0, the structure growth phase of updating the fuzzy neural network is stable.

[0123] Step 503: verifying the stability of the structure reduction phase of the updated fuzzy neural network;

[0124] Specifically, the number of normalized layer nodes in the structure reduction phase of the updated fuzzy neural network is reduced from P to (P-1). The expression of the Lyapunov function is set as:

[0125]

[0126] in, is the output error of the qth output node at time t after a redundant normalization layer node is deleted, and its calculation formula is:

[0127]

[0128] in, is the output value of the qth output node of a network with P normalization layer nodes, is the output value of the qth output node of the network with (P-1) normalized layer nodes. From the above calculation formula, the calculation formula of the derivative of V3(t) can be obtained as:

[0129]

[0130] Integrating the above calculation formulas, we can get:

[0131]

[0132] Through the stability verification of the fixed structure, it is known that therefore At the same time, because V3(t) is greater than 0, the structure reduction phase of updating the fuzzy neural network is stable.

[0133] Step 504: Evaluate the overall control performance of the updated fuzzy neural network and determine the recursive fuzzy neural network according to the evaluation result.

[0134] Specifically, the integrated square error (ISE), integrated absolute error (IAE) and maximum absolute error Dev are used. max As evaluation indicators, the respective calculation formulas are as follows:

[0135]

[0136] Dev max = max{e(t)};

[0137] Among them, K is the total number of samples.

[0138] Step 600: Perform variable control on the sewage treatment process through a recursive fuzzy neural network.

[0139] In Example 1, the data of process control using recursive fuzzy neural network (CE-RFNN) were compared with the data using PID, fuzzy neural network (FNN) and self-organizing recursive fuzzy neural network (SORFNN), and the comparison results of dissolved oxygen concentration and nitrate nitrogen concentration as fixed set values ​​are shown in Table 1:

[0140] Table 1 Comparison of data for dissolved oxygen concentration and nitrate nitrogen concentration with fixed set values

[0141]

[0142] The comparison results of dissolved oxygen concentration and nitrate nitrogen concentration as dynamic setting values ​​are shown in Table 2:

[0143] Table 2 Comparison of data of dissolved oxygen concentration and nitrate nitrogen concentration as dynamic setting values

[0144]

[0145]

[0146] In Example 2, the dissolved oxygen concentration in the sewage treatment process was set to a fixed value of 2 mg / L in the context of heavy rain. The data of process control using recursive fuzzy neural network were compared with those using proportional differential integral (PID), fuzzy neural network (FNN) and self-organizing fuzzy neural network (SOFNN). The trajectory of the dissolved oxygen concentration set value was compared. Figure 4 As shown in the figure, the tracking control error trajectory of dissolved oxygen concentration is compared with Figure 5 As shown in Figure 2. The trajectory of the set value of nitrate nitrogen concentration is compared with Figure 6 As shown in Figure 2, the tracking control error trajectory of nitrate nitrogen concentration is compared with Figure 7As shown in Figure 2, the change in the number of recursive fuzzy neural rules is as follows: Figure 8 shown.

[0147] In Example 3, with heavy rain as the background, the dissolved oxygen concentration in the sewage treatment process is dynamically set as follows: the initial setting value is 2 mg / L, which drops to 1.8 mg / L on the 3rd day, rises to 2 mg / L on the 6th day, increases to 2.2 mg / L on the 8th day, and drops to 2 mg / L on the 12th day. The nitrate nitrogen concentration in the sewage treatment process is dynamically set as follows: the initial setting value is 1 mg / L, which drops to 0.8 mg / L on the 3rd day, rises to 1 mg / L on the 6th day, increases to 1.2 mg / L on the 8th day, and drops to 1 mg / L on the 12th day. The data obtained by using a recursive fuzzy neural network for process control are compared with the data obtained by using proportional differential integral (PID), fuzzy neural network (FNN), and self-organizing fuzzy neural network (SOFNN). The set value tracking trajectory of the dissolved oxygen concentration is compared. Fig. 9 As shown in Figure 2, the tracking error trajectory of dissolved oxygen concentration is compared with Fig.10 As shown in the figure, the set value tracking trajectory of nitrate nitrogen concentration is compared with Fig.11 As shown in Figure 2, the tracking error trajectory of nitrate nitrogen concentration is shown in Figure 2. Fig.12 The change in the number of recursive fuzzy neural rules is shown in Fig.13 As shown, when the set value changes dynamically, the number of recursive fuzzy neural network rules can be changed accordingly to obtain the best dynamic effect.

[0148] The beneficial effects of the present invention are as follows:

[0149] 1) The ability to capture nonlinear information in the sewage treatment process is improved by constructing a recursive fuzzy neural network structure;

[0150] 2) The event-triggered structural adjustment mechanism improves the generalization ability of the recursive fuzzy neural network and suppresses the interference of noise and outliers;

[0151] 3) Online parameter updating through relevant entropy improves the adaptability, learning performance, anti-interference ability and stability of the network structure.

[0152] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0153] The present invention uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the method and core ideas of the present invention. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. An event-triggered recursive fuzzy neural network multivariable control method based on correlation entropy, characterized in that: The steps include: Acquiring sewage treatment process data; the sewage treatment process data includes dissolved oxygen concentration and nitrate nitrogen concentration; constructing an initial fuzzy neural network through the sewage treatment process data; The initial fuzzy neural network comprises: an input layer, a radial basis function layer, a rule layer, a normalization layer and an output layer; The input layer, the radial basis function layer, the regularization layer, the normalization layer and the output layer are connected in sequence; Constructing an event-triggered structural adjustment mechanism, and adjusting the structure of the initial fuzzy neural network through the event-triggered structural adjustment mechanism to obtain an optimized fuzzy neural network; The parameters of the optimized fuzzy neural network are updated by using the correlation entropy to obtain an updated fuzzy neural network; The updated fuzzy neural network is verified by a Lyapunov function to obtain a recursive fuzzy neural network; The recursive fuzzy neural network is used to control the variables of the sewage treatment process.

2. The event-triggered recursive fuzzy neural network multivariable control method based on correlation entropy according to claim 1 is characterized in that: The input variables of the input layer include: dissolved oxygen concentration error, dissolved oxygen concentration error change rate, nitrate nitrogen concentration error and nitrate nitrogen concentration error change rate; The calculation formula of the membership degree of the radial basis function layer is: Among them, a ij (t) is the membership of the i-th input variable to the j-th radial basis function, c ij (t) is the membership function center of the i-th input variable to the j-th radial basis function, σ ij (t) is the membership function width of the i-th input variable to the j-th radial basis function, o i (t) is the output vector of the i-th neuron in the input layer; The expression of the rule layer is: Among them, φ j (t) is the output of the jth rule layer node at the current moment, f j (t) is the sigmoid function, n is the number of neurons in the input layer; The expression of the normalization layer is: in, is the output of the jth normalization layer neuron, and P is the total number of radial basis function layer neurons; The expression of the output layer is: Among them, y q (t) is the output of the qth output layer node at time t, is the output weight, and Q is the number of output nodes.

3. The event-triggered recursive fuzzy neural network multivariable control method based on correlation entropy according to claim 1, characterized in that: Constructing an event-triggered structural adjustment mechanism, and adjusting the structure of the initial fuzzy neural network through the event-triggered structural adjustment mechanism to obtain an optimized fuzzy neural network, including: Construct the maximum relevant entropy criterion; Obtaining a loss function through the maximum relevant entropy criterion; Obtaining event definition through the loss function; The structure of the initial fuzzy neural network is adjusted according to the event definition to obtain the optimized fuzzy neural network.

4. The event-triggered recursive fuzzy neural network multivariable control method based on correlation entropy according to claim 3 is characterized in that: The calculation formula of the maximum correlation entropy criterion is: Among them, H(t) is the maximum correlation entropy, y k is the expected output, is the actual output, σ g is the kernel width, e k is the sample error, and K is the total number of samples.

5. The event-triggered recursive fuzzy neural network multivariable control method based on correlation entropy according to claim 4 is characterized in that: The calculation formula of the loss function is: Among them, J(t) is the loss function, e(t) is the composite error, σ g is the kernel width.

6. The event-triggered recursive fuzzy neural network multivariable control method based on correlation entropy according to claim 3 is characterized in that: The expression of the event definition is: Among them, Eve1 is the first event, Eve2 is the second event, Eve3 is the third event, J(t) is the loss function, ξ1 and ξ2 are both event triggering thresholds, δ is the time period, γ δ (t) is the change of loss function within δ, t is the current time, t E is the last triggering moment, T E is the maximum safety limit threshold.

7. The event-triggered recursive fuzzy neural network multivariable control method based on correlation entropy according to claim 6 is characterized in that: The initial fuzzy neural network is structurally adjusted according to the event definition to obtain the optimized fuzzy neural network, including: When the first event is triggered, the number of neurons in the regular layer is increased; When the second event is triggered, removing the normalization layer node; When the third event is triggered, the structure of the initial fuzzy neural network remains unchanged.

8. The event-triggered recursive fuzzy neural network multivariable control method based on correlation entropy according to claim 1, characterized in that: The parameters of the optimized fuzzy neural network are updated by using the correlation entropy to obtain an updated fuzzy neural network, specifically: the parameters of the optimized fuzzy neural network are updated by using the gradient descent method; the calculation formula of the gradient descent method is: Among them, Θ(t +1 ) is the parameter matrix of the next moment, Θ(t) is the parameter matrix of the current moment, β(t) is the positive learning rate, J k (t) is the cost function.

Images