Fountain solution recovery efficiency evaluation system and method based on image recognition

Abstract

The invention relates to the technical field of image recognition, in particular to a fountain solution recovery efficiency evaluation system and method based on image recognition, which adopts a generative adversarial network based on nonlinear feature deentanglement and allows input features to be mapped to a high-dimensional quantum feature space. The characterization capability of data is improved through superposition and entanglement effects of quantum states, so that the diversity and quality of a data set are effectively improved, a three-layer full-connection neural network of an adaptive wolf pack optimization algorithm is adopted to perform feature extraction, and the social behavior of the wolf pack is simulated to optimize neural network parameters, so that the robustness of the algorithm is improved. The self-encoding neural network based on mask enhancement is adopted to perform feature dimension reduction, and random masks are applied to selectively hide partial data features in an input layer, so that the network learns from partial visible data and reconstructs complete data, and the search efficiency and the solution accuracy of the algorithm in a high-dimensional data space are improved. And the learning ability of the model to the data implicit relationship is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and particularly to a dampening solution recovery efficiency evaluation system and method based on image recognition. Background Art

[0002] In traditional dampening solution recovery systems, although there are certain monitoring and evaluation methods, these methods are often limited by small data volume, insufficient data diversity, and limitations of processing technologies, making it difficult to achieve efficient and accurate evaluation of dampening solution quality. In addition, since the quality of the dampening solution directly affects printing quality and cost, effective recovery and reuse of it become particularly important;

[0003] Chinese invention patent with application number CN202410135161.5 proposes an elastic metal plastic bearing lubrication state recognition sensor, monitoring method and application. The lubrication state recognition sensor is installed in the area with the highest theoretical dynamic pressure lubricating oil film temperature of the elastic metal plastic bearing, and the monitoring end of the probe is flush with the bearing surface; the monitoring method includes: the probe monitors the resistance value and resistance change value of the fluid medium between the bearing and the mating pair in real time; obtaining the lubrication state monitoring according to the resistance change value, and then obtaining the resistance value data of fluid lubrication, mixed lubrication and solid lubrication respectively; mixed lubrication refers to the lubrication state in which fluid lubrication and solid lubrication alternate with each other; according to the Stribeck curve theory, simulation tests are carried out to identify the verification of the fluid lubrication, mixed lubrication and solid lubrication states of the elastic metal plastic bearing, and the red, yellow and green safe operation standards under different load and speed conditions are established to realize the early warning of fault prediction.

[0004] The above technical solution is innovative, but at the same time, the following problems still need to be further solved:

[0005] 1. In the prior art, data augmentation methods usually cannot fully utilize the high-dimensional feature space, resulting in insufficient diversity and quality of the generated data, and unable to effectively support the training and generalization of the model.

[0006] 2. In the prior art, feature extraction methods are often inefficient when dealing with high-dimensional data, and are easily limited by local optimal solutions, resulting in poor model performance and stability.

[0007] 3. In the prior art, feature dimensionality reduction techniques fail to effectively utilize the internal structure of the data, and key information may be lost during the dimensionality reduction process, affecting the decision-making ability and accuracy of the subsequent model.

[0008] 4. In the prior art, classifiers fail to adapt to the distribution characteristics of data on a specific geometric structure, resulting in low information transmission efficiency and insufficient accuracy during the classification process.

[0009] To solve these problems, this application designs a dampening solution recovery efficiency evaluation system and method based on image recognition. Summary of the Invention

[0010] In order to overcome the defects and deficiencies existing in the prior art, the present invention provides a dampening solution recovery efficiency evaluation system and method based on image recognition.

[0011] In order to achieve the above object, the present invention adopts the following technical solutions:

[0012] In a first aspect, the present invention provides a dampening solution recovery efficiency evaluation method based on image recognition, including the following steps:

[0013] Step S1: Collect the data collected by the equipment for using dampening solution, the recovery equipment and the relevant monitoring sensors, and perform data annotation;

[0014] Step S2: Generate samples based on the generative adversarial network algorithm for non-linear feature disentanglement, so as to perform data augmentation;

[0015] Step S3: Input the augmented data into the feature extraction model for training the feature extraction model;

[0016] Step S4: Input the data after feature extraction into the feature dimensionality reduction model for training the feature dimensionality reduction model;

[0017] Step S5: Input the data after dimensionality reduction into the classifier for training the classifier model;

[0018] Step S6: Use the trained feature extraction model, data dimensionality reduction model and classifier model to process new samples and generate the final output result.

[0019] In an implementation manner of the present invention, the training process of the generative adversarial network algorithm for non-linear feature disentanglement is as follows:

[0020] S201: Initialize the model parameters of the generator G c and the discriminator D c ;

[0021] S202: Map the original training data to a high-dimensional Hilbert space through the quantum feature mapping method;

[0022] S203: The generator G c receives the mapped high-dimensional features and generates new data samples through its neural network structure;

[0023] S204: The discriminator D c classifies the generated data and the real data;

[0024] S205. Adjust the parameters of the generator according to the feedback of the discriminator, and constrain the parameter update of the discriminator through an adaptive loss function;

[0025] S206. Repeat the above steps iteratively until the preset iteration stop condition is met, which indicates that the model training is completed.

[0026] In an implementation manner of the present invention, the step S3 includes the following specific steps:

[0027] S301. Initialize the parameters of the neural network, where the parameters of the neural network include weights and biases;

[0028] S302. Adjust the parameters of the neural network by simulating the social behavior of the wolf pack based on the adaptive wolf pack optimization algorithm;

[0029] S303. In each algorithm iteration, adjust the search strategy according to the optimization results of the previous generation;

[0030] S304. Update the weights and biases of the neural network;

[0031] S305. Repeat the above steps iteratively until the preset iteration stop condition is met, which indicates that the model training is completed.

[0032] In an implementation manner of the present invention, the step S4 includes the following specific steps:

[0033] S401. Initialize the parameters of the autoencoder network, including the weight and bias parameters of the autoencoder network;

[0034] S402. According to the framework of the greedy algorithm, set the independent training cycles and learning rates for each layer;

[0035] S403. At the beginning of each training cycle, for each input vector X th , randomly generate a mask matrix m for randomly masking part of the input data;

[0036] S404. After the input data is masked, it is sent to the encoder, and the encoder processes the data through an activation function and compresses it into a low-dimensional feature representation;

[0037] S405. The low-dimensional features are sent to the decoder, and the decoder attempts to reconstruct the original input data. By calculating the loss function, the backpropagation algorithm updates the parameters in the network;

[0038] S406. Use the greedy algorithm to optimize the parameters of each layer;

[0039] S407. When the training of all layers is completed, integrate the models of each layer to form a complete autoencoder model;

[0040] S408. Repeat the above steps iteratively until the preset iteration stop condition is met, indicating that the model training is completed.

[0041] In one implementation of the present invention, in step S5, inputting the dimension-reduced data into a classifier for training the classifier model includes the following specific contents:

[0042] S501. Initialize the parameters of the Riemannian neural network.

[0043] After the dimension-reduced data is input, first, a preprocessing layer performs a preliminary linear transformation on the data, and an adaptive feature recalibration strategy is adopted to dynamically adjust its transmission weight in the network according to the activation response of each feature.

[0044] S503. At each hidden layer node, calculate the inner product defined by the Riemannian metric of the input features as the weighted input before neuron activation to adapt to the Riemannian manifold structure of the data.

[0045] S504. Use an activation function to transform the weighted input to generate an activation output, which will be sent to the next layer or the output layer.

[0046] S505. At the end of the network, complete class determination through an output layer. The output layer uses a maximum function to calculate the classification probability and optimizes the model through a loss function.

[0047] S506. Repeat the above steps iteratively until the preset iteration stop condition is met, indicating that the model training is completed.

[0048] In one implementation of the present invention, step S6 includes the following specific steps:

[0049] Use the trained feature extraction model, data dimension reduction model, and classifier model to process new samples and generate a final output result to achieve the evaluation of the dampening solution recovery efficiency.

[0050] In one implementation of the present invention, the parameter update method of the discriminator in S205 is a backpropagation method based on gradient descent, and an adaptive loss function combines the quantum fidelity between the generated samples and the real samples.

[0051] In a second aspect, the present invention also provides a dampening solution recovery efficiency evaluation system based on image recognition, including:

[0052] A data collection and annotation module for collecting data of the equipment for using the dampening solution, the recovery equipment, and related monitoring sensors.

[0053] A feature extraction module that inputs the augmented data into a feature extraction model for training the feature extraction model.

[0054] A feature dimensionality reduction module that inputs the data after feature extraction into a feature dimensionality reduction model for training the feature dimensionality reduction model;

[0055] A classification module that inputs the dimensionally reduced data into a classifier for training the classifier model.

[0056] In a third aspect, an electronic device provided by the present invention includes: a processor and a memory, wherein a computer program callable by the processor is stored in the memory, and the processor executes a method for evaluating the recovery efficiency of dampening solution based on image recognition by calling the computer program stored in the memory.

[0057] In a fourth aspect, a computer-readable storage medium provided by the present invention stores instructions that, when run on a computer, cause the computer to execute a method for evaluating the recovery efficiency of dampening solution based on image recognition.

[0058] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0059] 1. The present invention adopts a generative adversarial network based on non-linear feature disentanglement, which allows input features to be mapped to a high-dimensional quantum feature space, and improves the data representation ability through the superposition and entanglement effects of quantum states, thereby effectively increasing the diversity and quality of the data set.

[0060] 2. The present invention uses a three-layer fully connected neural network with an adaptive wolf pack optimization algorithm for feature extraction, simulating the social behavior of wolf packs to optimize the neural network parameters, and improving the search efficiency and solution accuracy of the algorithm in the high-dimensional data space.

[0061] 3. The present invention uses a self-encoding neural network based on mask enhancement for feature dimensionality reduction, applying a random mask in the input layer to selectively hide some data features, enabling the network to learn from partially visible data and reconstruct the complete data, and enhancing the model's learning ability for the implicit relationship of the data.

[0062] 4. The present invention uses a Riemann neural network based on the response metric of activation functions, using the Riemann metric to optimize the connection weights between nodes, and improving the data transfer efficiency and classification accuracy in the network according to the distribution characteristics of the data on the Riemann manifold. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] By reading the detailed description of the non-limiting embodiments with reference to the following drawings, other features, objects, and advantages of the present invention will become more apparent:

[0064] Figure 1 It is a schematic diagram of the training process of the neural network algorithm based on the adaptive wolf pack optimization algorithm for the method of the present invention;

[0065] Figure 2 This is the flowchart of the training process of the method of the present invention based on the masked autoencoder neural network algorithm. Specific embodiments

[0066] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. Without conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0067] Embodiment 1

[0068] As Figure 1 and Figure 2 shown, this embodiment provides a method for evaluating the recovery efficiency of dampening solution based on image recognition, which specifically includes the following steps:

[0069] S1. Data collection and annotation. The data of the present invention is collected from the dampening solution using equipment, recovery equipment, and related monitoring sensors. The data is stored in JSON format. In one embodiment, the specific data attributes include:

[0070] R a1 is the temperature of the dampening solution (degrees), R a2 is the pH value of the dampening solution, R a3 is the recovery rate (liters per hour), R a4 is the color parameter of the dampening solution, R a5 is the turbidity of the recovered liquid, R a6 is the running time of the using equipment (hours), R a7 is the number of equipment maintenance times, R a8 is the viscosity of the dampening solution (mPa·s), R a9 is the humidity of the equipment environment (%), R a10 is the temperature of the equipment environment (degrees).

[0071] It should be noted that this embodiment is only to illustrate a data format and type of the present invention. In actual applications, the data attributes are usually more than 10 attributes, and the number of data attributes may reach dozens or even hundreds.

[0072] Furthermore, the collected data is annotated. The annotation method of the present invention is manual annotation. In one embodiment, the annotation categories include 4 levels: "excellent", "good", "medium", and "poor".

[0073] S2. Data expansion. It is understandable that in the task of the present invention, the acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model and affect the accuracy of the model. The present invention adopts a generative adversarial network algorithm based on nonlinear feature disentanglement for sample generation, and uses quantum feature mapping to improve the quality and diversity of data expansion. The generative adversarial network based on nonlinear feature disentanglement includes two parts: the generator G c and the discriminator D c The generator is responsible for generating data that looks real, while the discriminator tries to distinguish the generated data from the real data. In order to enhance the performance of the generative adversarial network and solve the problem of feature entanglement, the present invention c Quantum circuits are embedded in the system, which can map input features into a high-dimensional quantum feature space and use the superposition and entanglement effects of quantum states to increase the data representation capability.

[0074] Specifically, the training process of the generative adversarial network algorithm based on nonlinear feature disentanglement is as follows:

[0075] S201, initialize generator G c and the discriminator D c Model parameters, specifically, the initialization parameters include the generator G c Weight Discriminator D c Weight Generator G c Bias Discriminator D c Bias It is expressed as:

[0076]

[0077] In the formula, is a normal distribution, is the variance of weights and biases at initialization. Preferably, Set to 0.01.

[0078] S202, map the original training data to the high-dimensional Hilbert space by quantum feature mapping. Specifically, for the input data X c The way to perform quantum feature mapping is expressed as:

[0079]

[0080] Where, X c is the original data vector, is the data vector after quantum mapping, φ c () is the quantum characteristic mapping function, θ cis the mapping parameter, θ c is the quantum Hamiltonian operator. Preferably, the mapping parameter θ c is set to π / 8.

[0081] In one embodiment, the calculation of the quantum Hamiltonian operator H c is expressed as:

[0082]

[0083] wherein, σ z,j and σ x,j+1 respectively represent the Pauli Z and X matrices acting on the j-th and (j + 1)-th qubits, and N c represents the number of qubits.

[0084] And, the calculation of the mapping parameter θ c is expressed as:

[0085]

[0086] wherein, f c represents the mapping function from the data X c to the quantum state |ψ c >, x c,i is the i-th component of the input data vector X c and d c is the data dimension. Preferably, the mapping function performs projection mapping in the way of principal component analysis.

[0087] S203. The generator G c receives the mapped high-dimensional features and generates new data samples through its neural network structure. Specifically, the generator G c generates new samples using the mapped data in the following way:

[0088]

[0089] wherein, G c () is the generator function, is the generated new sample, is the tanh activation function adjusted based on slope and amplitude.

[0090] In one embodiment, the calculation of the tanh activation function adjusted based on slope and amplitude is expressed as:

[0091]

[0092] wherein, α c and β care learnable parameters used to adjust the slope and amplitude of the activation function. Preferably, the learnable parameter α c and β c are learned in the way of error backpropagation.

[0093] Meanwhile, the generator tries to maximize the similarity with the real data and optimize it using an adaptive loss function to ensure the quality of the generated data.

[0094] S204. Discriminator D c classifies the generated data and the real data, and tries to distinguish which are the generated data and which are the real data. Specifically, the discriminator D c evaluates the authenticity of the generated samples and the real samples as follows:

[0095]

[0096] In the formula, D c () is the discriminator function, and Sig() is an adjustable Sigmoid activation function.

[0097] Furthermore, the adjustable Sigmoid function is calculated as follows:

[0098]

[0099] In the formula, γ c is an adjustable parameter, and zs is the input of the adjustable Sigmoid function. Preferably, γ c is set to 0.01.

[0100] Meanwhile, the training objective of the discriminator is to minimize the classification error and optimize the adaptive loss function at the same time.

[0101] S205. Adjust the parameters of the generator according to the feedback of the discriminator, constrain the parameter update of the discriminator through the adaptive loss function. The parameter update method of the discriminator is the backpropagation method based on gradient descent. Moreover, the adaptive loss function combines the quantum fidelity between the generated samples and the real samples, and the calculation method is as follows:

[0102]

[0103] In the formula, is the expected symbol, L(G c , D c ) is the adaptive loss function, λ c is the weight of the quantum fidelity term, F() is the fidelity function between quantum states, is a regularization term based on the generator weights, Tr() is the trace operation of a matrix, and ρ c and σ c are the density matrices of the generated and real data respectively. Preferably, λ c is set to 3.

[0104] In one embodiment, the regularization term based on the generator weights dynamically adjusts its strength according to the performance of the discriminator D c in the current training stage, and the calculation method is expressed as:

[0105]

[0106] In the formula, represents the L2 norm of the generator weights, ρ c represents the accuracy of the discriminator D c in classifying real samples in the most recent training batch, κ c is a positive scaling factor, and η c is the baseline strength of the regularization term. Preferably, κ c is set to 10 and η c is set to 0.01.

[0107] S206. Repeat the above steps iteratively until the preset stop iteration condition is satisfied, which indicates that the model training is completed. In one embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0108] S3. Feature extraction model training. Input the augmented data into the feature extraction model for training the feature extraction model. The present invention uses a 3-layer fully connected neural network for feature extraction. In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, problems such as gradient disappearance, gradient explosion, or getting stuck in local optimal solutions may be encountered, affecting the stability of training and the performance of the model. The present invention uses a neural network based on the adaptive wolf pack optimization algorithm. Based on the traditional wolf pack optimization algorithm, an adaptive mechanism is adopted to adjust the search strategy according to the optimization results in the early stage, improving the search efficiency and accuracy of the algorithm in the high-dimensional data space.

[0109] Specifically, the training process of the neural network algorithm based on the adaptive wolf pack optimization algorithm is as Figure 1 shown:

[0110] S301. Initialize the parameters of the neural network. The parameters of the neural network include weights and biases. In one embodiment, the initialization method is expressed as:

[0111]

[0112] In the formula, represents the initial value of the neural network parameters; σ cs is the initialized standard deviation; G t (0, I) represents a Gaussian distribution with a mean of 0 and a covariance of the identity matrix. Preferably, σ cs is set to 0.01.

[0113] S302. Based on the adaptive wolf pack optimization algorithm, adjust the parameters of the neural network by simulating the social behavior of the wolf pack. Each wolf represents a set of potential solutions of the neural network, and the social behavior of the wolf pack helps to explore the solution space and find the optimal solution. The way to update the position of the wolf pack is expressed as:

[0114]

[0115] In the formula, and are the positions of the wolves at the t-th and (t + 1)-th iterations respectively; x cp is the individual best position in the wolf pack; x cg is the global best position; α c and β c are parameters for balancing exploration and exploitation. Preferably, α c and β c are set to 0.3 and 0.5 respectively.

[0116] S303. In each iteration of the algorithm, adjust the search strategy according to the optimization results of the previous generation. The way of adaptive adjustment is expressed as:

[0117]

[0118] In the formula, is the updated α c ; λ c is the adjustment factor used to adjust the magnitude of α c ; ∥∥ represents the Euclidean distance.

[0119] In one embodiment, the adjustment factor λ c is adjusted based on the feedback of the historical optimization performance, and the calculation method is expressed as:

[0120]

[0121] In the formula, T ve is the number of historical iterations considered, indicating how many generations of performance in the past affect the current adjustment; is the decay factor, γ c is the decay rate; is the value of the loss function at the t-th iteration; is the neural network parameter for the t-th iteration, that is, the individual best position x in the wolf pack for the t-th iteration cp corresponding neural network parameter. Preferably, γ c is set to 3.

[0122] S304. Update the weights and biases of the neural network, and the update method is expressed as:

[0123]

[0124] In the formula, and are the neural network parameters for the t-th iteration and the (t + 1)-th iteration respectively; η c is the learning rate; is the gradient of the loss function at the parameter . Preferably, η c is set to 0.01.

[0125] Furthermore, the calculation method of the gradient is expressed as:

[0126]

[0127] In the formula, M is the number of samples in each batch of training data; x cm is the input data of the m-th sample; y cm is the true label of the m-th sample; is the cross-entropy loss function; f(x cm ; θ c ) is the predicted output of the neural network model.

[0128] 305. Repeat the above steps iteratively until the preset iteration stop condition is met, which indicates that the model training is completed. In one embodiment, the preset iteration stop condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0129] S4. Feature dimensionality reduction model training. Input the data after feature extraction into the feature dimensionality reduction model for training the feature dimensionality reduction model. The present invention uses a mask-enhanced autoencoder neural network as the feature dimensionality reduction algorithm. The mask-enhanced autoencoder neural network includes an encoder and a decoder, and realizes data compression and reconstruction by learning the effective representation of the input data. Based on the traditional autoencoder, the present invention selectively hides some features of the input data with a random mask at the input layer, forcing the network to reconstruct the complete output from the partially visible input data, thereby enhancing the model's learning ability for the implicit relationships in the data to ensure that the data after dimensionality reduction can still maintain the original data structure characteristics.

[0130] Specifically, the training process of the self - encoding neural network algorithm based on mask enhancement is as follows Figure 2 shown;

[0131] S401. Initialize the parameters of the auto - encoder network, including the weights and bias parameters of the auto - encoder network. Let Θ (l) be the parameter set of the l - th layer, including weights and biases, that is, Θ (l) ={W tu (l) , b tu (l)}, and W tu (l) is the weight matrix of the l - th layer, b tu (l) is the bias vector of the l - th layer. In one embodiment, the initialization method is random initialization.

[0132] S402. According to the framework of the greedy algorithm, set the independent training cycles and learning rates for each layer to ensure that each layer can reach the local optimum before being added to the overall network. The number of independent training loop iterations T l for the l - th layer is set as:

[0133]

[0134] where Ep is the total number of iterations, Ler is the total number of layers, and δ l,Ler is the Kronecker function to ensure that the last layer can have more training cycles.

[0135] S403. At the beginning of each training cycle, for each input vector X th , randomly generate a mask matrix m to randomly mask part of the input data. The way to generate the mask is expressed as:

[0136] m j ~Ber(p)

[0137] where m j is the j - th element of the mask vector m, and p is the mask ratio; Ber() is the Bernoulli distribution discrete function. The Bernoulli distribution discrete function is used to generate the mask matrix, and each element m j is independently drawn from the Ber(p) distribution, indicating that this element has a probability of p to be set to 1 (i.e., selected or activated) and a probability of 1 - p to be set to 0 (i.e., covered or not activated).

[0138] Furthermore, the calculation method of the mask ratio p is expressed as:

[0139]

[0140] where Var(X TH) represents the input data X TH is the variance of θ tu is a preset variance threshold, and min() represents the minimum value function.

[0141] S404. After the input data is masked, it is sent to the encoder. After being processed by the activation function, the encoder compresses the data into a low-dimensional feature representation. The forward propagation method of the encoder is expressed as:

[0142] z tu = Sig(W tu (enc) ⊙X th + b tu (enc) )

[0143] In the formula, z tu is the encoded feature vector, Sig() is the Sigmoid activation function, W tu (enc) and b tu (enc) are the weights and biases of the encoder, and ⊙ represents the Hadamard product.

[0144] S405. The low-dimensional feature is sent to the decoder. The decoder attempts to reconstruct the original input data. By calculating the loss function, the backpropagation algorithm updates the parameters in the network. The output of the decoder is expressed as:

[0145] X th ^ = Sig(W tu (dec) ·z tu + b tu (dec) )

[0146] In the formula, X th ^ is the output of the decoder, that is, the reconstructed input data.

[0147] Furthermore, the calculation method of the loss function is expressed as:

[0148]

[0149] In the formula, L au is the total loss, is the reconstruction error, λ ru is the weight controlling the isolation loss; ∥z tu ∥0 is the isolation loss, representing the number of non-zero elements of z tu used to measure the sparsity of the encoding. Preferably, λ ru is set to 0.3.

[0150] Furthermore, the isolation loss ∥z tuThe calculation method of ∥0 is expressed as:

[0151]

[0152] In the formula, n z is the dimension of the vector z tu ; is the encoded feature vector of the i-th dimension; 1() is the indicator function, which takes the value of 1 when and 0 otherwise.

[0153] S406. Optimize the parameters of each layer using the greedy algorithm. In one embodiment, the parameter update method uses the Adam optimizer, and the update method is expressed as:

[0154]

[0155] In the formula, and respectively represent the parameters of the l-th layer at the t-th iteration and the t+1-th iteration. Adam() is the Adam optimizer function, and η vu is the learning rate, and L au is the calculated loss, is the gradient of the parameters of the l-th layer at the t-th iteration with respect to the loss function. Preferably, η vu is set using a dynamically adjusted strategy, and its initial value is set to 0.01.

[0156] In one embodiment, let the learning rate at the t-th iteration be and its dynamic adjustment method is expressed as:

[0157]

[0158] In the formula, is the initial learning rate, is the initial reconstruction weight, is the reconstruction weight at the t-th iteration, and δ vu is the adjustment influence coefficient. Preferably, δ vu is set to 3, is set to 5.

[0159] Furthermore, the adjustment of the reconstruction weight at the t-th iteration is based on the current effect of the reconstruction error and the sparsity weight, and the calculation method is expressed as:

[0160]

[0161] In the formula, γ yc is the adjustment rate parameter, is the current reconstruction error, β yc is the sparsity weight, Xth is the input data, is the reconstruction output of the t-th iteration. Preferably, γ yc is set to 0.3, and β yc is set to 5.

[0162] Furthermore, the adjustment method of the gradient in the Adam optimizer is expressed as:

[0163]

[0164] In the formula, m^ t and v^ t are the first and second moment estimates after bias correction of the t-th iteration respectively, and ∈ is a small constant. Preferably, ∈ is set to 0.0001.

[0165] Furthermore, the first and second moment estimates m∧ t after bias correction of the t-th iteration and v^ t are calculated as follows:

[0166]

[0167] In the formula, m t and v t are the exponential moving average of the gradient and the exponential moving average of the square of the t-th iteration respectively, and are the attenuation rates of the t-th iteration. Preferably, and are set to 0.9 and 0.999 respectively.

[0168] S407. When the training of all layers is completed, integrate the models of each layer to form a complete autoencoder model and perform overall fine-tuning to optimize the global performance of the model. The fine-tuning method is expressed as:

[0169]

[0170] In the formula, Θ fine-tune is the fine-tuned parameter, N tu is the number of training samples, is the i-th training sample, represents the loss function with parameters and Θ.

[0171] S408. Repeat the above steps iteratively until the preset iteration stop condition is met, which indicates that the model training is completed. In one embodiment, the preset iteration stop condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0172] S5. Training of the classifier model: Input the data after dimensionality reduction into the classifier for training the classifier model. In the present invention, a Riemannian neural network based on activation function response metric is used as the classification algorithm, and the distance metric in the Riemannian space is utilized to calculate the connection weights between nodes, making it more adaptable to the distribution characteristics of the data on the Riemannian manifold, optimizing the data transmission efficiency in the network, and improving the classification accuracy.

[0173] Specifically, the training process of the Riemannian neural network algorithm based on activation function response metric is as follows:

[0174] S501. Initialize the parameters of the Riemannian neural network, including the weight W u and the bias b u , and the initialization method is expressed as:

[0175]

[0176] In the formula, and respectively represent the initialized weight and bias of the Riemannian neural network, is the initialized variance. Preferably, is set to 0.01.

[0177] S502. After the data after dimensionality reduction is input, first perform a preliminary linear transformation on the data through a preprocessing layer, and adopt an adaptive feature recalibration strategy to dynamically adjust the transmission weight of each feature in the network according to the activation response of the feature. The calculation method is expressed as:

[0178]

[0179] In the formula, x u,i is the input feature after dimensionality reduction, is the feature after adaptive feature recalibration, s u,i is the scaling factor dynamically adjusted according to the feature importance, α u and β u are respectively the parameters for scaling adjustment, δ u,i is the importance measure of the i-th feature. Preferably, α u is set to 2, and β u is set to 0.5.

[0180] In one embodiment, the importance measure δ u,i of the i-th feature characterizes the sensitivity of the feature x u,i to the loss function L, and the calculation method is expressed as:

[0181]

[0182] Wherein, L is the loss function of the Riemann neural network.

[0183] S503. At each hidden layer node, calculate the inner product defined by the Riemann metric for the input features, and use this as the weighted input before neuron activation to adapt to the Riemannian manifold structure of the data. The calculation method is expressed as:

[0184]

[0185] Wherein, w u,ji is the weight, G u,ij is an element in the Riemann metric matrix, used to adjust the contribution of the i-th feature to the j-th neuron, and z u,j is the weighted input.

[0186] Furthermore, the element G u,ij in the Riemann metric matrix is a metric matrix defined on the Riemannian manifold according to the geometric structure of the data, and is dynamically adjusted according to the local characteristics of the data distribution. The calculation method is expressed as:

[0187]

[0188] Wherein, ∥∥ is the L2 norm, and re is the Gaussian kernel width parameter used to adjust the metric strength. Preferably, re is set to 2.

[0189] S504. Use the activation function to transform the weighted input to generate an activation output, and this output will be sent to the next layer or the output layer. Furthermore, at the end of each layer, according to the quality of the output of the current layer, adjust the feature recalibration parameters of the previous layer through backpropagation to ensure the effectiveness of information transmission. Specifically, the calculation method for processing the output of each layer through the non-linear activation function defined in the Riemann space is expressed as:

[0190] y u,j = f u (z u,j )

[0191] f u (z) = ReLU(z) = max(0, z)

[0192] Wherein, f u () is the activation function.

[0193] In one embodiment, the calculation method of the activation function f u (z) is expressed as:

[0194] f u (z) = γ u ·max(0, z)

[0195] Wherein, γ u is the activation function slope. Preferably, γu Set to 3.

[0196] S505. At the end of the network, category determination is completed through an output layer. The output layer calculates the classification probability using the maximum function and optimizes the model through the loss function. The calculation method of the classification probability is expressed as:

[0197] P u (y|X u ) = softmax(α u ·θ u (X u ) + β u ·s u (X u ))

[0198] Where α u and β u are weight parameters that adjust the influence of the original decision score and sensitivity. P u (y|X u ) is the probability of classifying y given the input X u . Y u is the actual label, K u is the number of classes, θ u (X) is the decision boundary of the model, and s u (X) is the boundary sensitivity of the sample. Preferably, α u and β u are set to 0.3 and 0.7 respectively.

[0199] In one embodiment, a probability recalibration module based on dynamic adjustment of the decision boundary is used to dynamically adjust the decision boundary during the training process to better adapt to the distribution characteristics of the data. Specifically, in each iteration, the decision boundary θ u (X u ) of the current model is dynamically calculated based on the current model parameters and input features. The calculation method is expressed as:

[0200] θ u (X u ) = W u ·X u +b u

[0201] Where W u and b u represent the weight and bias of the model respectively.

[0202] Moreover, to evaluate the position of each sample relative to the decision boundary, the boundary sensitivity s u ((X u ) is calculated in the following way:

[0203]

[0204] where λ u is a parameter for adjusting sensitivity.

[0205] Moreover, the method for calculating the loss function of the output layer based on class weights is expressed as:

[0206]

[0207] where is the loss function of the Riemann neural network, and ω k is the weight for each category. Preferably, ω k The weight is the reciprocal of the percentage of the number of samples of the k-th category in the current batch of samples to the total number of samples.

[0208] Furthermore, calculate the error between the output and the actual label, and update the network parameters using the gradient descent method. The way to update the weights and biases of the Riemann neural network is expressed as:

[0209]

[0210] where η u is the learning rate, and are the gradients of the loss function with respect to the weights and biases respectively.

[0211] S506. Repeat the above steps iteratively until the preset stop iteration condition is satisfied, which indicates that the model training is completed. In one embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0212] S6. Evaluate the effectiveness of dampening solution recycling. Use the trained feature extraction model, data dimensionality reduction model, and classifier model to process new samples and generate the final output result to achieve the evaluation of the effectiveness of dampening solution recycling. In one embodiment, the newly collected sample data extracts key features through the feature extraction model. Further, the extracted features are input into the trained data dimensionality reduction model for feature dimensionality reduction, retaining the information that is most critical to the decision-making process. Further, the dimension-reduced features are input into the classifier model for classification. In this embodiment, the classification categories include 4 levels: "excellent", "good", "medium", and "poor".

[0213] It has innovations in the following aspects compared with the prior art:

[0214] 1. The present invention adopts a generative adversarial network based on non-linear feature disentanglement, which allows input features to be mapped into a high-dimensional quantum feature space, and improves the data representation ability through the superposition and entanglement effects of quantum states, thereby effectively increasing the diversity and quality of the data set.

[0215] 2. The present invention uses a three-layer fully connected neural network with an adaptive wolf pack optimization algorithm for feature extraction, simulating the social behavior of wolf packs to optimize the parameters of the neural network, and improving the search efficiency and solution accuracy of the algorithm in the high-dimensional data space.

[0216] 3. The present invention uses a self-encoding neural network based on mask enhancement for feature dimensionality reduction, applying a random mask in the input layer to selectively hide some data features, enabling the network to learn and reconstruct complete data from partially visible data, and enhancing the model's learning ability for implicit data relationships.

[0217] 4. The present invention uses a Riemannian neural network based on activation function response metrics, using Riemannian metrics to optimize the connection weights between nodes, and improving the data transfer efficiency and classification accuracy in the network according to the distribution characteristics of data on the Riemannian manifold.

[0218] Embodiment 2

[0219] This embodiment provides an evaluation system for the recycling efficiency of dampening solution based on image recognition, including:

[0220] A data collection and annotation module for collecting data of the equipment using dampening solution, the recycling equipment, and related monitoring sensors;

[0221] A feature extraction module that inputs the expanded data into a feature extraction model for training the feature extraction model;

[0222] A feature dimensionality reduction module that inputs the data after feature extraction into a feature dimensionality reduction model for training the feature dimensionality reduction model;

[0223] A classification module that inputs the data after dimensionality reduction into a classifier for training the classifier model.

[0224] For the parameters and the steps of each unit module in the above-mentioned evaluation system for the recycling efficiency of dampening solution based on image recognition of the present invention to achieve the corresponding functions, reference can be made to the parameters and steps in the embodiment of the method for evaluating the recycling efficiency of dampening solution based on image recognition in the foregoing text, and details are not described herein again.

[0225] Embodiment 3

[0226] An electronic device according to an embodiment of the present invention includes: a processor and a memory. Among them, a computer program that can be called by the processor is stored in the memory. The processor executes a method for evaluating the recovery efficiency of dampening solution based on image recognition by calling the computer program stored in the memory. It should be noted that: all computer programs of a method for evaluating the recovery efficiency of dampening solution based on image recognition are implemented in the C language. Among them, the data acquisition module, the pollutant intrusion degree analysis module, the damage degree analysis module, the stain diffusion degree analysis module, the stain risk assessment module, the book processing warning module, and the control module are all controlled by a remote server.

[0227] Embodiment 4

[0228] This embodiment provides a computer-readable storage medium, on which a rewritable computer program is stored;

[0229] When the computer program runs on a computer device, it causes the computer device to execute the above-mentioned method for evaluating the recovery efficiency of dampening solution based on image recognition.

[0230] The embodiments of the present invention are all described in a progressive manner. For the same or similar parts among the embodiments, reference can be made to each other. The key points of each embodiment are the differences from other embodiments. In particular, for the embodiments of the Internet of Things devices and media, since they are basically similar to the method embodiments, the description is relatively simple. For the relevant parts, reference can be made to the partial description of the method embodiments.

[0231] The systems, media, and methods provided by the embodiments of the present invention are in one-to-one correspondence. Therefore, the systems and media also have beneficial technical effects similar to those of their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the systems and media will not be elaborated here.

[0232] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) that contain computer-usable program code.

[0233] The present invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each flow and / or block in the flowchart illustrations and / or block diagrams, and combinations of flows and / or blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions executed by the processor of the computer or other programmable data processing device create means for implementing the functions specified in the flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0234] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable memory produce a manufacture including instruction means that implement the functions specified in the flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0235] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and memory.

[0236] The memory may include non-permanent memory in the form of computer-readable media, random access memory (RAM) and / or non-volatile memory such as read only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0237] Computer-readable media includes both permanent and non-permanent, removable and non-removable media implemented by any method or technology for storing information. The information may be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), flash memory or other memory technologies, compact disc read only memory (CD-ROM), digital versatile discs (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transitory media that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media such as modulated data signals and carrier waves.

[0238] It should also be noted that the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, commodity or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, commodity or device. Without further limitation, an element defined by the statement "comprising one..." does not exclude the presence of additional identical elements in the process, method, commodity or device comprising the element.

[0239] The above are only embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. A method for evaluating the recovery efficiency of dampening solution based on image recognition, characterized in that, It includes the following steps: Step S1: Collect the data collected by the devices for using and recycling the dampening solution and the relevant monitoring sensors, and perform data annotation; Step S2: Generate samples based on the generative adversarial network algorithm for non-linear feature disentanglement, so as to perform data augmentation; Step S3: Input the augmented data into the feature extraction model to train the feature extraction model; Step S4: Input the data after feature extraction into the feature dimensionality reduction model to train the feature dimensionality reduction model; Step S5: Input the data after dimensionality reduction into the classifier to train the classifier model; Step S6: Use the trained feature extraction model, data dimensionality reduction model and classifier model to process new samples and generate the final output result.

2. The method for evaluating the recovery efficiency of dampening solution based on image recognition according to claim 1, wherein The training process of the generative adversarial network algorithm based on non-linear feature disentanglement is as follows: S201. Initialize the model parameters of generator G c and discriminator D c . S202: Map the original training data to a high-dimensional Hilbert space through the method of quantum feature mapping; S203. Generator G c Receives the mapped high-dimensional features and generates new data samples through its neural network structure; S204, Discriminator D c Classify the generated data and the real data; S205: Adjust the parameters of the generator according to the feedback of the discriminator, and constrain the parameter update of the discriminator through the adaptive loss function; S206: Repeat the above steps iteratively until the preset iteration stop condition is met, which means the model training is completed.

3. The method for evaluating the recovery efficiency of dampening solution based on image recognition according to claim 2, wherein The said step S3 includes the following specific steps: S301: Initialize the parameters of the neural network, where the parameters of the neural network include weights and biases; S302: Based on the adaptive wolf pack optimization algorithm, adjust the parameters of the neural network by simulating the social behavior of the wolf pack; S303: In each algorithm iteration, adjust the search strategy according to the optimization results of the previous generation; S304: Update the weights and biases of the neural network; S305: Repeat the above steps iteratively until the preset iteration stop condition is met, which means the model training is completed.

4. The method for evaluating the recovery efficiency of dampening solution based on image recognition according to claim 3, wherein The said step S4 includes the following specific steps: S401: Initialize the parameters of the autoencoder network, including the weight and bias parameters of the autoencoder network; S402: According to the framework of the greedy algorithm, set the independent training cycles and learning rates of each layer; S403. At the beginning of each training cycle, for each input vector X th , randomly generate a mask matrix m for randomly masking part of the input data; S404: After the input data is masked, it is sent to the encoder, and the encoder compresses the data into a low-dimensional feature representation through the activation function; S405: The low-dimensional features are sent to the decoder, and the decoder tries to reconstruct the original input data. By calculating the loss function, the backpropagation algorithm updates the parameters in the network; S406: Use the greedy algorithm to optimize the parameters of each layer; S407: When the training of all layers is completed, integrate the models of each layer to form a complete autoencoder model; S408: Repeat the above steps iteratively until the preset iteration stop condition is met, which means the model training is completed.

5. The method for evaluating the recovery efficiency of dampening solution based on image recognition according to claim 4, characterized in that, The training of the classifier model by inputting the data after dimensionality reduction into the classifier in the said step S5 includes the following specific contents: S501: Initialize the parameters of the Riemannian neural network; S502: After the data after dimensionality reduction is input, first perform a preliminary linear transformation on the data through a preprocessing layer, and adopt an adaptive feature recalibration strategy to dynamically adjust its transmission weight in the network according to the activation response of each feature; S503. At each hidden layer node, calculate the inner product defined by the Riemannian metric for the input features, and use this as the weighted input before neuron activation to adapt to the Riemannian manifold structure of the data; S504. Use an activation function to transform the weighted input to generate an activation output, which will be sent to the next layer or the output layer; S505. At the end of the network, complete class determination through an output layer. The output layer uses a maximum function to calculate the classification probability and optimizes the model through a loss function; S506. Repeat the above steps iteratively until the preset stop iteration condition is met, which indicates that the model training is completed.

6. The method for evaluating the recycling efficiency of dampening solution based on image recognition according to claim 5, characterized in that, The step S6 includes the following specific steps: Use the trained feature extraction model, data dimensionality reduction model, and classifier model to process new samples and generate a final output result to achieve the evaluation of the fountain solution recovery efficiency.

7. The method for evaluating the recovery efficiency of dampening solution based on image recognition according to claim 6, wherein In S205, the parameter update method of the discriminator is the backpropagation method based on gradient descent, and the adaptive loss function combines the quantum fidelity between the generated samples and the real samples.

8. An evaluation system for the recycling efficiency of dampening solution based on image recognition, which is implemented based on the method for evaluating the recycling efficiency of dampening solution based on image recognition described in any one of claims 1-7, and is characterized in that The system includes: A data collection and annotation module for collecting data from the fountain solution usage equipment, recovery equipment, and related monitoring sensors; A feature extraction module that inputs the augmented data into the feature extraction model for training the feature extraction model; A feature dimensionality reduction module that inputs the data after feature extraction into the feature dimensionality reduction model for training the feature dimensionality reduction model; A classification module that inputs the data after dimensionality reduction into the classifier for training the classifier model.

9. An electronic device, comprising: A processor and a memory. The memory stores a computer program that can be called by the processor. It is characterized in that the processor executes a method for evaluating the fountain solution recovery efficiency based on image recognition according to any one of claims 1-7 by calling the computer program stored in the memory.

Images