A new neural network-based temperature control method

Abstract

This invention discloses a novel neural network-based auto-disturbance rejection (ADRS) temperature control method. Under conditions of unknown external environmental interference and internal system disturbances, a novel neural network with a five-layer network structure is used to optimize the ADRS temperature control deviation parameter and obtain the ADRS temperature control gain coefficient. Based on the ADRS temperature control gain coefficient, the temperature control signal is corrected in conjunction with a signal conditioner and a high-order deviation detector to adjust the cold chain temperature. The method includes the following steps: collecting current temperature data to obtain the ADRS temperature control deviation parameter; optimizing the neural network parameters of the novel neural network to obtain the optimal solution for the neural network parameters; optimizing the ADRS temperature control deviation parameter based on the optimal solution for the neural network parameters to obtain the ADRS temperature control gain coefficient; and correcting the temperature control signal based on the ADRS temperature control gain coefficient to adjust the cold chain temperature.

Description

Technical Field

[0001] The present invention relates to the field of temperature control, and in particular to an auto-disturbance rejection temperature control method based on a novel neural network. Background Art

[0002] As people's living standards continue to improve, the demand for green fresh food is growing. However, local high-quality fresh products are very limited and far from meeting the freshness needs of modern consumers. The cold chain is a logistics network equipped with specialized equipment to maintain the quality of fresh and frozen foods, keeping them at low temperatures throughout the production, transportation, and consumption process. Therefore, the demand for cold chain logistics for fresh products is increasing. However, most existing product cold chain logistics transportation temperature control systems can only provide an approximate temperature for transporting fresh products. The product temperature control during transportation is not precise enough and lacks real-time and accuracy. This leads to the corruption and deterioration of a large number of fresh products, resulting in huge losses and seriously affecting the construction and development of product cold chain logistics systems. Therefore, intelligent temperature control of products during cold chain transportation and the establishment of a comprehensive cold chain logistics temperature control system are crucial to the development of cold chain logistics.

[0003] Temperature control is an essential component of cold chain logistics, and optimizing the temperature during transportation is of great significance. Currently, temperature control systems in my country's cold chain logistics are still primarily based on traditional PID controllers. While these controllers achieve relatively satisfactory control results for general temperature control systems, refrigerated trucks often encounter unique conditions during transportation, which can interfere with the temperature control system and potentially cause instability. Due to the nonlinear and uncertain relationship between the control input and output of temperature control systems, the control performance of conventional linear PID controllers deteriorates when subjected to external disturbances or changes in the parameters of the controlled object, making satisfactory control impossible. Active disturbance rejection control (ADRC) inherits the classic PID control method and is independent of the controlled object. It is a novel control method developed to address the inherent shortcomings of the classic PID control method. ADRC boasts a simple algorithm and a wide range of parameter adaptability, making it an effective approach for addressing nonlinearity, uncertainty, strong disturbance coupling, and large time delays. It offers advantages such as strong adaptability, robustness, and ease of operation. However, the ADRC controller employing the ADRC algorithm requires numerous parameters to be tuned, which directly impacts controller performance. To this end, a fuzzy controller can be used to modify the parameters of the ADRC online. Fuzzy controllers are highly robust to the controlled object and process and can effectively control nonlinear systems. However, extracting rules from fuzzy controllers is difficult. By leveraging the learning properties of artificial neural network technology, fuzzy rules can be automatically extracted and fuzzy membership functions generated. This overcomes the difficulties of determining the neural network structure and the lack of self-learning in fuzzy control, making the fuzzy system adaptive and improving the adaptive capabilities of the ADRC. Summary of the Invention

[0004] The following is a brief summary of one or more aspects to provide a basic understanding of these aspects. This summary is not an exhaustive overview of all conceivable aspects and is neither intended to identify key or critical elements of all aspects nor to define the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form as a prelude to the more detailed description that will be provided later.

[0005] The present invention aims to address the difficulty of achieving satisfactory control results in traditional control systems due to the various interference conditions associated with temperature control. Combining the advantages and disadvantages of active disturbance rejection control (ADRC), this paper provides a novel neural network-based ADRC temperature control method. This method enables better control of the controlled object under conditions of unknown external environmental interference and internal system disturbances. Furthermore, the present invention employs a novel neural network optimized using the Raven algorithm to adjust the ADRC temperature control deviation parameters, thereby improving the controller's adaptability and reducing the difficulty of parameter tuning.

[0006] The technical solution of the present invention is:

[0007] The present invention provides a novel neural network-based auto-disturbance rejection temperature control method, comprising the following steps:

[0008] Collect current temperature data and obtain the temperature control deviation parameters of the active disturbance rejection;

[0009] Optimize the neural network parameters of the new neural network and obtain the optimal solution of the neural network parameters;

[0010] Optimize the ADRC temperature control deviation parameter based on the optimal solution of the neural network parameters to obtain the ADRC temperature control gain coefficient;

[0011] The temperature control signal is corrected based on the ADRC temperature control gain coefficient to adjust the cold chain temperature.

[0012] According to one embodiment of the novel neural network-based auto-interference rejection temperature control method of the present invention, the novel neural network-based temperature anti-interference control method collects current temperature data through a temperature sensor, and determines whether to perform temperature control based on the current temperature data; if so, the auto-interference rejection temperature control parameters are calculated based on the expected temperature parameters obtained by the signal conditioner and the high-order deviation detector; if not, the temperature data continues to be collected.

[0013] According to an embodiment of the self-adaptive temperature control method based on a new neural network of the present invention, the expected temperature parameters include an expected temperature approximation, an expected temperature differential signal, an expected temperature approximation detection value, and an expected temperature differential signal detection value; wherein, when the current temperature data exceeds a temperature range preset in the signal processor, the signal processor sends the expected temperature to the signal conditioner, and the expected temperature approximation and the expected temperature differential signal are calculated by the signal conditioner, and then the corresponding expected temperature approximation detection value and the expected temperature differential signal are obtained through the high-order deviation detector.

[0014] According to an embodiment of the active disturbance rejection temperature control method based on a novel neural network of the present invention, the active disturbance rejection temperature control deviation parameter includes an output deviation e1, an output differential deviation e2, and an integral of the output deviation e3. The active disturbance rejection temperature control deviation parameter is calculated using the desired temperature parameter, and the calculation formula is as follows:

[0015] e1=x1-z1

[0016] e2=x2-z2

[0017]

[0018] ; where x1 represents the expected temperature approximation,

[0019] x2 represents the expected temperature differential signal,

[0020] z1 represents the expected temperature approximate detection value,

[0021] z2 represents the expected temperature differential signal detection value,

[0022] t represents the desired temperature.

[0023] According to an embodiment of the novel neural network-based auto-disturbance rejection temperature control method of the present invention, the novel neural network uses a crow algorithm to iterate neural network parameters to obtain an optimal solution for the neural network parameters.

[0024] According to an embodiment of the novel neural network-based auto-disturbance rejection temperature control method of the present invention, the crow algorithm iteratively optimizes the neural network parameters, including the following steps:

[0025] Initialize the neural network parameters and the crow algorithm parameters, and randomly generate a crow population; wherein the crows are the neural network parameters to be optimized, and the crow algorithm parameters include the step size fl, the perception probability AP, the number of iterations iter, the crow population size N, the problem dimension D, the crow position and the crow memory;

[0026] Calculate and determine the optimal crow population, and randomly select tracking crows based on the optimal crow population;

[0027] Update crow positions and crow memories based on tracking crows;

[0028] Determine whether the current number of iterations has reached the maximum number of iterations; if so, output the current best crow position as the optimal solution for the neural network parameters; if not, recalculate and determine the best crow population for iterative update.

[0029] According to an embodiment of the novel neural network-based auto-disturbance rejection temperature control method of the present invention, the crow algorithm calculates and determines the optimal crow population size in each iteration using the following formula:

[0030]

[0031] , where iter represents the iter-th iteration,

[0032] K iter represents the optimal number of crow populations at the iter iteration,

[0033] K max represents the maximum number of optimal crow population,

[0034] K min represents the minimum number of optimal crow population,

[0035] iter maxIndicates the maximum number of iterations.

[0036] According to an embodiment of the novel neural network-based auto-disturbance rejection temperature control method of the present invention, the crow algorithm determines the crow position update mode by judging whether the tracking crow is discovered.

[0037] According to an embodiment of the novel neural network-based auto-disturbance rejection temperature control method of the present invention, when a tracking crow does not find itself being tracked, the crow's position is updated using the following formula:

[0038] X i,iter+1 =X i,iter +r i ×fl i,iter ×(M i,iter -X i,iter )×c1

[0039]

[0040] ; Among them, X i,iter+1 represents the position of crow i at the iter+1th iteration,

[0041] r i represents a random number between [0,1],

[0042] M i,iter represents the crow memory of crow i at the iterth iteration,

[0043] X i,iter represents the position of crow i at the iterth iteration,

[0044] fl i,iter represents the flight length of crow i at the iterth iteration,

[0045] c1 represents the convergence factor.

[0046] According to an embodiment of the novel neural network-based auto-disturbance rejection temperature control method of the present invention, when a tracking crow finds that it is being tracked, the position of crow i is updated using the following formula:

[0047] X i,iter+1 =a random position

[0048] ; Among them, X i,iter+1 represents the position of crow i at the iter+1th iteration,

[0049] X i,iter Indicates the crow position of crow i at iteration iter.

[0050] According to an embodiment of the novel neural network-based auto-disturbance rejection temperature control method of the present invention, after the crow algorithm updates the crow position based on the tracking target, the crow memory is updated using the following formula:

[0051]

[0052] , where M i,iter+1 represents the crow memory of crow i at the iter+1th iteration,

[0053] X i,iter+1 represents the position of crow i at the iter+1th iteration,

[0054] f(X i,iter+1 ) represents the crow position fitness of crow i in the iter+1th iteration,

[0055] X i,iter represents the position of crow i at the iterth iteration,

[0056] M i,iter Represents the crow memory of crow i at iteration iter.

[0057] According to one embodiment of the novel neural network-based auto-disturbance rejection temperature control method of the present invention, after the novel neural network uses the crow algorithm to iteratively obtain the optimal solution for the neural network parameters, the auto-disturbance rejection temperature control deviation parameters are optimized through novel neural network training to obtain the auto-disturbance rejection temperature control gain coefficient; wherein the optimal solution for the neural network parameters includes the optimal initial weight parameters and membership function parameters of the neural network.

[0058] According to one embodiment of the auto-disturbance rejection temperature control method based on a novel neural network of the present invention, the novel neural network includes an input layer, a fuzzification layer, a rule layer, a rule reinforcement layer, and a clarification layer. The five-layer novel neural network adaptively adjusts the input auto-disturbance rejection temperature control deviation parameter and outputs the auto-disturbance rejection temperature control gain coefficient.

[0059] According to an embodiment of the active disturbance rejection temperature control method based on a novel neural network of the present invention, the input layer includes the output deviation e1, the output differential deviation e2, and the integral of the output deviation e3; the membership functions of the fuzzification layer, whose inputs are the output deviation e1, the output differential deviation e2, and the integral of the output deviation e3, are F1, F2, and F3; wherein the fuzzification layer fuzzifies the active disturbance rejection temperature control deviation parameter using the following membership function:

[0060]

[0061] ; Among them, a and k are constant parameters of the neural network parameters. When x takes the value of e1, e2, and e3, the corresponding function values are F1, F2, and F3.

[0062] According to an embodiment of the novel neural network-based automatic disturbance rejection temperature control method of the present invention, the rule layer calculates the activation strength of the neuron fuzzy rule output by each node in the rule layer by the following formula:

[0063]

[0064] ; Where bj represents the activation strength of the neuron fuzzy rule output by node j,

[0065] i represents the number of neuron layers,

[0066] n represents the number of neuron nodes,

[0067] F ij x(n) represents the weighted value of the j-th node x in the i-th layer.

[0068] F ij is the weight,

[0069] x(n) represents the input value.

[0070] According to an embodiment of the novel neural network-based auto-disturbance rejection temperature control method of the present invention, the rule reinforcement layer strengthens the neuron fuzzy rule by the following formula:

[0071]

[0072] ; Among them, y j Represents the weighted result of the neuron fuzzy rule output by node j,

[0073] i represents the number of neuron layers,

[0074] n represents the number of neuron nodes,

[0075] F′ ij x(n′) represents the weighted value of the j-th node x in the i-th layer,

[0076] F′ ij is the weight,

[0077] x(n′) represents an input value.

[0078] According to an embodiment of the novel neural network-based auto-disturbance rejection temperature control method of the present invention, the rule reinforcement layer further strengthens the neuron fuzzy rules through an error function based on time domain difference:

[0079]

[0080]

[0081]

[0082] Among them, TD t represents the sample at time t,

[0083] γ represents the discount coefficient,

[0084] represents the reward value of the j-th neuron fuzzy rule of agent i at time t in state s,

[0085] represents the output action of agent i at time t,

[0086] rt represents the reinforcement value returned to the learning system by the environment at time t,

[0087] rV(s t+1 ) represents the reward that the agent can get after multiplying the discount coefficient,

[0088] v(s) represents the reward that agent i can obtain by adopting the jth rule.

[0089] According to an embodiment of the novel neural network-based automatic disturbance rejection temperature control method of the present invention, the output action of the agent is determined by multiple neuron fuzzy rules, and the output action of the agent is calculated by the following formula:

[0090]

[0091]

[0092] in, represents the output action of agent i,

[0093] represents the weight of each neuron fuzzy rule,

[0094] l represents the number of rules,

[0095] Represents each rule R l The action selected by the corresponding reinforcement function,

[0096] q t-1 (s,R l ,a) represents the reward value of the jth neuron fuzzy rule of agent i at time t-1 in state s.

[0097] According to an embodiment of the novel neural network-based auto-disturbance rejection temperature control method of the present invention, the neuron fuzzy rule weight is calculated using the following formula:

[0098]

[0099] ; Where k and b represent the adjustment parameters of the neuron fuzzy rule,

[0100] j represents the jth neuron,

[0101] represents the output action of agent j,

[0102] f represents the starting item of the rule,

[0103] Represents the action selected by the reinforcement function through the f-term rule.

[0104] According to an embodiment of the novel neural network-based auto-disturbance rejection temperature control method of the present invention, the neuron fuzzy rule is updated by the following formula:

[0105]

[0106] ;in, represents the reward value of the jth neuron fuzzy rule of agent i at time t+1 in state s,

[0107] represents the reward value of the j-th neuron fuzzy rule of agent i at time t in state s,

[0108] a represents the selected action,

[0109] Indicates A sample of the moment,

[0110] Represents the weight of each neuron fuzzy rule.

[0111] According to an embodiment of the novel neural network-based active disturbance rejection temperature control method of the present invention, the clarity layer performs clarity processing on each neural network output value using a clarity calculation method; wherein the neural network output value is the active disturbance rejection controller parameters k1, k2, and k3, and the calculation formula is as follows:

[0112]

[0113] ; where Z represents a set of integers,

[0114] u c (z) represents the degree of membership,

[0115] z1 represents the result after the first processing,

[0116] z2 represents the result after the second processing,

[0117] z3 represents the result after the third processing,

[0118] z0 represents the exact output result after sharpening.

[0119] According to an embodiment of the novel neural network-based auto-disturbance rejection temperature control method of the present invention, after the novel neural network outputs the auto-disturbance rejection temperature control gain coefficient through the clarity layer, a pruning index is used to determine whether to delete neurons in the rule layer; wherein the pruning index is calculated as follows:

[0120] I s (t+1)=I s (t)exp(τ)

[0121] ; Among them, I s represents the activation strength of the neuron,

[0122] I s (t+1) represents the activation intensity of the neuron at time t+1,

[0123] I s (t) represents the activation intensity of the neuron at time t,

[0124] τ represents the decay constant of the neuron.

[0125] According to an embodiment of the novel neural network-based automatic disturbance rejection temperature control method of the present invention, if the activation intensity of the neurons in the regular layer is I s If the value is less than the preset pruning threshold, the neurons in the regular layer are pruned and the neural network parameters are updated using the following formula:

[0126]

[0127] ; Where v represents the neuron with the smallest Euclidean distance from the pruned neuron s,

[0128] a v 、k v 、w v are the neural network parameters before neuron v is adjusted,

[0129] are the neural network parameters after neuron v is adjusted,

[0130] Represents the neural network parameters after the pruned neurons s are adjusted.

[0131] According to an embodiment of the active disturbance rejection temperature control method based on a novel neural network of the present invention, the novel neural network updates the neural network parameters by a gradient descent calculation method, and the calculation formula is as follows:

[0132]

[0133]

[0134]

[0135]

[0136] ; Where E(k) represents the objective function value of the new neural network,

[0137] a ij (k), k ij (k), w ij (k) represents the parameters of the jth neuron in the i-th layer at the k-th iteration,

[0138] a ij (k+1), k ij (k+1), w ij (k+1) represents the parameters of the jth neuron in the i-th layer at the k+1th iteration,

[0139] η represents the learning rate,

[0140] y(k) represents the predicted value,

[0141] y m (k) represents the true value.

[0142] According to an embodiment of the auto-disturbance rejection temperature control method based on a novel neural network of the present invention, after the novel neural network obtains the auto-disturbance rejection temperature control gain coefficient, the auto-disturbance rejection temperature control gain coefficient is sent to a signal correction module for temperature control signal correction; wherein the signal correction module includes a signal regulator and a high-order deviation detector.

[0143] According to an embodiment of the novel neural network-based auto-disturbance rejection temperature control method of the present invention, the signal conditioner combines the detection value of the high-order deviation detector to perform disturbance compensation on the control component. The calculation formula is as follows:

[0144]

[0145] ; where T1 represents the disturbance observed by the high-order deviation detector,

[0146] T2 represents the compensation given by the control law of the signal conditioner,

[0147] T0 represents the initial temperature,

[0148] m and M both represent disturbance coefficients.

[0149] According to an embodiment of the novel neural network-based auto-disturbance rejection temperature control method of the present invention, the signal conditioner introduces a tracking differentiator to construct a maximum signal conditioner, which extracts the input signal containing random noise and its differential signal through the tracking differentiator. The calculation formula of the maximum signal conditioner is as follows:

[0150]

[0151] ; Among them, u is the fastest control comprehensive function,

[0152] h is the integration step length,

[0153] x1(k) represents the tracking input signal of x,

[0154] x2(k) represents the first-order differential of x1(k),

[0155] x represents the input signal of the fastest signal conditioner,

[0156] r represents the speed factor,

[0157] h0 represents the filter factor of the signal conditioner.

[0158] According to an embodiment of the active disturbance rejection temperature control method based on a novel neural network of the present invention, the calculation formula of the fastest control comprehensive function u is as follows:

[0159]

[0160] ; Among them, d, d0, y, a0, a are all intermediate variables,

[0161] h0 represents the filter factor of the signal conditioner,

[0162] r represents the speed factor,

[0163] x represents the input signal.

[0164] According to an embodiment of the novel neural network-based auto-disturbance rejection temperature control method of the present invention, the signal correction module combines the order variable to construct a new high-order deviation detector, and the calculation formula is as follows:

[0165]

[0166] ; Among them, f x (), are perturbation functions,

[0167] represents the possible disturbance term.

[0168] According to an embodiment of the novel neural network-based auto-disturbance rejection temperature control method of the present invention, the signal correction module adjusts the new high-order deviation detector by adjusting the factor; wherein the calculation formula of the new high-order deviation detector is as follows:

[0169]

[0170] ; Among them, θ1, θ2, θ3 are adjustment factors,

[0171] β1, β2, β3, β4 are gain coefficients,

[0172] b0 is the control input coefficient,

[0173] z1, z2, z3, z4, and z5 are the estimated values of states x1, x2, x3, x4, and the total disturbance, respectively.

[0174] According to an embodiment of the novel neural network-based active disturbance rejection temperature control method of the present invention, the value range of the adjustment factor is as follows:

[0175]

[0176] According to an embodiment of the novel neural network-based active disturbance rejection temperature control method of the present invention, the signal correction module calculates a total disturbance estimate in real time using a high-order deviation detector, and compensates the control law using the total disturbance estimate; wherein the compensated control law is calculated using the following formula:

[0177]

[0178] ; Among them, k1, k2, k3 are gain coefficients,

[0179] e1, e2, e3 are the state errors of states x1, x2, x3,

[0180] z1, z2, z3 are the estimated values of states x1, x2, x3 respectively,

[0181] b0 is the control input coefficient,

[0182] u0 represents the control component,

[0183] U represents the final result of the control law.

[0184] Compared with the prior art, the present invention has the following beneficial effects: the present invention adopts a new neural network with a five-layer network structure to optimize the self-disturbance rejection temperature control deviation parameter, obtains the self-disturbance rejection temperature control gain coefficient, and then corrects the temperature control signal based on the self-disturbance rejection temperature control gain coefficient in combination with a signal conditioner and a high-order deviation detector to adjust the cold chain temperature. Through the new neural network of the present invention, online adjustment of the self-disturbance rejection temperature control parameters is achieved, and the robustness to the controlled object and the control process is increased. In addition, the present invention also adopts an improved crow algorithm to optimize the new neural network, adjust the network parameters, speed up the convergence speed of the new neural network, and avoid the situation where the neural network falls into a local optimal solution, thereby improving the adaptability of the temperature controller and reducing the difficulty of parameter setting. BRIEF DESCRIPTION OF THE DRAWINGS

[0185] The above features and advantages of the present invention will be better understood after reading the detailed description of the embodiments of the present disclosure in conjunction with the following drawings. In the drawings, the components are not necessarily drawn to scale, and components with similar related properties or characteristics may have the same or similar reference numerals.

[0186] Figure 1 1 is a general flow chart showing an embodiment of the auto-disturbance rejection temperature control method based on a novel neural network of the present invention.

[0187] Figure 2 FIG. 1 is a flow chart illustrating an embodiment of the crow algorithm of the present invention.

[0188] Figure 3 1 is an overall network diagram showing an embodiment of the novel neural network of the present invention. DETAILED DESCRIPTION

[0189] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Note that the various aspects described below with reference to the accompanying drawings and specific embodiments are merely exemplary and should not be construed as limiting the scope of protection of the present invention.

[0190] An embodiment of an auto-disturbance rejection temperature control method based on a novel neural network is disclosed herein. Figure 1 This is a general flow chart showing an embodiment of the present invention's automatic anti-disturbance temperature control method based on a novel neural network. Figure 1 ,The following is a detailed description of each step of the auto-disturbance rejection temperature control method based on the new neural network.

[0191] Step S1: Collect current temperature data and obtain ADSR temperature control deviation parameters.

[0192] In this embodiment, before temperature control is implemented, the refrigerated truck first collects current temperature data from the temperature sensor in the cold chain cargo box. Based on this data, it then determines whether to implement temperature control. If necessary, the auto-disturbance rejection (ADRS) temperature control parameters are calculated using the desired temperature parameters obtained by the signal conditioner and the high-order deviation detector. If not, the temperature sensor continues to collect temperature data.

[0193] In one embodiment, after the temperature sensor collects current temperature data, it can be transmitted to a signal processor via wireless sensor network technology. The signal processor then compares the received temperature data with a pre-set temperature range. If the cargo box temperature is suitable, monitoring continues. Otherwise, the signal processor and a high-order deviation detector calculate the active interference rejection temperature control parameters, preparing for subsequent temperature control processes.

[0194] Specifically, the desired temperature parameters include the desired temperature approximation, the desired temperature differential signal, the desired temperature approximation detection value, and the desired temperature differential signal detection value. When the current temperature data exceeds the temperature range preset in the signal processor, the signal processor sends the desired temperature T to the signal conditioner, which then calculates the desired temperature approximation x1 and the desired temperature differential signal x2. A high-order deviation detector then obtains the desired temperature approximation detection value z1 and the desired temperature differential signal detection value z2 corresponding to the desired temperature approximation x1 and the desired temperature differential signal x2. Finally, the ADRC temperature control deviation parameter is calculated using the following formula to prepare for the subsequent temperature control process:

[0195] e1=x1-z1

[0196] e2=x2-z2

[0197]

[0198] Here, t represents the desired temperature.

[0199] Step S2: Optimize the neural network parameters of the new neural network to obtain the optimal solution of the neural network parameters.

[0200] In this embodiment, a novel neural network is used to optimize and train the temperature control deviation parameters of the auto-disturbance rejection system. Before optimizing and training the temperature control deviation parameters of the auto-disturbance rejection system, the neural network parameters of the novel neural network must first be optimized to obtain the optimal solution for the neural network parameters.

[0201] In one embodiment, an improved crow algorithm is used to optimize the neural network parameters of the new neural network. Figure 2This is a flow chart showing an embodiment of the Crow Algorithm of the present invention. Figure 2 ,The following is a detailed description of each step of the crow algorithm.

[0202] Step S21: Initialize the neural network parameters and the crow algorithm parameters, and randomly generate a crow population. Crows represent the neural network parameters to be optimized, and the crow algorithm parameters include the step size fl, the perception probability AP, the number of iterations iter, the crow population size N, the problem dimension D, and the crow position X. i and Crow Memory M i .

[0203] Specifically, in this embodiment, after the crow algorithm completes the initialization of the neural network parameters and the crow algorithm parameters, the fitness value of the crow, i.e., the neural network parameters to be optimized, is obtained according to the fitness value function and the optimal solution is found, and the position of the optimal individual, i.e., the optimal initial weight parameter w of the neural network, is obtained. ij As well as the membership function parameters a and k, the adaptive adjustment of the neural network parameters is completed.

[0204] Step S22: Calculate and determine the best crow population, and randomly select tracking crows based on the best crow population.

[0205] In this embodiment, for each iteration of the minimization problem of crow i, the K best crows in the population are selected as the optimal crow population, and then crow i randomly selects a tracking crow in the K population to track. K is calculated as follows:

[0206]

[0207] Among them, iter represents the iter-th iteration, K iter Indicates the optimal number of crow populations for the iter iteration, K max represents the maximum number of optimal crow population, K min Indicates the minimum number of optimal crow population, iter max = represents the maximum number of iterations. During each crow operation, M of the best crows in the population are selected as the duck-billed crow population. The other crows then randomly select one of these crows to track. The crows' selection of the best crow population is rotated from large to small, avoiding the problem of poor tracking quality due to a large population or falling into a local optimum due to a small population. By selecting crows from the population of high-quality crows as tracking targets, the possibility of selecting inappropriate crows is reduced.

[0208] Step S23: Update the crow position and crow memory based on the tracked crow.

[0209] In this embodiment, in order to solve the problems of slow convergence speed and easy entrapment in local optimal solutions of the crow algorithm, a convergence factor c1 is introduced and the range M of the tracking group of crows for the next time is divided in advance to optimize the crow algorithm. Among them, the crow algorithm determines the position update method of crows by judging whether the tracking crow is found to be tracked.

[0210] Specifically, in a D-dimensional space, a crow group with a group size of M is distributed, and the vector (Xi, iter) represents the position of crow i in the iter-th iteration. Among them, the position update of crow i is divided into two states:

[0211] First, when the tracking crow is not found to be tracked, that is, when rj (the tracked crow j) ≥ AP (perception probability), the position of crow i is updated by the following formula:

[0212] X i,iter+1 =X i,iter +r i ×fl i,iter ×(M i,iter -X i,iter )×c1

[0213]

[0214] . Among them, X i,iter+1 represents the position of crow i in the (iter + 1)-th iteration, r i represents a random number between [0, 1], M i,iter represents the crow memory of crow i in the iter-th iteration, X i,iter represents the position of crow i in the iter-th iteration, fl i,iter represents the flight length of crow i in the iter-th iteration, and c1 represents the convergence factor. Among them, the convergence factor c1 can prevent the entire algorithm from being easily trapped in local optimality in the initial stage and improve the convergence speed in the later stage, thereby enhancing the efficiency of the algorithm.

[0215] Second: When the tracking crow is found to be tracked, that is, when rj (the tracked crow j) < AP (perception probability), the position of crow i is updated by the following formula:

[0216] X i,iter+1 =a random position

[0217] ; among them, X i,iter+1 represents the randomly updated position of crow i in the (iter + 1)-th iteration, and X i,iter represents the position of crow i in the iter-th iteration.

[0218] In addition, in this embodiment, after the crow algorithm updates the positions of all crows based on the tracking target, it updates the crow memory using the following formula:

[0219]

[0220] Among them, M i,iter+1 represents the crow memory of crow i at the iter+1th iteration, X i,iter+1 represents the position of crow i at the iter+1th iteration, f(X i,iter+1 ) represents the crow position fitness of crow i in the iter+1th iteration, X i ,iter represents the position of crow i at the iterth iteration, M i,iter Represents the crow memory of crow i at iteration iter.

[0221] Specifically, in this embodiment, when a crow updates its position, it checks the feasibility of the new position, evaluates the new position, and updates its memory. If the crow's position fitness at the new position is greater than or equal to the crow's memory fitness at the new position, the current crow memory is updated to the updated crow position. If not, the current crow memory is updated to the crow memory at the new position.

[0222] Step S24: Determine whether the current number of iterations has reached the maximum number of iterations; if so, output the current best crow position as the optimal solution for the neural network parameters; if not, recalculate and determine the best crow population for iterative update.

[0223] Step S3: Optimizing the ADRC temperature control deviation parameter based on the optimal solution of the neural network parameters to obtain the ADRC temperature control gain coefficient.

[0224] In this embodiment, after the new neural network uses the crow algorithm to iterate and obtain the optimal solution of the neural network parameters, the new neural network training is used to optimize the temperature control deviation parameters of the active disturbance rejection and obtain the temperature control gain coefficient of the active disturbance rejection. The optimal solution of the neural network parameters includes the optimal initial weight parameter w of the neural network. ij And membership function parameters a, k.

[0225] Specifically, in this embodiment, the optimal initial weight parameter w of the neural network is obtained by the improved crow algorithm. ij After the membership function parameters a and k are obtained, the deviation obtained by the signal conditioner is input into the new neural network. The input data is processed by a 5-layer network model, and the error back propagation is used to further adjust the weights to complete the adjustment of the self-disturbance rejection temperature control deviation parameters. Figure 3 This is an overall network diagram showing an embodiment of the novel neural network of the present invention. Figure 3, the present invention is further described.

[0226] like Figure 3 As shown, in this embodiment, the new neural network includes an input layer, a fuzzification layer, a rule layer, a rule reinforcement layer and a clarification layer. The membership function parameters and output weight parameters obtained by the improved crow algorithm are applied through the new neural network with a five-layer structure.

[0227] Specifically, in this embodiment, before adjusting the self-disturbance rejection temperature control deviation parameters, the initial DO setting value, learning rate η, pruning threshold I, and the initial value of the number of neurons in the rule layer of the new neural network are first set to 7. Then the new neural network is trained using simulation data. The network input is the output deviation e1, the output differential deviation e2, and the integral of the output deviation e33. The input fuzzification layer is fuzzified to obtain the fuzzy quantity. At the same time, the neuron fuzzy rules are strengthened using the reinforcement learning formula of the rule strengthening layer, and the rule layer parameter values are updated. In addition, this embodiment can also determine whether the rule layer neurons meet the deletion conditions based on the pruning index. If so, the current rules are deleted and the neuron parameters are updated. Then, the network output at this moment is obtained by clarification. If the data training is completed, the loop ends; otherwise, the new neural network is repeatedly trained using simulation data. The following is a detailed description of the adjustment of each layer of the new neural network:

[0228] The first layer is the input layer, and the inputs are the output deviation e1, the output differential deviation e2 and the integral of the output deviation e3 to the ADRC.

[0229] The second layer is the fuzzification layer, which takes as input the membership functions F1, F2, and F3 of the output deviation e1, the output differential deviation e2, and the integral of the output deviation e3, and is used to fuzzify the input variables. The fuzzification layer fuzzifies the auto-disturbance rejection temperature control deviation parameter using the following membership functions:

[0230]

[0231] Among them, a and k are constant parameters of the neural network parameters, which are used to control the shape of the function. When x takes the value of e1, e2, e3, the corresponding function value is F1, F2, F3

[0232] The third layer is the rule layer, where the neuron fuzzy rules are converted from the reinforcement function of Q reinforcement learning. The output of each node in the rule layer can be expressed as the activation strength of the fuzzy rule, calculated as follows:

[0233]

[0234] Where bj represents the activation strength of the neuron fuzzy rule output by node j, i represents the number of neuron layers, for example, i=2 represents the second layer of neurons. n represents the number of neuron nodes in this layer, i.e., the number of neurons, F ij x(n) represents the weighted value of the jth node x in the i-th layer. ij is the weight, x(n) represents the input value, which is generally the output value of the previous layer.

[0235] The fourth layer is the rule reinforcement layer, which strengthens the neuron fuzzy rules through an effective reinforcement mechanism. The calculation formula is as follows:

[0236]

[0237] Among them, y j represents the weighted result of the neuron fuzzy rule output by node j, i represents the number of neuron layers, for example, i=3 represents the third layer of neurons. n represents the number of neuron nodes in this layer, i.e., the number of neurons, F′ ij x(n′) represents the weighted value of the jth node x in the i-th layer. F′ ij is the weight, and x(n′) represents the input value.

[0238] In one embodiment, the rule reinforcement layer can further train the neuron fuzzy rules through an error function based on time domain difference, thereby strengthening the neuron fuzzy rules and improving network performance. The calculation formula is as follows:

[0239]

[0240]

[0241]

[0242] Among them, TD t represents the sample at time t, γ represents the discount coefficient, Represents the reward value of the fuzzy rule of the jth neuron of agent i at time t in state s. The agent is a decision maker to decide which behavior to perform at the next moment. Therefore, represents the output action of agent i at time t, rt represents the reinforcement value returned to the learning system by the environment at time t, rV(s t+1 ) represents the reward that the agent can obtain after multiplying the discount coefficient, and v(s) represents the reward that agent i can obtain by adopting the jth rule.

[0243] In addition, in this embodiment, the output action of the agent is determined by multiple neuron fuzzy rules, and the output action of the agent is calculated by the following formula:

[0244]

[0245]

[0246] in, represents the output action of agent i, represents the weight of each neuron fuzzy rule, l represents the number of rules, Represents each rule R l The action selected by the corresponding reinforcement function, q t-1 (s,R l ,a) represents the reward value of the jth neuron fuzzy rule of agent i at time t-1 under state s. Among them, the neuron fuzzy rule weight Calculated using the following formula:

[0247]

[0248] Among them, k and b represent the adjustment parameters of the neuron fuzzy rule, j represents the jth neuron, represents the output action of agent j, f represents the starting item of the rule, represents the action selected by the reinforcement function through the f-term rule. At this time, the update function of the neuron fuzzy rule is:

[0249]

[0250] in, represents the reward value of the jth neuron fuzzy rule of agent i at time t+1 in state s, represents the reward value of the jth neuron fuzzy rule of agent i at time t in state s, a represents the selected action, Indicates A sample of the moment, Represents the weight of each neuron fuzzy rule.

[0251] The fifth layer is the clarity layer, which outputs the parameters k1, k2, and k3 required by the ADRC. The clarity calculation method is used to clarify the output values of each neural network. The output values of the neural network are the parameters k1, k2, and k3 of the ADRC, and the calculation formula is as follows:

[0252]

[0253] Where Z represents the set of integers, u c(z) represents the degree of membership, z1 represents the result of the first processing, z2 represents the result of the second processing, z3 represents the result of the third processing, and z0 represents the precise output after clarity, i.e., the final result. Through the above clarity calculation method, the output value can be clarified, and the fuzzy quantity can be selected according to a certain method to select a precise value that can represent the inference result as the output.

[0254] In addition, in this embodiment, after the new neural network outputs the self-disturbance rejection temperature control gain coefficient through the clarity layer, it also uses the pruning index to determine whether to delete the neurons in the rule layer. The pruning index calculation formula is as follows:

[0255] I s (t+1)=I s (t)exp(τ)

[0256] ; Among them, I s Indicates the activation strength of the neuron, I s (t+1) represents the activation intensity of the neuron at time t+1, I s (t) represents the activation strength of the neuron, and τ represents the decay constant of the neuron.

[0257] Specifically, if the activation intensity of the neurons in the regular layer is I s If the value is less than the preset pruning threshold, the neurons in the regular layer are pruned and the neural network parameters are updated using the following formula:

[0258]

[0259] Among them, v represents the neuron with the smallest Euclidean distance from the pruned neuron s, a v 、k v 、w v are the neural network parameters before neuron v is adjusted, are the neural network parameters after neuron v is adjusted, Represents the neural network parameters after the pruned neurons s are adjusted.

[0260] In addition, in this embodiment, the new neural network uses the gradient descent method with the introduction of momentum terms to enable the system to obtain a new neural network model with better parameters. The new neural network updates the neural network parameters through the gradient descent calculation method, and the calculation formula is as follows:

[0261]

[0262]

[0263]

[0264]

[0265] Among them, E(k) is the objective function of the new neural network, a ij (k), k ij (k), w ij (k) represents the parameter of the jth neuron in the i-th layer at the k-th iteration, a ij (k+1), k ij (k+1), w ij (k+1) represents the parameter of the jth neuron in the i-th layer at the k+1-th iteration, η represents the learning rate, y(k) represents the predicted value, and ym(k) represents the true value.

[0266] Step S4: Correct the temperature control signal based on the ADRC temperature control gain coefficient to adjust the cold chain temperature.

[0267] In this embodiment, the new neural network acquires the ADRC temperature control gain coefficients and sends them to the signal correction module for temperature control signal correction. The signal correction module, which includes a signal conditioner and a high-order deviation detector, applies the optimized parameters obtained from training to the control strategy, thereby completing the cold chain cargo box temperature control process.

[0268] Specifically, in this embodiment, the signal conditioner combines the detection values of the high-order deviation detector to perform disturbance compensation on the control component. A new control term is introduced into the signal correction law, enabling it to better control and compensate for temperature disturbances based on the error between the given signal and its derivative obtained by the signal conditioner and the system output and output derivative observed by the high-order deviation detector. The signal conditioner, high-order deviation detector, and signal correction law process data as follows:

[0269]

[0270] Where T1 represents the disturbance observed by the high-order deviation detector, T2 represents the compensation given by the control law of the signal conditioner, T0 represents the initial temperature, and m and M both represent the disturbance coefficients.

[0271] In addition, in this embodiment, the signal conditioner introduces a tracking differentiator to construct a fastest signal conditioner, which extracts the input signal containing random noise and its differential signal during the pre-arranged transition process, which can more smoothly resolve the contradiction between overshoot and rapidity. The calculation formula of the fastest signal conditioner is as follows:

[0272]

[0273] Where u is the fastest control synthesis function, h is the integration step size, x1(k) represents the tracking input signal of x, x2(k) represents the first-order differential of x1(k), x represents the input signal of the fastest signal conditioner, r represents the speed factor, and h0 represents the filter factor of the signal conditioner.

[0274] At the same time, this embodiment also introduces a new control function to improve the tracking differentiator. After the system enters steady state, the original maximum signal conditioner will produce high-frequency vibration. Therefore, a new maximum control synthesis function fhan(x1, x2, r, h0) is introduced. Its algorithm is as follows:

[0275]

[0276] Where d, d0, y, a0, and a are intermediate variables. h0 represents the filter factor of the signal conditioner, which acts as a filter. r represents the speed factor; a larger value indicates a faster approximation. x represents the input signal. By incorporating the new maximum speed control function into the signal conditioner and implementing the transition process through a tracking differentiator, the system can be optimized to reduce initial error and initial impact on the system, effectively resolving the conflict between overshoot and rapidity.

[0277] In this embodiment, the high-order deviation detector can estimate the system's feedback disturbance based on the system's feedback information and internal state information, and estimate the total disturbance online in real time. In order to better adjust the high-order deviation detector and improve its working efficiency, the signal correction module combines the order variable to construct a new high-order deviation detector, which has better anti-interference performance than the traditional high-order deviation detector and improves its tracking and estimation capabilities. Among them, the order variables x3 and x 4d The calculation formula is as follows:

[0278]

[0279] Among them, f x () is the perturbation function, For and f x () Different perturbation functions, Indicates possible disturbance terms. In actual calculations, there may be four disturbance terms, or other numbers. Here, four are used as an example, so four letters are used to represent them. Therefore, the various variables of the high-order deviation detector can be expressed by the following formula:

[0280]

[0281] . Where b is the control gain.

[0282] In addition, in this embodiment, the signal correction module can also adjust the new high-order deviation detector through the adjustment factor. The calculation formula of the new high-order deviation detector is as follows:

[0283]

[0284] Among them, θ1, θ2, θ3 are adjustment factors, β1, β2, β3, β4 are gain coefficients, b0 is the control input coefficient, z1, z2, z3, z4, z5 are the estimated values of states x1, x2, x3, x4 and the total disturbance respectively.

[0285] In order to ensure the stability of the system, the value range of the adjustment factor is as follows:

[0286]

[0287] In actual operation, we can first set θ1, θ2, and θ3 to 1 and roughly adjust the system to achieve a relatively reasonable performance of the observer. Then, we can fine-tune the values of fine adjustment and according to the range of the adjustment factors mentioned above to further improve the performance of the observer.

[0288] In addition, the signal correction module in this embodiment calculates the total disturbance estimate in real time through a high-order deviation detector, and uses the total disturbance estimate to compensate the control law to achieve anti-interference function. In order to further improve the performance of the control law, a new integral term is added to improve the control law. The calculation formula is as follows:

[0289]

[0290] Among them, k1, k2, k3 are gain coefficients, e1, e2, e3 are state errors of states x1, x2, x3, z1, z2, z3 are estimated values of states x1, x2, x3 respectively, b0 is the control input coefficient, u0 represents a certain form of control component, and U represents the final result of the control law.

[0291] The previous description of the disclosure is provided to enable any person skilled in the art to make or use the disclosure. Various modifications to the disclosure will be apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit or scope of the disclosure. Thus, the disclosure is not intended to be limited to the examples and designs described herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0292] Those skilled in the art will further appreciate that the various illustrative logic blocks, modules, circuits, and algorithmic steps described in conjunction with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or a combination of the two. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps are generally described above in terms of their functionality. Whether such functionality is implemented as hardware or software depends on the specific application and the design constraints imposed on the overall system. A skilled person may implement the described functionality in different ways for each specific application, but such implementation decisions should not be interpreted as resulting in a departure from the scope of the present invention.

[0293] The various illustrative logic blocks, modules, and circuits described in conjunction with the embodiments disclosed herein may be implemented or performed using a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general-purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration.

[0294] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. The software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor so that the processor can read and write information from / to the storage medium. In an alternative, the storage medium may be integrated into the processor. The processor and storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and storage medium may reside in a user terminal as discrete components.

[0295] In one or more exemplary embodiments, the functions described may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software as a computer program product, the functions may be stored on or transmitted via a computer-readable medium as one or more instructions or codes. Computer-readable media include both computer storage media and communication media, including any media that facilitates the transfer of a computer program from one location to another. A storage medium may be any available medium that can be accessed by a computer. By way of example and not limitation, such computer-readable media may include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer. Any connection is also properly referred to as a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using a coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwaves, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwaves are included in the definition of medium. As used herein, disk and disc include compact disc (CD), laser disc, optical disc, digital versatile disc (DVD), floppy disk and Blu-ray disc, where disks typically reproduce data magnetically, while discs reproduce data optically with lasers. Combinations of the above should also be included within the scope of computer-readable media.

[0296] The previous description of the disclosure is provided to enable any person skilled in the art to make or use the disclosure. Various modifications to the disclosure will be apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit or scope of the disclosure. Thus, the disclosure is not intended to be limited to the examples and designs described herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A novel neural network-based auto-disturbance rejection temperature control method, characterized in that: The following steps are included: Collect current temperature data and obtain the temperature control deviation parameters of the active disturbance rejection; Optimize the neural network parameters of the new neural network and obtain the optimal solution of the neural network parameters; Optimize the ADRC temperature control deviation parameter based on the optimal solution of the neural network parameters to obtain the ADRC temperature control gain coefficient; Correct the temperature control signal based on the ADRC temperature control gain coefficient to adjust the cold chain temperature; The novel neural network adopts the crow algorithm to iterate the neural network parameters to obtain the optimal solution of the neural network parameters; The crow algorithm iteratively optimizes the neural network parameters, including the following steps: Initialize the neural network parameters and the crow algorithm parameters, and randomly generate a crow population; wherein the crows are the neural network parameters to be optimized, and the crow algorithm parameters include the step size fl, the perception probability AP, the number of iterations iter, the crow population size N, the problem dimension D, the crow position and the crow memory; Calculate and determine the optimal crow population, and randomly select tracking crows based on the optimal crow population; Update crow positions and crow memories based on tracking crows; Determine whether the current number of iterations has reached the maximum number of iterations; if so, output the current best crow position as the optimal solution for the neural network parameters; if not, recalculate and determine the best crow population for iterative update; The crow algorithm calculates the optimal crow population size in each iteration using the following formula: Among them, iter represents the iter-th iteration, K iter represents the optimal number of crow populations at the iter iteration, K max represents the maximum number of optimal crow population, K min represents the minimum number of optimal crow population, iter max Indicates the maximum number of iterations; The crow algorithm determines the crow position update mode by judging whether the crow is found and tracked; When the tracking crow does not find that it is being tracked, the crow's position is updated using the following formula: X i,iter+1 =X i,iter +r i ×fl i,iter ×(M i,iter -X i,iter )×c1 Among them, X i,iter+1 represents the position of crow i at the iter+1th iteration, r i represents a random number between [0,1], M i,iter represents the crow memory of crow i at the iterth iteration, X i,iter represents the position of crow i at the iterth iteration, fl i,iter represents the flight length of crow i at the iterth iteration, c1 represents the convergence factor; After the novel neural network uses the crow algorithm to iteratively obtain the optimal solution for the neural network parameters, the new neural network is trained to optimize the auto-disturbance rejection temperature control deviation parameter to obtain the auto-disturbance rejection temperature control gain coefficient; wherein the optimal solution for the neural network parameters includes the optimal initial weight parameters and membership function parameters of the neural network; The novel neural network includes an input layer, a fuzzification layer, a rule layer, a rule reinforcement layer, and a clarification layer. The five-layer structure of the novel neural network adaptively adjusts the input auto-disturbance rejection temperature control deviation parameter and outputs the auto-disturbance rejection temperature control gain coefficient. The input layer includes the output deviation e1, the output differential deviation e2, and the integral of the output deviation e3. The membership functions of the fuzzification layer, whose inputs are the output deviation e1, the output differential deviation e2, and the integral of the output deviation e3, are F1, F2, and F3. The fuzzification layer fuzzifies the auto-disturbance rejection temperature control deviation parameter using the following membership function: Among them, a and k are constant parameters of the neural network parameters. When x takes the value of e1, e2, and e3, the corresponding function values are F1, F2, and F3; The rule reinforcement layer further strengthens the neuron fuzzy rules through an error function based on time domain difference: Among them, TD t represents the sample at time t, γ represents the discount coefficient, represents the reward value of the j-th neuron fuzzy rule of agent i at time t in state s, represents the output action of agent i at time t, rt represents the reinforcement value returned to the learning system by the environment at time t, rV(s t+1 ) represents the reward that the agent can get after multiplying the discount coefficient, v(s) represents the reward that agent i can obtain by adopting the jth rule; The output action of the agent is determined by multiple neuron fuzzy rules and is calculated using the following formula: in, represents the output action of agent i, represents the weight of each neuron fuzzy rule, l represents the number of rules, Represents each rule R l The action selected by the corresponding reinforcement function, q t-1 (s,R l ,a) represents the reward value of the jth neuron fuzzy rule of agent i at time t-1 in state s; The neuron fuzzy rule weight is calculated using the following formula: Among them, k and b represent the adjustment parameters of the neuron fuzzy rule, j represents the jth neuron, represents the output action of agent j, f represents the starting item of the rule, represents the action selected by the reinforcement function through the f-term rule; The clarification layer clarifies the output values of each neural network by a clarification calculation method; the output values of the neural network are the parameters k1, k2 and k3 of the active disturbance rejection controller, and the calculation formula is as follows: Where Z represents the set of integers, u c (z) represents the degree of membership, z1 represents the result after the first processing, z2 represents the result after the second processing, z3 represents the result after the third processing, z0 represents the precise output result after clarity; After the novel neural network obtains the self-disturbance rejection temperature control gain coefficient, the self-disturbance rejection temperature control gain coefficient is sent to the signal correction module for temperature control signal correction; wherein the signal correction module includes a signal conditioner and a high-order deviation detector; The signal conditioner performs disturbance compensation on the control component in combination with the detection value of the high-order deviation detector, and the calculation formula is as follows: Where T1 represents the disturbance observed by the high-order deviation detector, T2 represents the compensation given by the control law of the signal conditioner, T0 represents the initial temperature, m and M both represent disturbance coefficients; The signal conditioner introduces a tracking differentiator to construct a fastest signal conditioner, and extracts the input signal containing random noise and its differential signal through the tracking differentiator. The calculation formula of the fastest signal conditioner is as follows: Among them, u is the fastest control comprehensive function, h is the integration step length, x1(k) represents the tracking input signal of x, x2(k) represents the first-order differential of x1(k), x represents the input signal of the fastest signal conditioner, r represents the speed factor, h0 represents the filter factor of the signal conditioner; Among them, the calculation formula of the fastest control comprehensive function u is as follows: Among them, d, d0, y, a0, and a are all intermediate variables. h0 represents the filter factor of the signal conditioner, r represents the speed factor, x represents the input signal.

2. The novel neural network-based auto-disturbance rejection temperature control method according to claim 1, characterized in that: The temperature anti-interference control method based on the novel neural network collects current temperature data through a temperature sensor and determines whether to perform temperature control based on the current temperature data; if so, the self-anti-interference temperature control parameters are calculated based on the expected temperature parameters obtained by the signal regulator and the high-order deviation detector; if not, the temperature data collection continues.

3. The novel neural network-based auto-disturbance rejection temperature control method according to claim 2, characterized in that: The expected temperature parameters include an expected temperature approximation, an expected temperature differential signal, an expected temperature approximation detection value, and an expected temperature differential signal detection value; wherein, when the current temperature data exceeds the temperature range preset in the signal processor, the signal processor sends the expected temperature to the signal conditioner, and the signal conditioner calculates the expected temperature approximation and the expected temperature differential signal, and then obtains the corresponding expected temperature approximation detection value and expected temperature differential signal detection value through the high-order deviation detector.

4. The novel neural network-based auto-disturbance rejection temperature control method according to claim 2, characterized in that: The ADRC temperature control deviation parameter includes the output deviation e1, the output differential deviation e2, and the integral of the output deviation e3. The ADRC temperature control deviation parameter is calculated using the desired temperature parameter. The calculation formula is as follows: e1=x1-z1 e2=x2-z2 e3=∫0 t e1(t)dπ Where x1 represents the expected temperature approximation, x2 represents the expected temperature differential signal, z1 represents the expected temperature approximate detection value, z2 represents the expected temperature differential signal detection value, t represents the desired temperature.

5. The novel neural network-based auto-disturbance rejection temperature control method according to claim 1, characterized in that: When the tracking crow finds that it is being tracked, the position of crow i is updated by the following formula: X i,iter+1 =arandomposition Among them, X i,iter+1 represents the position of crow i at the iter+1th iteration, X i,iter Indicates the crow position of crow i at iteration iter.

6. The novel neural network-based auto-disturbance rejection temperature control method according to claim 1, characterized in that: After the crow algorithm updates the crow's position based on the tracking target, it updates the crow's memory using the following formula: Among them, M i,iter+1 represents the crow memory of crow i at the iter+1th iteration, X i,iter+1 represents the position of crow i at the iter+1th iteration, f(X i,iter+1 ) represents the crow position fitness of crow i in the iter+1th iteration, X i,iter represents the position of crow i at the iterth iteration, M i,iter Represents the crow memory of crow i at iteration iter.

7. The novel neural network-based auto-disturbance rejection temperature control method according to claim 1, characterized in that: The rule layer calculates the activation strength of the neuron fuzzy rule output by each node in the rule layer by the following formula: Among them, bj represents the activation strength of the neuron fuzzy rule output by node j, i represents the number of neuron layers, n represents the number of neuron nodes, F ij x(n) represents the weighted value of the j-th node x in the i-th layer. F ij is the weight, x(n) represents the input value.

8. The novel neural network-based auto-disturbance rejection temperature control method according to claim 1, characterized in that: The rule reinforcement layer strengthens the neuron fuzzy rules by the following formula: Among them, y j Represents the weighted result of the neuron fuzzy rule output by node j, i represents the number of neuron layers, n represents the number of neuron nodes, F′ ij x(n') represents the weighted value of the jth node x in the i-th layer, F′ ij is the weight, x(n′) represents an input value.

9. The novel neural network-based auto-disturbance rejection temperature control method according to claim 1, characterized in that: The neuron fuzzy rule is updated by the following formula: in, represents the reward value of the jth neuron fuzzy rule of agent i at time t+1 in state s, represents the reward value of the j-th neuron fuzzy rule of agent i at time t in state s, a represents the selected action, Indicates A sample of the moment, Represents the weight of each neuron fuzzy rule.

10. The novel neural network-based auto-disturbance rejection temperature control method according to claim 1, characterized in that: After the novel neural network outputs the auto-disturbance rejection temperature control gain coefficient through the clarity layer, it determines whether to delete neurons in the rule layer through the pruning index. The pruning index calculation formula is as follows: I s (t+1)=I s (t)exp(τ) Among them, I s represents the activation strength of the neuron, I s (t+1) represents the activation intensity of the neuron at time t+1, I s (t) represents the activation intensity of the neuron at time t, τ represents the decay constant of the neuron.

11. The novel neural network-based auto-disturbance rejection temperature control method according to claim 10, characterized in that: If the activation intensity of the neurons in the regular layer is I s If the value is less than the preset pruning threshold, the neurons in the regular layer are pruned and the neural network parameters are updated using the following formula: Among them, v represents the neuron with the smallest Euclidean distance from the pruned neuron s, a v 、k v 、w v are the neural network parameters before neuron v is adjusted, are the neural network parameters after neuron v is adjusted, Represents the neural network parameters after the pruned neurons s are adjusted.

12. The novel neural network-based auto-disturbance rejection temperature control method according to claim 1, characterized in that: The new neural network updates the neural network parameters by the gradient descent calculation method, and the calculation formula is as follows: Among them, E(k) represents the objective function value of the new neural network, a ij (k), k ij (k), w ij (k) represents the parameters of the jth neuron in the i-th layer at the k-th iteration, a ij (k+1), k ij (k+1), w ij (k+1) represents the parameters of the jth neuron in the i-th layer at the k+1th iteration, η represents the learning rate, y(k) represents the predicted value, y m (k) represents the true value.

13. The novel neural network-based auto-disturbance rejection temperature control method according to claim 1, characterized in that: The signal correction module combines the order variable to construct a new high-order deviation detector, and the calculation formula is as follows: Among them, f x (), are perturbation functions, t,w represent possible disturbance terms.

14. The novel neural network-based auto-disturbance rejection temperature control method according to claim 1, characterized in that: The signal correction module adjusts the new high-order deviation detector by adjusting the factor; wherein the calculation formula of the new high-order deviation detector is as follows: Among them, θ1, θ2, and θ3 are adjustment factors. β1, β2, β3, β4 are gain coefficients, b0 is the control input coefficient, z1, z2, z3, z4, and z5 are the estimated values of states x1, x2, x3, x4, and the total disturbance, respectively.

15. The novel neural network-based auto-disturbance rejection temperature control method according to claim 1, characterized in that: The value range of the adjustment factor is as follows:

16. The novel neural network-based auto-disturbance rejection temperature control method according to claim 1, characterized in that: The signal correction module calculates the total disturbance estimate in real time through a high-order deviation detector and compensates the control law based on the total disturbance estimate. The compensated control law is calculated using the following formula: Among them, k1, k2, k3 are gain coefficients, e1, e2, e3 are the state errors of states x1, x2, x3, z1, z2, z3 are the estimated values of states x1, x2, x3, b0 is the control input coefficient, u0 represents the control component, U represents the final result of the control law.

Images