Method and system for identifying abnormal operation behavior on cut tobacco production process site

Abstract

The invention discloses a method and a system for identifying abnormal operation behaviors on a cut tobacco production process site. The method comprises the following steps: receiving a real-time video stream from the cut tobacco production process site; extracting key visual features in the real-time video stream by using a feature extraction model; inputting the key visual features into a data dimension reduction model to obtain a first feature vector after dimension reduction; inputting the dimensionality-reduced first feature vector into a classifier model, and outputting a classification result of the operation behavior, namely normal operation or abnormal operation; wherein when the classifier model is trained, the weight and bias of the extreme learning machine are optimized based on a differential evolution algorithm. According to the method, feature extraction and feature dimension reduction are performed on each group of input data to improve the precision of the input data of the classifier model, and the weight and bias of the extreme learning machine are optimized based on the differential evolution algorithm to improve the precision and efficiency of the classification model, so that the classification accuracy is improved. Therefore, the problems of low accuracy and high false alarm rate in a complex industrial environment are solved.

Description

Technical Field

[0001] The present application relates to the technical field of cigarette shred production, and more specifically, to a method and system for identifying abnormal operation behaviors at a shred production process site. Background Art

[0002] In modern industrial production, especially in high-speed silk production lines, real-time and effective monitoring of operating behaviors is crucial to ensure production safety and efficiency. Although conventional video surveillance systems can record in real time, they usually lack effective automated analysis functions to identify abnormal operations, which may lead to potential safety risks not being discovered and handled in a timely manner. Traditional abnormal behavior recognition technologies mostly rely on simple image processing techniques and basic pattern recognition algorithms. These methods often face problems such as low accuracy and high false alarm rates in complex industrial environments.

[0003] In addition, due to the lack of sufficient training samples and data diversity, the generalization and robustness of these models are often insufficient, making it difficult for them to cope with changing working environments and complex operational behaviors. Summary of the invention

[0004] The present application provides a method and system for identifying abnormal operation behaviors at a silk-making process site. For each set of input data, feature extraction and feature dimension reduction are performed through a trained model to improve the accuracy of the input data of the classifier model, and the weights and biases of the extreme learning machine are optimized based on a differential evolution algorithm to improve the accuracy and efficiency of the classification model, thereby solving the problems of low accuracy and high false alarm rate in complex industrial environments.

[0005] The present application provides a method for identifying abnormal operation behaviors at a silk-making process site, comprising:

[0006] Receive real-time video stream from the silk-making process site;

[0007] Extract key visual features from real-time video streams using feature extraction models;

[0008] Input the key visual features into the data dimension reduction model to obtain the first feature vector after dimension reduction;

[0009] The first feature vector after dimension reduction is input into the classifier model, and the classification result of the operation behavior is output, that is, normal operation or abnormal operation;

[0010] When training the classifier model, the weights and biases of the extreme learning machine are optimized based on the differential evolution algorithm.

[0011] Preferably, before training the feature extraction model, training data is first obtained, specifically including:

[0012] receiving an image set collected in a silk-making process, wherein each image in the image set has annotation information;

[0013] Train a visual saliency-based generative adversarial network as a data augmentor;

[0014] Based on the image set, a data expander is used to expand the image to obtain expanded data.

[0015] Preferably, when training the feature extraction model, a neural network parameter optimization algorithm based on the fusion phenomenon principle is used to optimize the parameters of the feature extraction model.

[0016] Preferably, the data dimension reduction model uses an improved autoencoder algorithm to perform feature dimension reduction.

[0017] Preferably, training the data dimensionality reduction model includes repeatedly iterating the following steps until a preset stop iteration condition is met, indicating that the model training is completed:

[0018] Randomly initialize the encoder and decoder weights and biases according to a normal distribution;

[0019] The second feature vector is input into the encoder to achieve dimensionality reduction;

[0020] Use nonlinear activation functions to process hidden layer outputs, increase the nonlinear capabilities of the model and improve the efficiency of information expression;

[0021] The hidden layer output passes through the decoder to obtain the third feature vector;

[0022] Compare the third eigenvector with the second eigenvector and calculate the reconstruction error using the mean square error loss function;

[0023] According to the reconstruction error, the weights and biases of the encoder and decoder are adjusted, and the influence of quantum state regularization is still considered during the optimization process;

[0024] The network parameters are updated according to the results of back-propagation to reduce the reconstruction error and enhance the stability of the model.

[0025] The present application also provides a system for identifying abnormal operation behaviors at a silk-making process site, comprising a first receiving module, a feature extraction module, a dimensionality reduction module, and a classification module;

[0026] The first receiving module is used to receive the real-time video stream from the silk making process site;

[0027] The feature extraction module is used to extract key visual features in real-time video streams using a feature extraction model;

[0028] The dimension reduction module is used to input the key visual features into the data dimension reduction model to obtain the first feature vector after dimension reduction;

[0029] The classification module is used to input the first feature vector after dimension reduction into the classifier model, and output the classification result of the operation behavior, that is, normal operation or abnormal operation;

[0030] When training the classifier model, the weights and biases of the extreme learning machine are optimized based on the differential evolution algorithm.

[0031] Preferably, the abnormal operation behavior recognition system at the silk-making process site further includes a second receiving module, a first training module, and an expansion module;

[0032] The second receiving module is used to receive an image set collected in the silk making process, each image in the image set having annotation information;

[0033] The first training module is used to train a visual saliency-based generative adversarial network as a data expander;

[0034] The expansion module is used to expand the image based on the image set using the data expander to obtain the expanded data.

[0035] Preferably, the abnormal operation behavior identification system at the silk-making process site also includes a second training module, which is used to optimize the parameters of the feature extraction model by using a neural network parameter optimization algorithm based on the fusion phenomenon principle when training the feature extraction model.

[0036] Preferably, the data dimension reduction model uses an improved autoencoder algorithm to perform feature dimension reduction.

[0037] Preferably, the abnormal operation behavior recognition system at the silk-making process site further includes a third training module, which includes an initialization module, an encoding module, a hidden layer processing module, a decoding module, a reconstruction error calculation module, an optimization module, and a network parameter updating module;

[0038] The initialization module is used to randomly initialize the weights and biases of the encoder and decoder according to the normal distribution;

[0039] The encoding module is used to input the second feature vector into the encoder to achieve dimensionality reduction;

[0040] The hidden layer processing module is used to process the hidden layer output using a nonlinear activation function, thereby increasing the nonlinear capability of the model and improving the efficiency of information expression;

[0041] The decoding module is used to pass the hidden layer output through the decoder to obtain the third feature vector;

[0042] The reconstruction error calculation module is used to compare the third eigenvector and the second eigenvector, and calculate the reconstruction error using a mean square error loss function;

[0043] The optimization module is used to adjust the weights and biases of the encoder and decoder according to the reconstruction error, and the influence of quantum state regularization is still considered during the optimization process;

[0044] The network parameter update module is used to update the network parameters according to the results of back propagation to reduce the reconstruction error and enhance the stability of the model.

[0045] Other features and advantages of the present application will become apparent from the following detailed description of exemplary embodiments of the present application with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate embodiments of the application and, together with the description, serve to explain the principles of the application.

[0047] Figure 1 A flow chart of a method for identifying abnormal operation behavior at a silk-making process site provided in this application;

[0048] Figure 2 This is a structural diagram of the abnormal operation behavior identification system at the silk-making process site provided in this application. DETAILED DESCRIPTION

[0049] Various exemplary embodiments of the present application will now be described in detail with reference to the accompanying drawings. It should be noted that unless otherwise specifically stated, the relative arrangement of components and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present application.

[0050] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present application, its application, or uses.

[0051] Technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered part of the specification.

[0052] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0053] The present application provides a method and system for identifying abnormal operation behaviors at a silk-making process site. For each set of input data, feature extraction and feature dimension reduction are performed through a trained model to improve the accuracy of the input data of the classifier model, and the weights and biases of the extreme learning machine are optimized based on a differential evolution algorithm to improve the accuracy and efficiency of the classification model, thereby solving the problems of low accuracy and high false alarm rate in complex industrial environments.

[0054] like Figure 1 As shown, the method for identifying abnormal operation behavior at the silk-making process site provided by the present application includes:

[0055] S110: receiving a real-time video stream from a silk-making process site, where the real-time video stream has undergone basic preprocessing, such as formatting and standardization.

[0056] S120: Extract key visual features from the real-time video stream using a feature extraction model.

[0057] S130: Input the key visual features into the data dimension reduction model to obtain the first feature vector after dimension reduction, further compress the data scale and improve the processing efficiency.

[0058] S140: Inputting the first feature vector after dimension reduction into a classifier model, and outputting a classification result of the operation behavior, that is, a normal operation or an abnormal operation;

[0059] It should be noted that the feature extraction model, data dimensionality reduction model and classifier model are all trained models.

[0060] Preferably, before training the feature extraction model, training data is first obtained, specifically including:

[0061] P1: Receive an image set collected during the silk-making process, where each image in the image set has annotation information.

[0062] The image set of this application is collected from the monitoring system of the silk production line, using a high-resolution camera fixed at a key operating position to record the operation of the silk production process in real time. The collected image data is stored in JPEG format. When the collected image set is labeled with categories, the labeled categories are "normal operation" and "abnormal operation".

[0063] P2: Training a visual saliency-based generative adversarial network as a data augmentor.

[0064] It is understandable that the collection, acquisition, annotation 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, while affecting the accuracy of the model. This application uses a generative adversarial network algorithm based on visual saliency for data expansion. The generative adversarial network algorithm based on visual saliency adopts the basic framework of a generative adversarial network, including a generator (G) and a discriminator (D). The generator is responsible for generating fake image data that is visually indistinguishable, increasing data diversity; the discriminator attempts to distinguish between real images and fake images generated by the generator to train the generator to improve generation quality.

[0065] Specifically, the training process of the generative adversarial network algorithm based on visual saliency is as follows:

[0066] P21: Initialize the network parameters of the generator and discriminator. The initialization method can be expressed as:

[0067]

[0068] In the formula, bθ g is the parameter of the generator; bθ d are the parameters of the discriminator; It indicates a normal distribution with a standard deviation of 0.02.

[0069] P22: Use a generator to generate fake images from random noise data. Specifically, the generator G generates a fake image bI from random noise bz g can be expressed as:

[0070] I g =G(bz;bθ g )

[0071] In the formula, is the random noise of the input; G is the generator function.

[0072] P23: For the generated fake image, use the visual saliency algorithm to calculate its saliency map, compare it with the saliency map of the real image, and calculate the saliency loss L between the generated fake image and the real image sal , which can be expressed as:

[0073]

[0074] Where S represents the significance calculation function; bI gi is the generated fake image; bI ri is the corresponding real image; N is the number of images. Preferably, the saliency calculation function S is implemented based on a local contrast method.

[0075] P24: The discriminator classifies the fake images and real images from the generator, trains its ability to distinguish, and calculates the discrimination loss L d , which can be expressed as:

[0076]

[0077] Where D is the discriminator function; log is the natural logarithm function.

[0078] Furthermore, the calculation method of the discriminator function D can be expressed as:

[0079]

[0080] Where W d represents the weight matrix of the discriminator; b dis the bias term of the discriminator; e is the base of the natural exponential; bx represents the input image data.

[0081] P25: According to the feedback from the discriminator, the parameters of the generator are adjusted to optimize the quality of the generated image and the accuracy of the salient features, which can be expressed as:

[0082]

[0083] In the formula, bθ g ' is the updated generator parameter, bθ g is the generator parameter before updating; α tr is the learning rate; represents the gradient of the generator parameters; L d is the judgment loss; L sal is the significance loss; tr is the loss weight. Preferably, α tr Set to 0.01, λ tr Set to 0.3.

[0084] Furthermore, the gradient of the discriminative loss to the generator parameters is calculated can be expressed as:

[0085]

[0086] In the formula, is the back-propagation term in the generator network.

[0087] Furthermore, the gradient of the saliency loss with respect to the generator parameters is calculated can be expressed as:

[0088]

[0089]

[0090] In the formula, S ′ (bI g ) represents the derivative of the saliency function with respect to the image.

[0091] P26: Repeat the above steps until the preset stop iteration condition is met, indicating that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0092] P3: Based on the image set, use the data expander to expand the image and obtain the expanded data.

[0093] After data expansion, training data is obtained and the feature extraction model can be trained. Some existing solutions use neural networks for feature extraction. In some neural network structures, gradient vanishing, gradient exploding or falling into local optimal solutions may occur, affecting the stability of training and the performance of the model.

[0094] This application uses a 3-layer fully connected neural network for feature extraction. When training the feature extraction model, a neural network parameter optimization algorithm based on the principle of fusion phenomenon is used to optimize the parameters of the feature extraction model. Inspired by the fusion phenomenon in nature, lighter atomic nuclei merge under high temperature and high pressure to form more stable heavy atomic nuclei, while releasing huge energy. The neural network parameter optimization algorithm based on the principle of fusion phenomenon optimizes the weights and biases of the network by simulating the fusion process. The weights and biases are regarded as atomic nuclei, and their merging and transformation processes are designed to achieve the optimal state of the loss function. Unlike traditional optimization algorithms based on gradient descent, the neural network parameter optimization algorithm based on the principle of fusion phenomenon does not rely on gradient information, thereby avoiding the problem of gradient disappearance or explosion.

[0095] Specifically, the training process of the neural network parameter optimization algorithm based on the fusion phenomenon principle is as follows:

[0096] Q1. Define the structure of the neural network. The three-layer fully connected neural network used includes an input layer and two hidden layers, and a preset output layer for category supervision training, which are defined as follows:

[0097] Input layer, assuming the dimension of the input vector is d, then the input layer has d neurons, corresponding to the d elements in the vector;

[0098] Hidden layer, both hidden layers have 128 neurons and use ReLU as the activation function;

[0099] Output layer,The output layer uses the softmax function as the activation function.

[0100] Furthermore, the neural network parameters are initialized and assigned a random energy state. Specifically, the weights w and biases b of the neural network are assigned initial values, and the energy state E(w,b) is defined to describe the potential optimization ability of the parameters, which can be expressed as:

[0101]

[0102] In the formula, x i Represents input data, y i is the corresponding target output, N is the number of samples in the dataset, and K is the number of feature scales. k(t) is the adaptive weight of the kth scale, which depends on the training time t and the current performance state of the network, ensuring that the focus of the loss function is adjusted at different stages of training.

[0103] Furthermore, α k (t) is dynamically updated at each iteration, which allows the model to focus on different feature levels at different training stages, thereby more effectively learning the multi-level structure of complex data. Specifically, the weight α of each scale k The update of (t) depends on the effect of the scale under the current training state. The update method of the adaptive weight can be expressed as:

[0104]

[0105] In the formula, E k (t) is the loss value of the kth scale at time t; β k is the importance adjustment coefficient of the kth scale, which is preset artificially according to the characteristics of the scale.

[0106] It should be noted that the output layer is only used for energy state supervision during training, and the output of the feature extraction model is the output of the second hidden layer of the neural network, that is, the feature vector after feature extraction.

[0107] Q2, perform nuclear synthesis cycle operation, and check the fusion candidate for each pair of parameters (weight and bias). If the synthesis of the two parameters can reduce the overall system energy, fusion is allowed, otherwise it remains independent. Specifically, for each pair of parameters w j and b k An assessment is made to determine whether the total energy of the system can be reduced through fusion, which can be expressed as:

[0108] ΔE=E(w j ′ ,b ′ k )-E(w j ,b k )

[0109] Moreover, the dramatic change behavior of the parameters can be expressed as:

[0110] w j ′ =w j +δ tr

[0111] b ′ k =b k +∈ tr

[0112] In the formula, w j ′and b ′ k are the updated neural network parameters respectively. If ΔE<Θ(t), the new parameter w is accepted. j ′ and b ′ k , otherwise keep the original parameters unchanged. Θ(t) is the energy threshold of the current iteration.

[0113] Furthermore, the energy threshold Θ(t) is updated at each iteration. The energy changes during the training process are monitored in real time, and the energy threshold is adjusted dynamically to control the sensitivity and frequency of parameter updates. Specifically, the update rule of the energy threshold is defined as:

[0114]

[0115] Where Θ(t) and Θ(t+1) are the energy thresholds for the current and next time steps respectively; ν eo is the learning rate of threshold adjustment, which controls the speed of adjustment; E(t) is the average energy of the current time step; E target is the target energy, which is preset according to the task requirements; σ E is the standard deviation of the energy variation; tanh is the hyperbolic tangent function. Preferably, ν eo Set to 0.01, E target Set to 0.1.

[0116] Furthermore, the weight adjustment amount δ tr and bias adjustment ∈ tr The calculation method can be expressed as:

[0117]

[0118] In the formula, σ to and η to is the sensitivity and scaling factor of the fusion adjustment. Preferably, σ to and η to Set them to 0.1 and 0.3 respectively.

[0119] Q3. Perform energy evaluation on the synthesized new parameter set, update the parameter state, and push the parameter set in the high energy state to the low energy state to achieve the desired state stability. Specifically, the synthesized parameter set is re-evaluated through the energy function E to ensure the energy is minimized, which can be expressed as:

[0120]

[0121] In the formula, p(y i |x i ; w, b) means that given the input x i With parameters w, b, output yi probability.

[0122] Furthermore, logp(y i |x i ; The calculation method of w, b) can be expressed as:

[0123]

[0124] In the formula, σ std Indicates the standard deviation of the output data.

[0125] Q4. Adjust the nuclear reaction temperature according to the current network performance and stability, simulate the temperature control in physical fusion, and adjust the fusion temperature T, which can be expressed as:

[0126] T new =T old ·exp(-α te ΔE)

[0127] In the formula, α te is the temperature adjustment factor, T new and T old are the temperatures before and after adjustment, and ΔE is the energy change.

[0128] Furthermore, α te Based on dynamic adjustment of energy changes, its calculation method can be expressed as:

[0129]

[0130] Here, κ te is a preset adjustment coefficient, |ΔE| is the absolute energy change, E avg is the average value of all current parameter energies.

[0131] Q5. Check whether the energy of the network parameters reaches a dynamic equilibrium state. If equilibrium is reached, continue the optimization process; if not, perform parameter re-fusion processing. In one embodiment, the judgment condition can be expressed as:

[0132] Var(E)<∈ mr

[0133] Where Var(E) is the variance function of energy E, preferably, ∈ mr Set to 10 -5 .

[0134] Q6. Decide whether to update the current parameter configuration based on the energy optimization principle, which can be expressed as:

[0135]

[0136] bnew =b ′ k +β rate Δb

[0137] In the formula, β rate To update the learning rate, is the fusion adjustment factor. Preferably, the learning rate β is set to 0.001.

[0138] Furthermore, the weight and bias parameters are updated using momentum terms, and the calculation method can be expressed as:

[0139]

[0140] In the formula, μ ea is the momentum coefficient, Δw prev and Δb prev are the weight and bias updates of the previous iteration respectively.

[0141] Furthermore, the fusion adjustment factor According to the historical changes of weights and the error rate of the current layer, it is dynamically set and can be expressed as:

[0142]

[0143] In the formula, is the difference between the current weight and the weight of the previous iteration; is the standard deviation of the weight change of the current layer; E l is the error rate of the current layer; ρ yt and λ yt is the adjustment coefficient, which controls the influence intensity of weight change rate and error rate respectively; tanh is the hyperbolic tangent function. Preferably, ρ yt and λ yt Set them to 0.3 and 0.5 respectively.

[0144] Q7. Determine whether the algorithm has converged by the change in parameter energy after multiple consecutive iterations. If converged, output the final optimized parameters; if not, return to step 2 to continue iterating. The convergence judgment method can be expressed as:

[0145]

[0146] In the formula, E new is the energy state after parameter update, E old is the energy state before parameter update, γ ew Preferably, the convergence threshold γ ew Set to 10 -6 .

[0147] After the features in the training set are extracted using the feature extraction model, they are input into the data dimensionality reduction model for training. Among them, the data dimensionality reduction model uses an improved autoencoder algorithm for feature dimensionality reduction. The improved autoencoder algorithm consists of two main parts: an encoder and a decoder. The encoder is responsible for reducing the input feature vector to a hidden layer of a lower dimension, and the decoder is responsible for restoring the low-dimensional hidden layer data to the original dimension to evaluate the effect of dimensionality reduction. The present application uses quantum state regularization terms in the encoder and decoder to stabilize the learning process and prevent overfitting.

[0148] Training the data dimensionality reduction model includes:

[0149] Repeat the following steps until the preset stop iteration condition is met, indicating that the model training is complete:

[0150] R1: Randomly initialize the weights and biases of the encoder and decoder according to the normal distribution. Specifically, the network parameters are initialized using the normal distribution, which can be expressed as:

[0151]

[0152] In the formula, pW is the network weight; pb is the network bias; represents a normal distribution with a mean of 0 and a standard deviation of 0.01.

[0153] R2: Input the second feature vector into the encoder to achieve dimensionality reduction.

[0154] In specific implementation, the forward propagation method can be expressed as:

[0155] ph=Sig(pW e ·px+pb e )

[0156] Where ph is the hidden layer output; Sig() represents the Sigmoid nonlinear activation function; pW e and pb e are the weights and biases of the encoder respectively.

[0157] Furthermore, the encoder weight pW e The update method can be expressed as:

[0158] pW e =pW init -η aw ·ΔpW e

[0159] Where, pW init is the initial weight, η aw is the learning rate, ΔpW eis the gradient of the weight, which is calculated by error back propagation and can be expressed as:

[0160]

[0161] Wherein, Backprop() indicates that the gradient is calculated using the back propagation algorithm. Preferably, η aw Set to 0.01.

[0162] R3: Use nonlinear activation function to process hidden layer output, increase the nonlinear ability of the model and improve the efficiency of information expression. The activation function of the hidden layer can be expressed as:

[0163] pa=ReLU(ph)

[0164] Where pa is the output of the hidden layer after activation; ReLU() is the ReLU activation function, defined as:

[0165] ReLU(x)=max(0,x)

[0166] Where x is the input of the activation function.

[0167] R4: The hidden layer output passes through the decoder and is restored to data close to the original feature vector to obtain the third feature vector.

[0168] The decoding process can be expressed as:

[0169]

[0170] In the formula, is the decoded output (i.e., the third eigenvector), which is as close as possible to the original input px; pW d and pb d are the decoder weights and biases.

[0171] Furthermore, the decoder weight pW d The calculation method can be expressed as:

[0172] pW d =pW d_init -η aw ·ΔpW d

[0173] Where, pW d_init is the initial weight of the decoder, ΔpW d It is obtained by back propagation similar to the encoder and can be expressed as:

[0174]

[0175] R5: Compare the third eigenvector and the second eigenvector, and use the mean square error loss function to calculate the reconstruction error, which can be expressed as:

[0176]

[0177] Where pL is the reconstruction error; n is the number of samples; px is the original input; is the decoded output.

[0178] R6: According to the reconstruction error, adjust the weights and biases of the encoder and decoder, and continue to consider the influence of quantum state regularization during the optimization process. It can be expressed as:

[0179]

[0180] Where, pW e '、pb e ' is the updated encoder weight and bias, pW d '、pb d ' is the updated decoder weight and bias, α ut is the learning rate; are the gradients of the decoder weights and biases before the update; is the gradient of the encoder weights and biases before the update. Preferably, α ut Set to 0.005.

[0181] Taking the gradient calculation of the encoder weight as an example, its calculation method can be expressed as:

[0182]

[0183] In the formula, It means that in the i-th sample, the reconstructed output For the encoder weight pW e The partial derivatives of are calculated using the chain rule.

[0184] R7: Update the network parameters according to the results of back propagation to reduce the reconstruction error and enhance the stability of the model.

[0185] Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0186] The reduced-dimensional data is input into the classifier to train the classifier model. In this application, when training the classifier model, the weights and biases of the extreme learning machine are optimized based on the differential evolution algorithm. The extreme learning machine based on the differential evolution algorithm includes the following layers:

[0187] Input layer: receives the feature vector obtained from the feature dimension reduction model training step;

[0188] Hidden layer: Use multiple neurons to process input data. The number and activation function are adjusted according to the actual data.

[0189] Output layer: maps the processing results of the hidden layer to the predetermined classification labels.

[0190] Specifically, the training process of the extreme learning machine algorithm based on the differential evolution algorithm is as follows:

[0191] L1: Randomly initialize the weights and biases of the neural network according to the differential evolution algorithm. Specifically, initialize the weights qW and biases qb of the neural network. The initial value setting method can be expressed as:

[0192] QW init ,qb init =DiffEvol(pop_size,dim,iter)

[0193] Where DiffEvol() represents the differential evolution function; pop_size is the population size; dim is the problem dimension; iter is the number of iterations.

[0194] L2: Assume that the feature vector of the input data after feature dimension reduction is expressed as qx is the input data matrix; qN is the number of samples; qd is the data dimension.

[0195] Furthermore, the data is processed in the hidden layer, and the activation function used in the hidden layer is the Sigmoid activation function, which can be expressed as:

[0196] qh=Sig(qW·qx+qb)

[0197] Where qh is the hidden layer output; Sig() is the Sigmoid activation function; qW is the weight matrix; qb is the bias vector.

[0198] Furthermore, the output of the hidden layer is transmitted to the output layer, and the output layer makes classification decisions based on the trained parameters, which can be expressed as:

[0199] qy=softmax(qW o ·qh+qb o )

[0200] Where qy is the output layer result; softmax() is the Softmax function; qW o and qb o are the weight and bias of the output layer, respectively.

[0201] Furthermore, the gradient descent method is used to update the parameters, evaluate the classification accuracy of the model during the optimization process, and adjust the weights and biases of the neural network, which can be expressed as:

[0202]

[0203] Where qW' and qb' are the weights and biases of the extreme learning machine after the update, qW and qb are the weights and biases of the extreme learning machine before the update, and α hu is the learning rate; and are the gradients of the loss function qL with respect to weights and biases, respectively. Preferably, α hu Set to 0.01.

[0204] Furthermore, the loss function qL uses the mean square error, and the calculation method can be expressed as:

[0205]

[0206] In the formula, qy i is the ith predicted output, is the i-th actual label, and qN is the total number of samples.

[0207] Furthermore, the corresponding gradient calculation method can be expressed as:

[0208]

[0209] In the formula, the partial derivative and It is calculated using the chain rule.

[0210] L3: Repeat the above steps until the preset stop iteration condition is met, indicating that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0211] Based on the above, the present application also provides a system for identifying abnormal operation behaviors at the silk-making process site. Figure 2 As shown, the abnormal operation behavior identification system includes a first receiving module 210, a feature extraction module 220, a dimension reduction module 230 and a classification module 240.

[0212] The first receiving module 210 is used to receive a real-time video stream from the silk making process site.

[0213] The feature extraction module 220 is used to extract key visual features in the real-time video stream using a feature extraction model.

[0214] The dimension reduction module 230 is used to input the key visual features into the data dimension reduction model to obtain the first feature vector after dimension reduction.

[0215] The classification module 240 is used to input the first feature vector after dimension reduction into the classifier model, and output the classification result of the operation behavior, that is, normal operation or abnormal operation.

[0216] When training the classifier model, the weights and biases of the extreme learning machine are optimized based on the differential evolution algorithm.

[0217] Preferably, the abnormal operation behavior recognition system at the silk-making process site further includes a second receiving module 250 , a first training module 260 , and an expansion module 270 .

[0218] The second receiving module 250 is used to receive an image set collected in the silk making process, and each image in the image set has annotation information.

[0219] The first training module 260 is used to train a visual saliency-based generative adversarial network as a data expander.

[0220] The expansion module 270 is used to perform image expansion based on the image set using a data expander to obtain expanded data.

[0221] Preferably, the abnormal operation behavior identification system at the silk-making process site also includes a second training module 280, which is used to optimize the parameters of the feature extraction model by using a neural network parameter optimization algorithm based on the fusion phenomenon principle when training the feature extraction model.

[0222] Preferably, the data dimension reduction model uses an improved autoencoder algorithm to perform feature dimension reduction.

[0223] Preferably, the abnormal operation behavior identification system at the silk making process site also includes a third training module 290, and the third training module 290 includes an initialization module 2901, an encoding module 2902, a hidden layer processing module 2903, a decoding module 2904, a reconstruction error calculation module 2905, an optimization module 2906 and a network parameter updating module 2907.

[0224] The initialization module 2901 is used to randomly initialize the weights and biases of the encoder and decoder according to a normal distribution.

[0225] The encoding module 2902 is used to input the second feature vector into the encoder to achieve dimensionality reduction.

[0226] The hidden layer processing module 2903 is used to process the hidden layer output using a nonlinear activation function, thereby increasing the nonlinear capability of the model and improving the efficiency of information expression.

[0227] The decoding module 2904 is used to pass the hidden layer output through the decoder to obtain the third feature vector.

[0228] The reconstruction error calculation module 2905 is used to compare the third eigenvector and the second eigenvector, and calculate the reconstruction error using a mean square error loss function.

[0229] The optimization module 2906 is used to adjust the weights and biases of the encoder and decoder according to the reconstruction error, and the influence of quantum state regularization continues to be considered during the optimization process.

[0230] The network parameter updating module 2907 is used to update the network parameters according to the results of back propagation to reduce the reconstruction error and enhance the stability of the model.

[0231] The beneficial effects of this application are as follows:

[0232] 1. Generative adversarial networks based on visual saliency are used for data augmentation. By comparing the saliency maps of real images and generated images, the quality of generated images is optimized. High-quality fake images generated by generative adversarial networks based on visual saliency can effectively expand the data set and enhance the model's generalization and recognition capabilities for different operating situations.

[0233] 2. The neural network parameter optimization algorithm based on the fusion phenomenon principle simulates the nuclear fusion process, optimizes the network weights and biases, does not rely on gradient information, improves the stability of the network training process, thereby avoiding the common gradient vanishing or explosion problems in traditional neural network training and improving training efficiency.

[0234] 3. The improved autoencoder algorithm adds a quantum state regularization term to the autoencoder to stabilize the learning process and prevent overfitting, which helps to improve the generalization ability of the model and make the model more stable and reliable.

[0235] 4. Based on the differential evolution algorithm, the weights and biases of the extreme learning machine are optimized to improve the accuracy and efficiency of the classification model, so that classification tasks can be processed more accurately and quickly.

[0236] Although some specific embodiments of the present application have been described in detail by way of example, it should be understood by those skilled in the art that the above examples are only for illustration, not for limiting the scope of the present application. It should be understood by those skilled in the art that the above embodiments may be modified without departing from the scope and spirit of the present application. The scope of the present application is defined by the appended claims.

Claims

1. A method for identifying abnormal operation behavior at a silk-making process site, characterized in that: include: Receive real-time video stream from the silk-making process site; Extracting key visual features from the real-time video stream using a feature extraction model; Inputting the key visual feature into a data dimension reduction model to obtain a first feature vector after dimension reduction; Inputting the first feature vector after dimension reduction into a classifier model, and outputting a classification result of the operation behavior, that is, normal operation or abnormal operation; Wherein, when training the classifier model, the weights and biases of the extreme learning machine are optimized based on the differential evolution algorithm.

2. The method for identifying abnormal operation behavior at a silk-making process site according to claim 1, characterized in that: Before training the feature extraction model, first obtain the training data, including: receiving an image set collected in a silk-making process, wherein each image in the image set has annotation information; Train a visual saliency-based generative adversarial network as a data augmentor; Based on the image set, a data expander is used to perform image expansion to obtain expanded data.

3. The method for identifying abnormal operation behavior at a silk-making process site according to claim 1, characterized in that: When training the feature extraction model, a neural network parameter optimization algorithm based on the fusion phenomenon principle is used to optimize the parameters of the feature extraction model.

4. The method for identifying abnormal operation behavior at a silk-making process site according to claim 1, characterized in that: The data dimension reduction model uses an improved autoencoder algorithm to perform feature dimension reduction.

5. The method for identifying abnormal operation behavior at a silk-making process site according to claim 4, characterized in that: Training the data dimensionality reduction model includes repeatedly iterating the following steps until the preset stop iteration condition is met, indicating that the model training is complete: Randomly initialize the encoder and decoder weights and biases according to a normal distribution; The second feature vector is input into the encoder to achieve dimensionality reduction; Use nonlinear activation functions to process hidden layer outputs, increase the nonlinear capabilities of the model and improve the efficiency of information expression; The hidden layer output passes through the decoder to obtain the third feature vector; Comparing the third eigenvector with the second eigenvector, and calculating a reconstruction error using a mean square error loss function; According to the reconstruction error, the weights and biases of the encoder and the decoder are adjusted, and the influence of quantum state regularization is continuously considered during the optimization process; The network parameters are updated according to the results of back-propagation to reduce the reconstruction error and enhance the stability of the model.

6. A system for identifying abnormal operation behavior at a silk-making process site, characterized in that: It includes a first receiving module, a feature extraction module, a dimension reduction module and a classification module; The first receiving module is used to receive a real-time video stream from the silk making process site; The feature extraction module is used to extract key visual features in the real-time video stream using a feature extraction model; The dimension reduction module is used to input the key visual features into a data dimension reduction model to obtain a first feature vector after dimension reduction; The classification module is used to input the first feature vector after dimension reduction into a classifier model, and output a classification result of the operation behavior, that is, normal operation or abnormal operation; Wherein, when training the classifier model, the weights and biases of the extreme learning machine are optimized based on the differential evolution algorithm.

7. The abnormal operation behavior recognition system at the silk-making process site according to claim 6, characterized in that: It also includes a second receiving module, a first training module, and an expansion module; The second receiving module is used to receive an image set collected in the silk making process, each image in the image set having annotation information; The first training module is used to train a visual saliency-based generative adversarial network as a data expander; The expansion module is used to perform image expansion based on the image set using a data expander to obtain expanded data.

8. The abnormal operation behavior recognition system at the silk-making process site according to claim 6, characterized in that: It also includes a second training module, which is used to optimize the parameters of the feature extraction model by using a neural network parameter optimization algorithm based on the principle of fusion phenomenon when training the feature extraction model.

9. The abnormal operation behavior recognition system at the silk-making process site according to claim 6, characterized in that: The data dimension reduction model uses an improved autoencoder algorithm to perform feature dimension reduction.

10. The abnormal operation behavior recognition system at the silk-making process site according to claim 9, characterized in that: It also includes a third training module, which includes an initialization module, an encoding module, a hidden layer processing module, a decoding module, a reconstruction error calculation module, an optimization module, and a network parameter updating module; The initialization module is used to randomly initialize the weights and biases of the encoder and decoder according to a normal distribution; The encoding module is used to input the second feature vector into the encoder to achieve dimensionality reduction; The hidden layer processing module is used to process the hidden layer output using a nonlinear activation function, thereby increasing the nonlinear capability of the model and improving the efficiency of information expression; The decoding module is used to output the hidden layer through a decoder to obtain a third feature vector; The reconstruction error calculation module is used to compare the third eigenvector and the second eigenvector, and calculate the reconstruction error using a mean square error loss function; The optimization module is used to adjust the weights and biases of the encoder and the decoder according to the reconstruction error, and the influence of quantum state regularization is continuously considered during the optimization process; The network parameter updating module is used to update the network parameters according to the result of back propagation to reduce the reconstruction error and enhance the stability of the model.

Images