Method for adaptive control of a device for sludge treatment

Abstract

The present application relates to the technical field of sludge treatment, and discloses a self-adaptive control method for a device for sludge treatment, comprising: acquiring data of a cleaning device for sludge treatment, wherein the data comprises various environmental and operating state parameters capable of monitoring in real time during device operation; constructing a generative adversarial network algorithm based on feature fusion and training to obtain a data expansion model after training; constructing a device state recognition model based on a federated learning architecture; monitoring the state of the sludge treatment device through the device state recognition model, and then making a corresponding control strategy. The present application has the beneficial effect that the state of the sludge treatment device is monitored through the device state recognition model, and then a corresponding control strategy is made.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sludge treatment, and particularly relates to a self-adaptive control method for a sludge treatment device. BACKGROUND

[0002] During the operation of a sludge treatment device, state monitoring and fault diagnosis have important influences on the safety, operation efficiency and maintenance cost of the device. However, the operation environment of the sludge treatment device is complex and changeable, involving multiple operation parameters (such as temperature, pressure, humidity, vibration, etc.) and multiple fault types (such as slight faults and serious faults). Precise identification and control of the device state are of great significance to optimize the operation efficiency, reduce the fault downtime and reduce the maintenance cost. In the prior art, device state identification and fault diagnosis mainly rely on centralized model training, which requires the centralized transmission of original operation data of the device to a central server for processing. However, the operation data of the sludge treatment device often involves sensitive information, and centralized transmission may lead to the problem of data privacy leakage. In addition, the operation data collection of the sludge treatment device has the phenomenon of insufficient sample quantity, especially under different devices and working conditions, the sample scarcity limits the generalization ability of the model. To solve the problem of insufficient data, the traditional data expansion method cannot effectively generate diversified high-quality data samples, which deviates significantly from the distribution of the actual device operation state, further affecting the identification accuracy of the model.

[0003] The existing Chinese invention patent CN117105502A proposes a self-adaptive sludge treatment control system and control method based on machine learning, which belongs to the technical field of sludge treatment. It solves the technical problems of performance and efficiency reduction, and even faults or accidents in the sludge treatment process. The sludge treatment control system includes a sludge treatment device for anaerobic digestion, dewatering and drying treatment of the incoming sludge; a sensor module arranged at different positions of the sludge treatment device for detecting various parameters of the sludge, including temperature, pressure, humidity, pH value, organic matter content and ammonia nitrogen content; a control module connected with the sensor module and the sludge treatment device for adjusting various operation parameters of the sludge treatment device according to the sludge parameters detected by the sensor module. The sludge treatment control method includes four steps. The present application can realize adaptive control of different types and states of sludge, and improve the energy consumption, efficiency and quality indicators of sludge treatment.

[0004] Although the above technical solution can realize adaptive control of different types and states of sludge, the following problems still need to be further solved:

[0005] 1. In the adaptive control task of the sludge treatment equipment, the traditional data augmentation method generates low-quality samples in the case of less sludge treatment equipment data samples, lacks consistent feature distribution with real data, and leads to insufficient generalization ability of model training.

[0006] 2. In the adaptive control task of the sludge treatment equipment, the traditional neural network optimization method mostly uses gradient descent algorithm, which is easy to fall into local optimum in complex nonlinear problems, and has low training efficiency, which affects the feature extraction effect and classification accuracy of sludge treatment equipment state recognition. SUMMARY

[0007] To solve the problems in the prior art, the present application provides a device adaptive control method for sludge treatment, which monitors the state of the sludge treatment equipment through a device state recognition model, and then makes a corresponding control strategy.

[0008] To achieve the above purpose, the following technical solutions are used:

[0009] A device adaptive control method for sludge treatment, characterized in that:

[0010] Data of a cleaning device for sludge treatment is obtained, which includes various environmental and operating state parameters that can be monitored in real time during the operation of the device;

[0011] A feature fusion-based generative adversarial network algorithm is constructed and trained to obtain a data augmentation model after training;

[0012] A device state recognition model based on a federated learning architecture is constructed, which has multiple clients, each client has a local device state recognition model, and the local device state recognition model is trained on the sludge treatment equipment data of the respective client, and the local device state recognition model and the central device state recognition model interact with each other;

[0013] The state of the sludge treatment equipment is monitored by the device state recognition model, and then a corresponding control strategy is made.

[0014] Further, the collected data includes but is not limited to the operating temperature, pressure, speed, current, voltage, vibration, humidity and other multi-dimensional information of the device.

[0015] Further, the training method of the feature fusion-based generative adversarial network algorithm is:

[0016] Initialize the network parameters of the generator and the discriminator;

[0017] In each iteration, the generator receives a random noise vector and constructs a feature vector through a series of convolutional layers;

[0018] In the feature fusion process, the generator uses the features extracted from the real sludge treatment equipment data samples to guide the generation of the features;

[0019] The generated features are sent to an adaptive manifold learning module, which dynamically adjusts the feature space according to the underlying geometry of the sludge treatment equipment data samples;

[0020] In each iteration, the discriminator receives the generated sludge treatment equipment data samples from the generator and the real sludge treatment equipment data samples, and calculates a discrimination result through its network structure;

[0021] The parameters of the generator and the discriminator are updated using the gradient descent method;

[0022] Repeat the above steps until the preset stopping iteration condition is met, indicating that the model training is complete.

[0023] Further, the central device state recognition model aggregates updates from the local device state recognition model, and the global device state recognition model updates are aggregated in a federated learning process to create a new improved global device state recognition model. After the global device state recognition model updates are completed, the final trained device state recognition model is obtained.

[0024] Further, the device state recognition model of each client uses a neural network based on simulated optical optimization, specifically: the initial parameters of the neural network are used as the emission points of the light sources, and each parameter corresponds to a light source;

[0025] The propagation of the light beams emitted by the simulated light sources in the neural network occurs, and upon reaching each neuron, the light beams refract according to the characteristics of the activation function, and the calculation of the refractive intensity improves the sensitivity of the model to the input signal through a nonlinear function;

[0026] In each layer of the neural network, the wavefront state of the light beam is calculated, including intensity, phase, and direction, to simulate the energy distribution of the light beam in that layer;

[0027] According to the wavefront detection results, the emission angle and intensity of the light source in the previous layer are adjusted to optimize the quality of light beam propagation;

[0028] Based on the spectral modulation theory in physical optics, the learning efficiency and stability of the neural network are improved by adjusting the spectral characteristics of the light source corresponding to each parameter in the neural network;

[0029] The deviation between the propagation path of the light beam and the expected output is calculated, and the optical backpropagation algorithm is used to optimize the light beam path to adjust the light beam angle and intensity to minimize the output error;

[0030] The above steps are repeated until a preset stopping iteration condition is met, indicating that the model training is completed.

[0031]

[0032] In the formula, M() is a mapping function; alpha ch is a manifold learning learning rate; Dist() represents a distance function between the generated feature and the manifold of sludge treatment equipment data; U ed is the manifold feature of the sludge treatment equipment data; represents the gradient of F' (l) .

[0033] Further, the wavefront state is calculated in the following manner:

[0034]

[0035] In the formula, Phi p，k is the wavefront state of the kth iteration, is the beam intensity of the ith neuron of the kth iteration, is the phase angle of the ith neuron of the kth iteration.

[0036] Further, the formula of the spectral characteristic adjustment function is as follows:

[0037]

[0038] In the formula, S p，k () is a spectral modulation function; omega p，k represents the spectral frequency of the kth layer parameter; and are the mean and standard deviation of the spectrum, respectively, controlling the center frequency and distribution width of the spectrum.

[0039] Compared with the prior art, the present application has the following advantages:

[0040] 1. In the adaptive control task of the equipment for sludge treatment, a feature fusion-based generative adversarial network is used for data augmentation, a feature fusion mechanism is proposed, and the characteristics of the real sludge treatment equipment data are combined to guide the generator to optimize the data samples, solving the problem of poor model generalization ability caused by insufficient equipment data samples.

[0041] 2. In the adaptive control task of the equipment for sludge treatment, a feature fusion-based generative adversarial network is used for data augmentation, and an adaptive manifold learning module is used to dynamically adjust the generated features according to the data geometric structure, improving the consistency and diversity of the generated data and the real data.

[0042] 3. In the adaptive control task of equipment used for sludge treatment, a neural network based on simulated optical optimization is used for feature extraction. The simulated optical optimization method converts the parameter optimization problem of the neural network into a light propagation process. By adjusting the light beam propagation path and intensity, the limitations of gradient descent in complex problems are avoided, thereby improving the efficiency of model optimization.

[0043] 4. In the adaptive control task of equipment used for sludge treatment, a neural network based on simulated optical optimization is used for feature extraction, and spectral modulation theory is used to refine the control of parameter updates, which improves the neural network's feature extraction ability and classification accuracy for sludge treatment equipment data. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Attachment Figure 1 This is the federal learning architecture diagram of our invention. DETAILED DESCRIPTION

[0045] Below in conjunction with specific embodiment, further set forth the present invention.Should be understood that these embodiments are only used to illustrate the present invention and are not used in limiting the scope of the present invention.In addition, should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the application equally.

[0046] The present invention proposes an adaptive control method for sludge treatment equipment. The method monitors the status of the sludge treatment equipment through an equipment status recognition model and then formulates a corresponding control strategy. The specific details are as follows:

[0047] S1. Acquire data from cleaning equipment used for sludge treatment. These equipment are equipped with multiple sensors that can monitor various environmental and operating status parameters of the equipment in real time. The collected data includes but is not limited to multi-dimensional information such as the equipment's operating temperature, pressure, speed, current, voltage, vibration, and humidity.

[0048] Utilizing a multi-point distributed sensor architecture, sensors are deployed on key components of sludge treatment equipment (such as the transmission system, filter press, and heating device) to ensure that real-time status information of the equipment can be captured from multiple dimensions. Through real-time sampling technology, all sensor data will be collected at fixed time intervals (for example, every second or every millisecond) and stored in the local server.

[0049] The collected data is stored in a timestamp-serialized structured data file (e.g., JSON or CSV format). Each record contains a timestamp and the corresponding reading of each sensor. In one embodiment, the attributes of the data include:

[0050] Ra is the operating temperature of the device (unit: ℃), Db is the current of the device (unit: A), Dc is the voltage of the device (unit: V), Dd is the rotating speed of the device (unit: rpm), De is the pressure of the device (unit: Pa), Df is the humidity of the device (unit: %), Dg is the vibration amplitude of the device (unit: m / s 2 ), Dh is the operating time of the device (unit: hour), Di is the fault indication of the device (unit: Boolean value, 0 is normal, 1 is fault), Dj is the working load of the device (unit: kg).

[0051] It should be noted that the embodiment is only to illustrate one data format and kind of the application, and in actual application, the number of data attributes is usually more than 10, and the number of data attributes can reach dozens or even hundreds.

[0052] Then, the collected data is labeled, and the labeling method of the application is artificial labeling. In one embodiment, the labeled categories include: normal state, slight fault, serious fault, and equipment maintenance, a total of 4 categories.

[0053] S2, since in the task of the application, the collection, labeling and preprocessing of the training data of the sludge treatment equipment are time-consuming and laborious, and the lack of training sludge treatment equipment data samples can easily lead to poor model generalization ability and affect the accuracy of the model, therefore, a feature fusion-based generative adversarial network algorithm is constructed and trained to obtain a data expansion model after training, and then the data expansion of the sludge treatment equipment is realized.

[0054] Specifically, the feature fusion-based generative adversarial network includes two main components: a generator and a discriminator; the generator is responsible for generating realistic sludge treatment equipment data samples, and the task of the discriminator is to distinguish between generated sludge treatment equipment data samples and real sludge treatment equipment data samples. In the traditional generative adversarial network, a feature fusion mechanism is used to guide the processing of generated features by extracting features from real sludge treatment equipment data samples, thereby improving the diversity and quality of generated sludge treatment equipment data samples.

[0055] Specifically, the training process of the feature fusion-based generative adversarial network algorithm is as follows:

[0056] S201, initialize the network parameters of the generator and the discriminator, and set the parameters of the generator as θ G , and the parameters of the discriminator as θ D In one embodiment, the initialization is performed in the form of normal distribution, which is represented as:

[0057]

[0058]

[0059] wherein σ vd 2 denotes the initialized variance, ~ denotes the subject to a certain distribution, denotes a normal distribution. Preferably, σ vd 2 is set to 0.01.

[0060] S202, in each iteration, the generator receives a random noise vector, constructs a feature vector through a series of convolutional layers, denoted as:

[0061]

[0062] wherein F (l) denotes the output feature vector of the l-th layer, z is a noise vector from a normal distribution, is the parameter of the generator at the l-th layer, f (l) is the convolution function of the l-th layer of the generator.

[0063] S203, in the feature fusion process, the generator uses the features extracted from the real sludge treatment equipment data samples to guide the processing of the generated features, denoted as:

[0064]

[0065] wherein, is the feature extracted by the generator from the real sludge treatment equipment data samples; F′ (l) denotes the output fusion feature of the l-th layer of the generator; λ cb is the fusion weight; x denotes the real sludge treatment equipment data samples. Preferably, λ cb denotes the real sludge treatment equipment data samples; is the parameter of the discriminator at the l-th layer.

[0066] S204, the generated features are sent to an adaptive manifold learning module, which dynamically adjusts the feature space according to the underlying geometric structure of the sludge treatment equipment data samples. Specifically, the generated features are adjusted through a mapping function to adapt to the manifold structure of the sludge treatment equipment data, denoted as:

[0067]

[0068] wherein M() is the mapping function; α ch is the manifold learning learning rate; Dist() denotes the distance function between the generated features and the sludge treatment equipment data manifold; U ed is the manifold feature of the sludge treatment equipment data; denotes the gradient of F′ (l) . Preferably, αch is set to 0.01.

[0069] Further, the distance function is used to calculate the distance between the generated feature and the sludge treatment equipment data manifold, and in one embodiment, the calculation method is represented as:

[0070]

[0071] In the formula, Ne is the dimension of the feature, is a nonlinear projection function that projects the feature vector to the manifold U ed ; and ||·|| is the L2 norm.

[0072] S205, the task of the discriminator is to evaluate the difference between the generated sludge treatment equipment data sample and the real sludge treatment equipment data sample, in each iteration, the discriminator receives the generated sludge treatment equipment data sample from the generator and the real sludge treatment equipment data sample, and calculates a discrimination result through its network structure, i.e. whether the sludge treatment equipment data sample is generated or real, and the loss function of the discriminator is calculated as:

[0073]

[0074] In the formula, L D is the loss function of the discriminator; D() represents the output of the discriminator; represents the sludge treatment equipment data generated by the generator.

[0075] And the goal of the generator is to deceive the discriminator as much as possible, so that it misjudges the generated sludge treatment equipment data sample as a real sludge treatment equipment data sample, while the goal of the discriminator is to correctly distinguish the real and generated sludge treatment equipment data sample The loss function of the generator is calculated as:

[0076]

[0077] In the formula, L G is the loss function of the generator; β vk is a coefficient that adjusts the importance of the two parts. Preferably, β vk is set to 0.3.

[0078] S206, the parameters of the generator and the discriminator are updated using the gradient descent method, and the update method is represented as:

[0079]

[0080] In the formula, γ G and γ D are the learning rates of the generator and the discriminator respectively; <- represents the parameter update operation. and respectively represent the parameter gradients of the generator and the discriminator. Preferably, γ G and γ d are both set to 0.03.

[0081] S207, repeat the above steps until a preset stop iteration condition is met, i.e., the model training is completed. In an embodiment, the preset stop iteration condition is that a preset maximum number of iterations is reached, and preferably, the preset maximum number of iterations is set to 1000 times.

[0082] After the sludge treatment equipment data augmentation model is trained, the sludge treatment equipment data augmentation model is used to increase the number of sludge treatment equipment data samples. In an embodiment, the original collected sludge treatment equipment data samples are 800, and the sludge treatment equipment data augmentation model generates 200 samples, so that the augmented sludge treatment equipment data set contains 1000 samples.

[0083] S3, construct a device state recognition model based on a federated learning architecture, the federated learning architecture has a plurality of clients, each client has a local device state recognition model, the local device state recognition model is trained on the sludge treatment equipment data of the respective client, and the local device state recognition model and the center device state recognition model interact.

[0084] Specifically, the federated learning architecture proposed by the present application has a plurality of clients, which are marked as client 0, client 1, client 2, and client 3, etc. These clients represent independent nodes in the federated learning network, each client has its own local sludge treatment equipment data, and each client has a local device state recognition model. These local device state recognition models are trained on the sludge treatment equipment data of the respective client, and the local device state recognition model and the center device state recognition model interact, i.e., the device state recognition model parameter update flows. Fed ' represents the updated device state recognition model parameter after local training, θ fed represents the parameter received from the center device state recognition model.

[0085] Wherein, the center device state recognition model aggregates updates from the local device state recognition model, the global device state recognition model update is a way of federated learning process aggregating individual updates to create a new improved global device state recognition model, and after the global device state recognition model update, the final trained device state recognition model is obtained.

[0086] In an embodiment, the device state recognition model parameter to exchange between the local device state recognition model and the central device state recognition model, which allows the central device state recognition model to aggregate updates and allows the local device state recognition model to receive new, aggregated parameters;

[0087] Further, through multiple iterations, the local device state recognition model is trained, the updates are sent to the central device state recognition model, the global updates are performed, and then the updated parameters are sent back to the local device state recognition model;

[0088] Based on the federated learning architecture, the privacy of the sludge treatment device data can be protected, and the original sludge treatment device data is not shared between the clients or with the central server during the training process, but only the device state recognition model parameters are exchanged for updates, i.e., the central device state recognition model assigns θ fed to each client.

[0089] Further, the device state recognition model of each client uses a neural network based on simulated optical optimization, which simulates the optical optimization process by regarding the parameter optimization problem of the neural network as a light propagation process, simulates the path selection and intensity adjustment of the light beam to optimize the parameter settings. Compared with the traditional gradient descent method, the simulated optical optimization method does not rely on error gradient information, but optimizes through the effect of light beam propagation, avoiding the limitations of gradient information in nonlinear complex problems and deep networks, thereby improving the optimization efficiency and reducing the risk of falling into local optimum, and improving the feature extraction capability of the neural network for sludge treatment device data.

[0090] Specifically, the training process of the neural network algorithm based on simulated optical optimization is as follows:

[0091] S301, the initial parameters of the neural network are used as the emission points of the light source, each parameter corresponds to a light source, and the initial value size of the parameter is represented by the light intensity of the light source. The initial intensity distribution of the light source is represented as:

[0092]

[0093] In the formula, I p，0 (w p，i , b p，j ) represents the initial intensity of the light source, w p，i and b p，j are the initial values of the weights and biases of the neural network, respectively, and σ p is a hyperparameter that controls the intensity distribution of the light beam. Preferably, σ p is set to 0.5 to ensure a moderate intensity distribution.

[0094] S302, the propagation of the light beam emitted by the light source in the neural network, and the light beam is refracted according to the characteristics of the activation function at each neuron, and the calculation of the refractive intensity improves the sensitivity of the model to the input signal through a nonlinear function, which is represented as:

[0095]

[0096] wherein R p，k represents the light beam intensity after the kth iteration of refraction, I p，k is the light intensity of the light source at the kth iteration, a p is a hyperparameter that adjusts the refractive index, and i w p，i ·x p，i +b p，j is a linear combination of the input signal and the weight; x p，i is the i-th feature input into the neural network, which can be Ra is the operating temperature of the device (in ℃), Db is the current of the device (in A), Dc is the voltage of the device (in V), Dd is the rotating speed of the device (in rpm), De is the pressure of the device (in Pa), Df is the humidity of the device (in %), Dg is the vibration amplitude of the device (in m / s 2 ), Dh is the running time of the device (in hours), Di is the fault indication of the device (in Boolean value, 0 for normal and 1 for fault), and Dj is the working load of the device (in kg). Preferably, a p is set to 2.

[0097] S303, at each layer of the neural network, the wavefront state of the light beam is calculated, including intensity, phase and direction, and the energy distribution of the light beam in the layer is simulated, and the calculation method of the wavefront state is represented as:

[0098]

[0099] wherein Φ p，k is the wavefront state at the kth iteration, is the light beam intensity of the i-th neuron at the kth iteration, is the phase angle of the i-th neuron at the kth iteration.

[0100] S304, according to the wavefront detection result, the emission angle and intensity of the light source of the previous layer are adjusted to optimize the light beam propagation quality, which characterizes the network performance, and the calculation method of the feedback control function is represented as:

[0101]

[0102] wherein G p，k () represents the feedback control function at the kth iteration, and γ pis the learning rate of the feedback control function; λ p is the adjustment term for enhancing the coherence of adjacent neuron beam angles; is the wavefront state of the i-th neuron in the k-th iteration; is the phase angle of the i-1-th neuron in the k-th iteration. Preferably, γ p is set to 0.01, λ [ is set to 0.2.

[0103] S305, based on the spectral modulation theory in physical optics, the learning efficiency and stability of the neural network are improved by adjusting the spectral characteristics of the light source corresponding to each parameter in the neural network. The spectral modulation strategy can more finely control the update speed and direction of each parameter. Specifically, it is realized by changing the spectral distribution of the light source. The spectral distribution determines the intensity and phase of the light beam when it reaches different layers of the network, thereby affecting the specific way of parameter update. The calculation method of the spectral modulation function is represented as:

[0104]

[0105] In the formula, S p，k () is the spectral modulation function; ω p，k represents the spectral frequency of the k-th layer parameter; and are the mean and standard deviation of the spectrum, respectively, controlling the center frequency and distribution width of the spectrum. Preferably, is set as an adjustable hyperparameter to dynamically adjust according to the training progress, maintains a constant to maintain the consistency of the modulation.

[0106] S306, the deviation between the light beam propagation path and the expected output is calculated, the light beam path is optimized using the optical backpropagation algorithm, and the light beam angle and intensity are adjusted to minimize the output error. The parameter update method of the neural network is represented as:

[0107]

[0108] In the formula, w p，k is the weight of the k-th iteration of the neural network; b p，k is the bias of the k-th iteration of the neural network; w p，k+1 is the weight of the k+1-th iteration of the neural network; b p，k+1 is the bias of the k+1-th iteration of the neural network; w p，k-1 is the weight of the k-1-th iteration of the neural network; b p，k-1 is the bias of the k-1-th iteration of the neural network; η p is the update step size; μ p is the momentum factor, which uses historical gradient information through the momentum term to enhance the stability of optimization; G is a gradient with respect to the bias of the neural network; G p，k is the value of the feedback control function at the kth iteration; G is a gradient with respect to the weight of the neural network. Preferably, η p is set to 0.01, μ p is set to 0.9.

[0109] S307, repeat the above steps until the preset stopping iteration condition is met, that is, the model training is completed. In an embodiment, the preset stopping iteration condition is that the preset maximum iteration number is reached, and preferably, the preset maximum iteration number is set to 1000 times.

[0110] S4, monitoring the state of the sludge treatment equipment by the equipment state recognition model, and then making a corresponding control strategy. Specifically, after the neural network training is completed, the trained neural network is used to extract features from the sludge treatment equipment data, and further, the preset Softmax function is used to classify the sludge treatment equipment data. The classification categories include: normal state, slight fault, serious fault, and equipment maintenance, a total of 4 categories. According to the state of the sludge treatment equipment, a corresponding control strategy is made. For example, when a slight fault is prompted, it is necessary to troubleshoot in time; when a serious fault is prompted, it is necessary to stop and process.

[0111] Finally, it should be noted that: the above is only the preferred embodiment of the present application, and is not used to limit the present application, although the present application has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. An adaptive control method for sludge treatment equipment, characterized by: acquiring data of a cleaning equipment for sludge treatment, the data including various environmental and operating state parameters that can be monitored in real time during the operation of the equipment; constructing a generative adversarial network algorithm based on feature fusion and training to obtain a data augmentation model after training; constructing an equipment state recognition model based on a federated learning architecture, the federated learning architecture having multiple clients, each client having a local equipment state recognition model, the local equipment state recognition model being trained on the sludge treatment equipment data of the respective client, and the local equipment state recognition model and the central equipment state recognition model interacting; monitoring the state of the sludge treatment equipment by the equipment state recognition model and making a corresponding control strategy accordingly; the equipment state recognition model of each client uses a neural network based on simulated optical optimization, specifically: the initial parameters of the neural network are taken as the emission points of light sources, and each parameter corresponds to a light source; the propagation of the light beams emitted by the simulated light sources in the neural network, and upon reaching each neuron, the light beams refract according to the characteristics of the activation function, and the calculation of the refractive intensity improves the sensitivity of the model to the input signal through a nonlinear function; at each layer of the neural network, the wavefront state of the light beam is calculated, including intensity, phase and direction, and the energy distribution of the simulated light beam in the layer is simulated; based on the wavefront detection results, the emission angle and intensity of the light source in the previous layer are adjusted to optimize the light beam propagation quality; based on the spectral modulation theory in physical optics, the learning efficiency and stability of the neural network are improved by adjusting the spectral characteristics of the light source corresponding to each parameter in the neural network; the deviation between the light beam propagation path and the expected output is calculated, the light beam path is optimized using an optical backpropagation algorithm, and the light beam angle and intensity are adjusted to minimize the output error; repeat the above steps until the preset stopping iteration condition is met, indicating that the model training is complete.

2. The adaptive control method for a device for sludge treatment according to claim 1, characterized in that: The collected data includes but is not limited to the running temperature, pressure, speed, current, voltage, vibration, and humidity of the equipment.

3. The adaptive control method for a device for sludge treatment according to claim 1, characterized in that, The training method of the generative adversarial network algorithm based on feature fusion is: initialize the network parameters of the generator and the discriminator; in each iteration, the generator receives a random noise vector and constructs a feature vector through a series of convolutional layers; in the feature fusion process, the generator uses the features extracted from the real sludge treatment equipment data samples to guide the generation of features; the generated features are sent to an adaptive manifold learning module, which dynamically adjusts the feature space according to the underlying geometric structure of the sludge treatment equipment data samples; in each iteration, the discriminator receives the generated sludge treatment equipment data samples from the generator and the real sludge treatment equipment data samples, and calculates a discrimination result through its network structure; update the parameters of the generator and the discriminator using gradient descent method; repeat the above steps until the preset stopping iteration condition is met, indicating that the model training is complete.

4. The adaptive control method for a device for sludge treatment according to claim 1, characterized in that, The central device state recognition model aggregates updates from the local device state recognition models, the global device state recognition model update is the way of federated learning process to aggregate individual updates to create a new improved global device state recognition model, and after the global device state recognition model update is completed, the final trained device state recognition model is obtained.

5. The adaptive control method for a device for sludge treatment according to claim 3, characterized in that: The generated features are adjusted by a mapping function to adapt to the manifold structure of the sludge treatment device data, and are represented as: ； wherein is a mapping function; is a manifold learning learning rate; denotes a distance function between the generated feature and the sludge treatment plant data manifold; is a sludge treatment plant data manifold feature; denotes a gradient of , denotes an output fusion feature of the generator's layer.

6. The adaptive control method for a device for sludge treatment according to claim 1, characterized in that, The calculation method of the wave front state is: ； Where, For the The wavefront state of the iteration, It is The first iteration The beam intensity of each neuron, It is The first iteration The phase angle of a neuron.

7. The device adaptive control method for sludge treatment according to claim 1, characterized in that: The formula of the spectral characteristic adjustment function is: ； where is the spectral modulation function; denotes the spectral frequency of the layer parameters; and are the mean and standard deviation of the spectrum, respectively, controlling the center frequency and the distribution width of the spectrum.

Images